Hybrid model integrating LeViT transformer and distillation techniques for pattern detection and dance classification Authors: Yanyan Wang Published: 2025 ABSTRACT Artificial intelligence (AI) has become an integral part of modern life, extending its impact into the preservation of cultural heritage. This study applies state-of-the-art vision transformer models for the classification of traditional Chinese dance types. The main aim is to introduce a lightweight Vision Transformer (LeViT) architecture integrated with distillation mechanism to capture complex cues and visual patterns from images. It is evaluated alongside the standard Vision Transformer (ViT) and a baseline Convolutional Neural Network (CNN) model, which remains a benchmark in computer vision research. The dataset comprises images of three Chinese dance forms that follows standard preprocessing techniques including noise removal, geometric augmentation, and contrast enhancement. After feature extraction and supervised training, the proposed LeViT model demonstrated superior performance in terms of accuracy and stability compared with ViT and CNN. The findings confirm that transformer-based architectures, particularly LeViT, effectively capture subtle motion cues and fine-grained spatial patterns within dance imagery, offering a promising framework for future research in AI-driven cultural preservation and visual arts classification. Introduction Culture has long been an integral part of human civilization, shaping traditions, identities, and artistic expressions. Among these, dance stands as one of the most profound and dynamic forms of cultural expression, reflecting historical narratives, regional diversity, and artistic creativity1,2. People engage with dance in various ways, from performance and entertainment to therapy, education, and ritualistic practices. As societies evolve, dance continues to serve as an essential medium for storytelling and emotional expression, influencing domains such as performing arts, social events, and even rehabilitation therapy3. It serves as a universal language that exceeds barriers, enhance social cohesion and emotional linking4. Dance is deeply embedded in cultural traditions, preserving history and storytelling across generations5. Moreover, it is a tool for education, therapy, and social activism, enabling individuals to express emotions, challenge societal norms, and advocate for change6. With the rise of digital media and globalization, preserving, analyzing, and classifying different dance forms has become an important endeavor, fostering cultural appreciation and ensuring that traditional practices are not lost over time7. The growing interest in dance classification has led to the development of computational techniques to systematically analyze and categorize dance movements8. Traditionally, classification relied on human expertise, where choreographers, dance instructors, or historians manually identified distinct styles based on movement patterns, attire, and music synchronization. However, these conventional approaches are inherently subjective, time-consuming, and require extensive domain knowledge9. With the advancement of AI, particularly in the fields of computer vision and deep learning, automated systems can now accurately classify dance styles with minimal human intervention10. Deep learning models and AI-based methods, in general, have shown substantial promise in identifying and categorizing11 numerous artistic and physical actions, including dance. Initial machine learning systems were based on manual features and classical computer vision algorithms, and it was usually restricted to represent the complex motion patterns12. In more recent times, convolutional neural networks (CNNs) have been used with great success in dance recognition tasks, with learners being able to more effectively learn hierarchical features of images and videos. Nonetheless, CNN-based techniques continue to have generalization and scalability issues with large and heterogeneous data13. Vision Transformers (ViTs) have been proposed to solve these shortcomings, becoming an effective alternative to dance classification. In contrast to CNNs, which pay attention to local spatial relationships, ViTs include self-attention mechanisms to obtain global context, and thus, they are most effective with tasks that require complex visual representations14. Recent papers have shown that transformer-based models are more powerful than classical CNN models used in image classification tasks, such as human activity recognition and artistic style classification. Being capable of capturing long-range correlations and storing complex information, Vision Transformers are a perfect solution to the problem of classifying dance forms with high accuracy and high strength15. A dataset used in this study based on variations of Chinese dance and the divided into three classes, which are China Fan Dance, Sword Dance, and Dragon Dance. The sample will consist of 3815 images that will record the various movement patterns and postures of each style of dance. To form a baseline of the performance, traditional machine learning models were taken as the baseline classifiers and then a large-scale comparison to various versions of Vision Transformers was performed. This aims to determine the performance of transformer-based architecture in learning and separation of minor differences between dance styles, both of which will ultimately enhance automated dance recognition. A novel dataset comprising three distinct Chinese dance styles is introduced, providing a valuable resource for dance classification research.A comparative analysis of traditional machine learning models, CNN-based architectures, and Vision Transformer variants is conducted to evaluate classification performance.The LeViT model achieves the highest accuracy of 96%, outperforming both traditional CNNs and other Vision Transformer models, demonstrating its effectiveness in fine-grained dance classification. A novel dataset comprising three distinct Chinese dance styles is introduced, providing a valuable resource for dance classification research. A novel dataset comprising three distinct Chinese dance styles is introduced, providing a valuable resource for dance classification research. A comparative analysis of traditional machine learning models, CNN-based architectures, and Vision Transformer variants is conducted to evaluate classification performance. A comparative analysis of traditional machine learning models, CNN-based architectures, and Vision Transformer variants is conducted to evaluate classification performance. The LeViT model achieves the highest accuracy of 96%, outperforming both traditional CNNs and other Vision Transformer models, demonstrating its effectiveness in fine-grained dance classification. The LeViT model achieves the highest accuracy of 96%, outperforming both traditional CNNs and other Vision Transformer models, demonstrating its effectiveness in fine-grained dance classification. The remainder of this paper is organized as follows: Sect. 2 provides an overview of related work in dance classification and deep learning approaches. Section 3 details the dataset collection, preprocessing techniques, and experimental setup. Section 4 presents the methodology, including model architecture and training procedures. Section 5 discusses the results and comparative analysis of different models. Finally, Sect. 6 concludes the study with key findings and future research directions. Related work The use of deep learning techniques in dance classification has highly evolved to the point where the identification of dance styles and movements just from images has made great progress. To improve accuracy and efficiency of dance classification, several researchers have investigated different architectures, including CNN architectures, DNN architectures, geometric deep learning, and vision transformer-based studies, etc. In the study of, Li et al.16 proposed a DNN based approach to folk dance training and classification to deep learning to improve movement identification from image data with the help of Enhancing feature extraction. However, this approach made only limited generality to other dance forms as it primarily worked with folk dances. In Dewan et al.17, they proposed a deep learning pipeline to Indian dance style classification which includes CNN based architectures to increase accuracy. Though their results were effective, their observation is important for the real world, where the data variability is higher and the applicability of the study of dance classification in a controlled dataset is raised. Naik et al.18 also made another great contribution by applying ResNet34 to classify Indian classical dance images. The main contribution of their study was to show the capability of CNNs to process complex visual forms observed in dance postures. Nevertheless, their analysis was limited to the dataset’s small size and variety, which may limit the generalization of their results to other dance forms. In the same way, Karumuri et al.19 investigated classifying emotional expressions in dance motions through CNNs. In their study, they encoded the full body motion capture (Mo-cap) data into an 8-bit RGB image format which demonstrates the potential of learning affective features from dance using CNNs. The study was innovative, yet as all dance applications may not have access to Mo-cap data, this study relied on it. In his study, Wang et al.20 integrate deep learning with the Internet of Things (IoT) for investigating high level dance movements. The intersection of over here is between AI and smart technology in dance analysis research, they underscored. The feasibility of large-scale implementation of ICT is, however, compromised by the need for IoT integration. In Castro et al.21, did a study of dance classification in video data where they study online dance videos, and the challenges of availability and quality of the dataset. Current large scale video datasets lack sufficient annotated dance movements for deep learning models, and they pointed out the drawbacks of such datasets. Other than the image-based approaches, Lin et al.22 had also used the geometric deep learning technique for classifying 3D Indian classical point cloud data. While this approach enhanced depth-aware feature extraction, it did not make the data acquisition process easily available to popular use. Mohanty et al.23 also took another approach to recognize movement of dance in controlled laboratory settings using Kinect using deep learning. However, their system, Nrityabodha, effectively extracted features from dance postures but today’s Kinect devices were required by their system to be only used in tickets conditions. In Feng et al.24 focused on dance movement arrangement where they used deep learning to improve the movement classification accuracy by incorporating a super pixel segmentation. Although their approach increases the precision of segmenting, the classification performance of complex dance sequences is still unclear. Similarly, Mohanty et al.25combined affective computing to analyze emotional recognition in Indian classical dance videos using the techniques of deep learning to class emotions presented in dance performances. In a hybrid system using CNN and SVM Gupta et al.26, classified respectively Classical, Bhangra, Popping, Hip Hop and Ballet with an accuracy of 94.2%. This is why the high performance of their model shows that combining deep learning with classic machine learning classifiers works well. Another study27 proposed a fuzzy deep learning technique with IoT for dancing pose recognition. While encouraging, the study did not offer explicitly defined accuracy metrics that could support comparison of performance. In Jain et al.12, the authors enhanced the architectures of traditional CNNs for classifying Indian classical dance forms with an accuracy of 95.6%. Their findings demonstrated that image thresholding was a pre-processing technique that produced improved classification accuracy. Mulyanto et al.28 performed an experiment on the use of a Video Vision Transformer with Tubelet Embedding for traditional dance style classification. The study had also shown improvement over the conventional CNN based methods in processing features in temporal dimension; this was in the cases of dance sequences. A hybrid model based on a transformer encoder with CNN was proposed by Qu et al.29 to evaluate dance movement quality but lacked high quality annotated datasets. In Fan et al.15 introduced a cross-modal transformer for dance style transfer, which aimed to adjust movements from one style to another while making them suitable to classify. Like this, in their work of choreographic quality transforms for dance performance assessment, Kim et al.30 use kinematic data using visual attention mechanisms, their approach classified the movement quality but with scalability due to the dependency on the sensor-based input. Zhong et al.31 developed DanceMVP, a self-supervised transformer model for multi-task primitive-based dance assessment. Among other choreographies, their study also included Urban, Jazz, and K–Pop, which show the versatility of transformers in multi modal learning. Nevertheless, data collection was challenging due to their reliance on large pretraining datasets. In study32, FCTNet is introduced, a fusion model of 3D CNNs and transformers applied to dance action recognition. Although it was very useful for discerning complex movements, the demands on computation made it infeasible for real time applications. To generate partner dance motion in virtual reality environments, Wu et al.33 employed a transformer-based model and synthesized complementary movements. But this approach that trained itself could be used for training and had not been validated with real world data. Similarly, in Pan et al.34, they employed a transformer controller for various dance synthesis by synthesizing key frame-based movements with good fidelity. Nevertheless, the method lacked generalization on other dance styles. In fact, Yin et al.35 presented a multimodal dance style transfer model with vision transformers for better classification and extraction. The study also demonstrated the advantage of using transformer-based motion adaptation, yet classification loss in the domain remained a big challenge. In this work, Neethu et al.36 use a fused Vision Transformer and CNN approach to understand and classify facial expressions in Kathakali dance. While their work showed the effectiveness of transformers in expression recognition, they needed a significant number of labeled datasets for optimization. Transformer based; In a work by Kritis et al.37 DanceConv’s dance motions are synthesized from music using a Transformer-based audio encoder. What they showed was the potential of transformers in music driven dance generation but were missing work on accuracy of movement synchronization. Likewise, DanceFormer38 is a parametric motion transformer for music conditioned 3D dance generation. Whilst this method improved movement coherence and fluidity that was hard for it to adapt to various dance styles. In these studies, the transformer is shown to be an effective tool for making predictions in dance classification and generation, but lessons are drawn to continue the investigation on real-time processing, dataset generalization, and computational efficiency39. More studies prove that deep learning is effective in various fields of medical and agricultural diagnostics. Extensive testing with CNN- and ViT-based models on the SIPaKMeD pap-smear dataset has demonstrated that ViT architectures, with data augmentation and ensemble learning, are better than conventional CNNs, and hold promising prospects of clinical application, allowing detection of cancer at an early and accurate stage40. Likewise, in the agricultural field, the MetaFormer architecture has been deployed to perform olive leaf disease detection, combining the separable self-attention with the lightweight architecture to attain state-of-the-art performance on two benchmark data sets and outperform state-of-the-art models Swin, MaxViT, and ResNet5041. These papers emphasize that models based on transformers combined with effective architectural design can also improve the development of computer-aided diagnostics in the fields by ensuring the balance of accuracy, interpretability, and computation efficiency. Most of these methods, despite the state of the art in the existing literature on this research area of classifying dances with the help of CNNs, RNNs, and transformer-based models have drawbacks related to scalability, generalization, and efficiency. Transformer-based neural networks used to handle fine-grained variations of poses, but they require large, labeled datasets and are computationally expensive and restrict their practical use to real-time applications. CNN-based methods do not handle fine-grained variations of poses, and their high computational cost makes them unsuitable for real-time applications. Additionally, some studies use artificial or manipulated datasets, instead of the real-life data of cultural dance, and this makes them less applicable to practice. Notably, previous studies have mostly failed to consider lightweight architectures that are balanced accuracy and yet are computationally efficient. These gaps demonstrate the necessity of a model that can be useful in identifying the presence of complex visual cues in dance and be viable to be used in educational and cultural preservation settings. The proposed LeViT with knowledge distillation framework is designed to address this unmet need by improving accuracy, reducing computational overhead, and adapting to real-world Chinese dance datasets. Proposed research methodology In this study, the adopted research methodology is a systematic approach to achieve the effective recognition and classification of dance with Vision Transformer based models. This methodology includes data collection, preprocessing, model selection, training, evaluation and performance validation, as flow defined in Fig. 1. First, preprocessing steps, including image normalization, and resizing, are employed to improve model generalization. Since, the study uses Vision Transformer (ViT) models with the use of self-attention mechanisms to increase the accuracy in feature extraction and classification. Various standard classification metrics such as accuracy, precision, recall, and F1-score were used to evaluate performance, and a thorough model effectiveness was assessed. Finally, computational efficiency, as well as the inference speed and real time applicability are also analyzed to evaluate if these models can be deployed into the retail and industrial environments. Fig. 1Framework of proposed methodology. Framework of proposed methodology. Data preprocessing The preprocessing phase is crucial for improving the quality and robustness of the dataset, and it needs to standardize the images for deep learning models42. This phase involves a sequence of transformations to remove noise, augment the data, and normalize the images, as sample images of applying preprocessing steps shown in Fig. 2. The first step, which is necessary to remove distortions and artifacts from the images, is noise removal of images. Gaussian filtering is applied to smooth images, which is mathematically represented as in Eq. (1). For salt-and-pepper noise, median filtering is applied, which replaces each pixel value with the median of its neighboring pixels. Table 1 shows a brief understanding of symbols used in equation representation.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{\text{filtered}}\left(x,y\right)=\frac{1}{2{\uppi\:}{{\upsigma\:}}^{2}}{\sum\:}_{i,j}I\left(x+i,y+j\right){e}^{-\frac{{i}^{2}+{j}^{2}}{2{{\upsigma\:}}^{2}}}$$\end{document} Next, the dataset is geometric augmented to widen the range of the dataset and enhance the model’s ability in generalizing over unseen images due to some eventual overfitting. It specifically includes horizontal and vertical flipping as well as all images are resized to the same resolution of 224 × 224 using bilinear interpolation to ensure consistency across the dataset, computed using Eq. (2).2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{\text{flipped}}\left(x,y\right)=I\left(W-x,y\right)\:\text{(horizontal\:flip)}$$\end{document} Rotation is also applied to the images by an angle \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document}, using the transformation matrix, defined in Eq. (3):3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[\begin{array}{c}{x}^{{\prime\:}}\\\:{y}^{{\prime\:}}\end{array}\right]=\left[\begin{array}{cc}cos\theta\:&\:-sin\theta\:\\\:sin\theta\:&\:cos\theta\:\end{array}\right]\:\left[\begin{array}{c}x\\\:y\end{array}\right]$$\end{document} Additionally, all images are resized to a uniform resolution of 224 × 224 using bilinear interpolation to ensure consistency across the dataset. Lastly, apply photometric augmentation to simulate variations in lighting conditions. Equation (4) is used to perform brightness adjustment.4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{\text{adjusted}}={\upalpha\:}I+{\upbeta\:}\:\left({\upalpha\:}\in\:\left[\text{0.8,1.2}\right],{\upbeta\:}\in\:\left[-\text{10,10}\right]\right)$$\end{document} These transformations ensure that the dataset is robust to real-world variations and improve the model’s generalization capabilities. These preprocessing steps collectively ensured that the dataset was diverse, balanced, and optimized for training, enabling the models to achieve robust and accurate predictions in classifying the dance forms43. Fig. 2Sample images of applied preprocessing phase. Sample images of applied preprocessing phase. Feature extraction and classification During this phase, deep learning models are used to classify between various types of dances using feature extraction and classification. The LeViT (Hybrid CNN-Vision Transformer) proposed model combines the advantages of CNNs and ViTs to strengthen within local and global features44. For comparative analysis, the baseline models are also implemented such as CNN, default vision transformer. LeViT proposed model Hybrid deep learning model LeViT which combines the merits of CNNs and ViTs towards classifying the dance types into relevant categories is introduced. Figure 3 shows the architecture of proposed model working. It is designed to deal with both local features (e.g., edges, textures), as well as global dependencies (e.g., overall layout, patterns) in images, and thus is very appropriate for tasks that need fine-grained understanding45. The architecture begins with the CNN backbone which extracts hierarchical features from input image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:I$$\end{document}. This is achieved through a series of convolutional layers, where each layer applies a set of learnable filters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:K$$\end{document} followed by a ReLU activation function. The output feature map is computed using Eq. (5).5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{CNN}={f}_{\text{CNN}}\left(I\right)={\sum\:}_{i=1}^{{C}_{\text{in}}}{\sum\:}_{j=1}^{k\times\:k}{W}_{i,j}\text{*}{I}_{i,j}+b\:\to\:\text{Conv2D}\left(K,\text{ReLU}\right)$$\end{document} The CNN backbone ensures that local special features are effectively captured, which is crucial for analyzing detailed and steps analysis as elements in dance variations. Moreover, the feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{CNN}$$\end{document} is divided into non overlapping patches, which are flattened and projected into a lower dimensional space using a learnable projection matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{p}$$\end{document}. This procedure is called batch embedding which is presented using Eq. (6): This ensures that transform the 2D feature map into a sequence of patch embeddings which are fed into the transformer encoder.6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{o}=\left[{x}_{p}^{1}{W}_{p},{x}_{p}^{2}{W}_{p},\dots\:\dots\:,{x}_{p}^{N}{W}_{p}\right]+{E}_{pos}\:$$\end{document} Furthermore, Encoder Transformer is composed of multiple layers of MSA and FFNs. Self-attention mechanism calculates attention scores among all the patch embeddings so that long range dependencies can also be learnt by the model. It is specified that attention mechanism is defined using Eq. (7).7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{z}}_{\text{l}}^{{\prime\:}}=\text{LayerNorm}\left({\text{z}}_{\text{l}}+\text{Concat}{\left(\text{Attention}\left({\text{z}}_{\text{l}}{\text{W}}_{\text{Q}}^{\text{i}},{\text{z}}_{\text{l}}{\text{W}}_{\text{K}}^{\text{i}},{\text{z}}_{\text{l}}{\text{W}}_{\text{V}}^{\text{i}}\right)\right)}_{\text{i}=1}^{\text{h}}{\text{W}}_{\text{O}}\right)$$\end{document} Finally, the classification head processes the output feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{z}}_{\text{l}}$$\end{document}​ through a fully connected layer with SoftMax activation to produce class probabilities, using Eq. (8).8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P\left(y|I\right)=\text{softmax}\left({W}^{T}{z}_{L}+b\right)$$\end{document} The SoftMax function makes the output probability distribution over the classes. Taken together, LeViT takes advantage of the difference between CNNs and transformers in patterns, style, their complementary strengths in feature extraction and classification. However, LeViT combines the two by enabling CNN backbone to represent local detail and transformer encoder to learn the global relationships, that makes it excellent for decomposition for centric analysis of various dance types. Fig. 3Proposed model architecture. Proposed model architecture. The novelty of this work lies in integration of Levit + distillation which was not explored in the existing studies of relevant literature. The main contribution with integrating the computational efficiency of the LeViT transformer with a specially designed distillation strategy, which passes knowledge of a larger, high-capacity teacher model to a lightweight student model. This design accomplishes two distinct innovations, specifically: (i) it allows powerful performance on complex and fine-grained Chinese dance motions where subtle visual signals are essential, and (ii) it demands much lower training and inference expenses than the standard ViT and CNN baselines, which makes it more feasible in practice in terms of real-time cultural preservation and learning. Therefore, the suggested LeViT + distillation model is not simply an incremental one, but it includes a new architecture, whose accuracy, efficiency, and cultural domain specificity are balanced. Baseline models To evaluate the effectiveness of the proposed Vision Transformer based model by matching it with some of the best deep learning architectures that are used for the task of dance types of recognition. At baseline, local feature extraction from CNN is followed by structured network scaling with simple vision transformer model is the union of Convolutional layers and transformer-based attention mechanisms. The proposed LeViT approach is benchmark references to evaluate the gains in classification accuracy, the feature extraction efficiency, as well as the computational performance over these models. Traditional CNN model Among deep learning models for image-based tasks, Convolutional Neural Networks (CNN) are the most popular model used. CNNs are particularly suitable for extracting spatial hierarchies in images (feature extraction), that can recognize the presence of logos, colors and positioning of text (elements in packaging)46. Convolution operation is the fundamental operation in CNN, the same is given by Eq. (9).9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{ij}^{l}=f\left({\sum\:}_{m}{\sum\:}_{n}{w}_{mn}^{l}{x}_{\left(i+m\right)\left(j+n\right)}^{l-1}+{b}^{l}\right)$$\end{document} The next step is to use the pooling layer (e.g. max pooling) which has reduce the dimension of spatial space and maintain key features that keep good computational computation efficiency and ward off overfitting. The output of the extracted features prior to the final stage of FC layers is flattened and passed through a SoftMax activation function for classification47. CNNs work exceptionally well to recognize images in image recognition tasks such as consumer attention modeling and packaging analysis, but the drawbacks of CNNs with long range dependencies and contextual relationships of images make it challenging to learn. Vision transformer Another model, vision Transformer (ViT), utilized as baseline for dance type classification, an image of input is flattened as a sequence of patches, and the patch are embedded linearly to a lower dimensional space. The latter sends these patch embeddings, together with position encodings to maintain spatial information, into the standard Transformer encoder. It has multiple layers of multi-head self-attention and feed forward NN layers which facilitates the model in performing global patch pairs relationships and understanding of intricate patterns of the image48. So, a special classification token is preceded to the sequence and the final representation of this sequence after passing through the encoder form its input is fed into a fully connected layer to predict the dance type49. ViT takes advantage of the Transformer’s capacity to learn the long-ranged temporal dependency structure at different levels of pooling, hence, it effectively classes dances based on both local and global visual features in the image50. Machine requirements A high-performance computing setup is needed to ensure efficient training and deployment of Vision Transformer based models for dance type recognition, as requirements are clearly defined in Table 2. To tackle huge quantity of image processing, the system needs to utilize a multi-core processor, high-end GPU that guarantees fast data access, with stable power supply as well as an advanced cooling system keep things from overheating during long training sessions. Finally, model implementation requires deep learning frameworks such as PyTorch or TensorFlow and is run on any OS. Furthermore, data retrieval over high-speed internet connectivity as well as cloud-based experimentation is required. Table 1Description of applied equation Symbols.SymbolsDescriptionSymbolsDescription \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{I}}_{\text{filtered}}\left(\varvec{x},\varvec{y}\right)$$\end{document} Intensity Pixel at location \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(x,y\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{W}}_{\text{O}}$$\end{document} Output head linearly projected \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\upsigma\:}$$\end{document} Standard deviation of Gaussian kernel \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:W\:\in\:{\mathbb{R}}^{D*C}$$\end{document} Learnable Weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{W}$$\end{document} Width of the images \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:b\:\in\:{\mathbb{R}}^{C}$$\end{document} Learnable Bias \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\upalpha\:}$$\end{document} Controls the contrast \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:C$$\end{document} No. of Classes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\upbeta\:}$$\end{document} Adjust the brightness \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{ij}^{l}$$\end{document} Output feature map at layer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:l$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{x}}_{\varvec{p}}^{\varvec{i}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i-th$$\end{document} patch \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{mn}^{l}$$\end{document} Convolutional kernel weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{W}}_{\varvec{p}}$$\end{document} Projection matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{\left(i+m\right)\left(j+n\right)}^{l-1}$$\end{document} Input feature map from previous layer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{E}}_{\varvec{p}\varvec{o}\varvec{s}}$$\end{document} Positional Embedding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f(.)$$\end{document} Activation function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{z}}_{\varvec{l}}$$\end{document} Input embeddings \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{A}_{ij}$$\end{document} Interconnection between features \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{h}$$\end{document} Parallel head \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{j}$$\end{document} Feature Vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{W}}_{\mathbf{Q}}^{\mathbf{i}}$$\end{document} Query \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{k}$$\end{document} Scaling Factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{W}}_{\varvec{k}}^{\mathbf{i}}$$\end{document} Key \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{W}}_{V}^{\text{i}}$$\end{document} Value Table 1Description of applied equation Symbols.SymbolsDescriptionSymbolsDescription \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{I}}_{\text{filtered}}\left(\varvec{x},\varvec{y}\right)$$\end{document} Intensity Pixel at location \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(x,y\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{W}}_{\text{O}}$$\end{document} Output head linearly projected \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\upsigma\:}$$\end{document} Standard deviation of Gaussian kernel \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:W\:\in\:{\mathbb{R}}^{D*C}$$\end{document} Learnable Weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{W}$$\end{document} Width of the images \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:b\:\in\:{\mathbb{R}}^{C}$$\end{document} Learnable Bias \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\upalpha\:}$$\end{document} Controls the contrast \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:C$$\end{document} No. of Classes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\upbeta\:}$$\end{document} Adjust the brightness \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{ij}^{l}$$\end{document} Output feature map at layer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:l$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{x}}_{\varvec{p}}^{\varvec{i}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i-th$$\end{document} patch \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{mn}^{l}$$\end{document} Convolutional kernel weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{W}}_{\varvec{p}}$$\end{document} Projection matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{\left(i+m\right)\left(j+n\right)}^{l-1}$$\end{document} Input feature map from previous layer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{E}}_{\varvec{p}\varvec{o}\varvec{s}}$$\end{document} Positional Embedding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f(.)$$\end{document} Activation function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{z}}_{\varvec{l}}$$\end{document} Input embeddings \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{A}_{ij}$$\end{document} Interconnection between features \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{h}$$\end{document} Parallel head \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{j}$$\end{document} Feature Vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{W}}_{\mathbf{Q}}^{\mathbf{i}}$$\end{document} Query \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{k}$$\end{document} Scaling Factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{W}}_{\varvec{k}}^{\mathbf{i}}$$\end{document} Key \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{W}}_{V}^{\text{i}}$$\end{document} Value Description of applied equation Symbols. Table 2Machie requirements for implementation.ComponentSpecificationDescriptionProcessor (CPU)Intel Core i9 / AMD Ryzen 9 (12 + Cores, 3.5 GHz+)High-speed computation for preprocessing and model training tasks.Graphics Card (GPU)NVIDIA RTX 3090 / A100 / Tesla V100 (24GB + VRAM)Essential for Vision Transformer model training and inference acceleration.RAM (Memory)32GB + DDR4/DDR5Handles large-scale image data and deep learning model execution.Storage1 TB NVMe SSDFast read/write speeds for handling large datasets and model checkpoints.Operating SystemUbuntu 20.04 / Windows 11 / macOS (Latest Version)Compatibility with deep learning frameworks (PyTorch, TensorFlow).Deep Learning FrameworkPyTorch 2.0 / TensorFlow 2.10+Supports Vision Transformer models with optimized GPU acceleration.Power Supply (PSU)800 W+ (Platinum Rated)Ensures stable power for high-performance computing components.Cooling SystemLiquid Cooling or High-Performance Air CoolingPrevents overheating during prolonged training sessions.NetworkingHigh-speed internet (1Gbps+)Facilitates dataset downloading, cloud-based model training, and updates. Table 2Machie requirements for implementation.ComponentSpecificationDescriptionProcessor (CPU)Intel Core i9 / AMD Ryzen 9 (12 + Cores, 3.5 GHz+)High-speed computation for preprocessing and model training tasks.Graphics Card (GPU)NVIDIA RTX 3090 / A100 / Tesla V100 (24GB + VRAM)Essential for Vision Transformer model training and inference acceleration.RAM (Memory)32GB + DDR4/DDR5Handles large-scale image data and deep learning model execution.Storage1 TB NVMe SSDFast read/write speeds for handling large datasets and model checkpoints.Operating SystemUbuntu 20.04 / Windows 11 / macOS (Latest Version)Compatibility with deep learning frameworks (PyTorch, TensorFlow).Deep Learning FrameworkPyTorch 2.0 / TensorFlow 2.10+Supports Vision Transformer models with optimized GPU acceleration.Power Supply (PSU)800 W+ (Platinum Rated)Ensures stable power for high-performance computing components.Cooling SystemLiquid Cooling or High-Performance Air CoolingPrevents overheating during prolonged training sessions.NetworkingHigh-speed internet (1Gbps+)Facilitates dataset downloading, cloud-based model training, and updates. Machie requirements for implementation. Experimental setup In this section experimental setup has been discussed to use a structured dataset and standardized evaluation metrics to evaluate the performance using standard measures of Accuracy, precision, recall and F1_score, are utilized to evaluate the performance of the model, and confusion matrix analysis, confidence distribution for model robustness and misclassification trends. Dataset The dataset used is for image classification of various traditional dance utilized to capture the fine steps, complex dance movements and distinctive features of a dance. It includes 3815 images limited to three dance categories, China Fan dance (1315 images), Sword Dance (1250 images) and Dragon Dance (1250 images). The dataset is of dancers in various poses, gestures, forms, including costumes, prop, and dynamic movement sequences. For model training and evaluation, the dataset is split into a 70:30 ratio into training and testing sets, ensuring a balanced distribution of images across all categories. The details of the dataset split are shown in Table 3: Table 3Description of dataset images.ClassesActual imagesTraining imagesValidation imagesTesting ImagesChina Fan Dance13151158157394Sword Dance12501100150375Dragon Dance12501100150375Total381533584571144 Table 3Description of dataset images.ClassesActual imagesTraining imagesValidation imagesTesting ImagesChina Fan Dance13151158157394Sword Dance12501100150375Dragon Dance12501100150375Total381533584571144 Description of dataset images. From three dances, each is a form of cultural heritage: China Fan Dance has beautiful fan moving, quick Mind, and vivid transitions; Sword Dance for accurate use of the weapon and martial arts influence; Dragon Dance features coordinated group movements with elaborate dragon props, as sample images of these classes shown in Fig. 4. The dataset enables such deep learning applications in motion analysis, performance evaluation, pose estimation, cultural heritage preservation. The application of the proposed method leads to its utilization for training computer vision models in dance recognition, human pose estimation, and real time performance tracking to support their application to AI for analysis of choreography and interactive media applications. It provides a balanced dataset that is sufficient for training to provide some diversity and prevent the model from overfitting and to have good ability discriminant fine grained features among the three dance types. Fig. 4Sample images of label from dataset. Sample images of label from dataset. Performance evaluation measures Several performance evaluation metrics are then used to assess how well deep learning models can classify dance. All these metrics give us the idea of model reliability, class wise predictive capability as well as overall classification capability51. Figure 5 outlines each measure, its significance, and the corresponding mathematical formulation. Fig. 5Analysis of performance measures. Analysis of performance measures. Results and discussion The performance is evaluated of the proposed LeViT model with baseline architectures. Finally, this analysis is performed based upon the accuracy, precision, recall, F1-score, and computational efficiency. Compared with conventional CNN models, the proposed model does better in predicting the classification and feature extraction at the same time, as well as in the aspect of better generalization. The model is also evaluated in terms of training convergence, loss trends, real time inference speed, and confidence scores and to predict whether it is robust enough and can be deployed in dance type recognition applications. The results in Table 4 indicate evident discrepancies between CNN and ViT and the suggested LeViT-based framework in the Chinese dance type recognition scenario. CNN has the lowest overall performance, with an accuracy of 89% and F1-score 89% meaning that the network is not very good at generalizing between dance classes. Whereas, CNN is relatively high in recall of some classes like China Fan Dance (98%), its precision is low (81%), which indicates that the method has a tendency of giving false positives. This is an imbalance that indicates that CNN has not always recorded the finer details of the cultural dances, whose minute differences between posture and costumes are essential in proper classification. ViT model is based on CNN with the added benefit of self-attention, which results in an overall accuracy of 93% and equal precision, recall, and F1-scores. As an illustration, it achieves almost perfect accuracy when it comes to Sword Dance (100%), and high F1-scores when it comes to Dragon Dance (97%). Nonetheless, the inconsistency between classes also indicates that ViT, despite being effective, may also fail to cope with domain-specific complexity of dance images, especially in the case of small datasets. The fact that it needs large-scale training information to be optimal means that it is not always well suited to real-world applications where annotated cultural datasets are typically scarce. Conversely, the introduced LeViT with knowledge distillation has the highest performance of 96% accuracy and 93% F1-score on all the classes. It does not have the same precision recall imbalance as CNN and in comparison, to ViT, it has consistently high results on China Fan, Sword, and Dragon Dance categories. This stability shows the two-fold advantage of the efficient transformer architecture of LeViT and the distillation approach that complement each other to support strong recognition of subtle motions of a dance with lower computational costs. The proposed model, therefore, supports accuracy, precision and efficiency and becomes a more viable and domain-oriented model in solving real world dance classification tasks. Table 4Results with proposed model LeViT using holdout.ClassesAccuracyPrecisionRecallF1-ScoreAccuracyPrecisionRecallF1-ScoreAccuracyPrecisionRecallF1-ScoreProposed model - LeViTViTCNNChina Fan Dance9692979593851009289819889Sword Dance9696969693100869289997585Dragon Dance969994969398959789919493Overall Results 96 93 93 93 9393939389898989 Table 4Results with proposed model LeViT using holdout.ClassesAccuracyPrecisionRecallF1-ScoreAccuracyPrecisionRecallF1-ScoreAccuracyPrecisionRecallF1-ScoreProposed model - LeViTViTCNNChina Fan Dance9692979593851009289819889Sword Dance9696969693100869289997585Dragon Dance969994969398959789919493Overall Results 96 93 93 93 9393939389898989 Results with proposed model LeViT using holdout. Further, these models also compare their performance as proven in the confusion matrix analysis, shown in Fig. 6. The CNN model (Fig. 6a) exhibits considerable misclassifications with 51 samples in Sword Dance being misclassified with China Fan Dance that CNN struggles to differentiate visually similar patterns. The ViT model (Fig. 6b) helps to lessen this confusion but continues to fail to classify 68 Sword Dance images, with some overlap of China Fan and Dragon Dance, indicating that it is partially sensitive to finer cultural characteristics. Conversely, the suggested LeViT version (Fig. 6c) has the best balanced and correct classification in all categories with a slight misclassification in Sword and Dragon Dances. This consistency highlights the power of the integration of LeViT lightweight transformer architecture with the process of distillation since the model can achieve the ability to identify finer details of the Chinese dance movement at high efficiency and robustness. Fig. 6Confusion matrix Analysis of (a) CNN (b) ViT (c) LeViT. Confusion matrix Analysis of (a) CNN (b) ViT (c) LeViT. The plots for training and Validation accuracy with different epoch ranges (1–25, 26–50, 51–75, 76–100) in Fig. 7, show that as time goes by the model performance also improves steadily. The model first learns quickly, and training accuracy rises steeply in the first 25 epochs up to greater than 90%. We observe however that validation accuracy increases along the same trends, however with much more fluctuations which could indicate generalization instability. In the more recent epochs (26–100), training accuracy settles close to 99%, while validation accuracy still fluctuates especially between 25 and 75 epochs suggesting possible overfitting. This is also reinforced in the loss curves. Validation loss jumps around the initial 20 epochs then falls and become flat, while training loss decreasing consistently to near zero. The fact that fluctuating validation loss indicates that the model cannot generalize means that while the model is learning the training data well, it is not learning too well in general. A good indication of the overfitting is the diverging training and validation loss in later epochs. The model performs very well on training accuracy, but the validation accuracy and loss show some fluctuations which indicate the need to develop methods to increase the generalization, such as regularization techniques, dropout, or early stopping, shown in Fig. 8. Fig. 7Training and validation accuracy analysis of model progress over epochs. Training and validation accuracy analysis of model progress over epochs. Fig. 8Loss trend analysis of model progress. Loss trend analysis of model progress. However, the model still shows a strong classification performance, high accuracy, however, the improvement of stability in the validation performance requires additional fine tuning. Furthermore, the resultant confidence plots give good indications about the model’s accuracy trend, loss trend and data splitting impact in Fig. 9. Clearly, there is higher confidence in the model with increasing epochs. It gets initially less accurate but gain in accuracy very quickly and level off at about 90%–95% after around 30 epochs. The general improvement with minor fluctuations is indicated in the trend line. The color gradient corresponding to the confidence values is an indication that most predictions have high confidence level as the training progresses. But occasionally, when the model goes slightly off with some data points, chances are it could not process those data points. The validation loss trend consists of large oscillations and a general decreasing trend from time to time, shown in Fig. 10. The loss is first high and quickly drops down, revealing that the model can learn well. Nevertheless, there are spikes in loss across epochs that show that the model gets unstable after some epochs, that is probably because it overfits to the training data. The color-coded loss values corroborate that some epochs have smaller loss, but some have often spiked loss value, as seen, and this can be due to challenging validation samples or possibly that there is utility in the data itself. The dataset confidence plot is used to visualize how the model’s confidence varies for different training, validation, and test dataset. Most predictions cluster around 95%–100%, and the training data is the most confident. Fig. 9Confidence plot analysis of model based on accuracy measure. Confidence plot analysis of model based on accuracy measure. Fig. 10Confidence plot analysis of model based on loss measure. Confidence plot analysis of model based on loss measure. Overall, the findings show that even though the proposed LeViT model has high overall accuracy and stability, there are still some challenges. The variation in validation and divergent loss curves indicate mild overfitting, which was partially countered using cross-validation, dropout, weight decay and data augmentation. The most striking misclassifications are within the Sword Dance group, when there is a visual homology between pose and costume with China Fan Dance and Dragon Dance resulting in confusion as indicated by the confusion matrices. These are some of the mistakes that point to the natural challenge of tasks on fine-grained cultural recognition. Furthermore, although the framework demonstrates potential in the context of efficiency and accuracy, the assertions of industrial applicability are to be viewed as potential. Before full applicability is certain, real-world deployment will need additional validation on larger and more diverse data sets and need to be run on mobile or edge devices in real cultural or educational settings. Further to confirm and analyze the proposed model performance, the confidence distributions of validation and test data are just a little less confident as the model becomes less certain when it encounters new data, shown in Fig. 11. One difference to note is in fact that there is a drop in confidence between non-training samples to the point that it is necessary to employ techniques like regularization or augmentation to close the gap. Based on the test results above, the model shows extremely high accuracy to classify Sword Dance images, and this accuracy always exceeds 99% in majority of cases. The predictions generalize strongly and even varied representation, such as artistic or real-world depictions, are correctly identified. But in some instances, the smallest drop in confidence (e.g. 81.24%) may point to the presence of ambiguity in taxing situations as we had low lighting or ambiguous compositions, as predictive results based on unseen data shown in Fig. 12. In real world testing, the model is highly accurate and reliable, which indicates its ability to work well in classification of such dance. Considering the overall performance, the model has high accuracy, however the fluctuations of loss and confidence scores across different datasets indicate a need for stability and generalization improvement. Fig. 11Confidence plot analysis of model based on data spilliting. Confidence plot analysis of model based on data spilliting. Fig. 12Predictive outcome using unseen data. Predictive outcome using unseen data. The cross-validation outcomes presented in Table 5 and substantiated with the confusion matrices shown in Fig. 13 give a comparison between CNN, ViT, and the proposed LeViT model to classify Chinese dance. The CNN baseline was the least performing in general, with an accuracy of 87% and balanced and moderate precision, recall, and F1-scores. The confusion matrix (Fig. 6a) that corresponds to the results shows that CNN has the tendency to misclassify visually similar classes, including Dragon Dance and China Fan Dance (6.2 out of 100), as well as Sword Dance (6.8 out of 100). This implies that CNN is dependent on local feature extraction thereby restricting its capability to recover the intricate fine-grained spatial dependencies that play a vital role in differentiating traditional dance movements. ViT model was a better approach as it reached 90% accuracy and better precision (93%), recall (93%), and F1-score (93%). ViT minimizes the misclassification of classes, as demonstrated in Fig. 13b, over CNN, especially in Dragon Dance, in which recall increased to 91.9%. Nevertheless, there is still an unresolved ambiguity between China Fan Dance and Sword Dance (5.1% and 5.0% respectively), suggesting that although ViT takes advantage of global self-awareness to learn long-range dependencies, it is also to some extent susceptible to differences in pose and costume in a limited set of data. The proposed LeViT model using distillation had the best overall performance, with an accuracy of 94% precision, recall, and F1-score at 93 consistently. The confusion matrix in Fig. 13c indicates that LeViT does not only reduce the errors in class level but also produces very reliable predictions regardless of the type of dance. As an example, Sword Dance is learned with a recall of 95.5 and slight overlap with other classes which reflects the capability of the model to learn the unique motion cues and cultural patterns within the fine-grained recognition of dance. In comparison to CNN and ViT, LeViT is more cost-efficiency and has strong accuracy, which makes it effective to be implemented in practice in cultural conservation and learning systems. The findings combined indicate that the proposed LeViT with distillation can be considered significant progress in the field of fine-grained dance classification as it allows to enhance the separability of classes, decrease the rate of misclassification, and provide consistent folds. The quantitative data are supported by the visual evidence of the confusion matrices (Fig. 13) that testify to the fact that the model serves its purpose in extracting the subtle visual and contextual aspects of traditional Chinese dance. Table 5Results with proposed model LeViT using cross validation.ClassesAccuracyPrecisionRecallF1-ScoreAccuracyPrecisionRecallF1-ScoreAccuracyPrecisionRecallF1-ScoreProposed model - LeViTViTCNNChina Fan Dance949392939090888987878787Sword Dance949396949091909187888888Dragon Dance949694959090929187868787Overall Results 94 93 93 93 9093939387878787 Table 5Results with proposed model LeViT using cross validation.ClassesAccuracyPrecisionRecallF1-ScoreAccuracyPrecisionRecallF1-ScoreAccuracyPrecisionRecallF1-ScoreProposed model - LeViTViTCNNChina Fan Dance949392939090888987878787Sword Dance949396949091909187888888Dragon Dance949694959090929187868787Overall Results 94 93 93 93 9093939387878787 Results with proposed model LeViT using cross validation. Fig. 13Confusion matrix Analysis of (a) CNN (b) ViT (c) LeViT. Confusion matrix Analysis of (a) CNN (b) ViT (c) LeViT. Figure 14 cross validation analysis shows that the proposed LeViT model is strong and consistent across different folds. Accuracy trends (Fig. 14a) reveal that the higher fold validations (7-fold and 10-fold) are characterized by a higher generalization rate in comparison to the lower fold validations (3-fold and 5-fold) because they portray increasing trends in accuracy. According to this, loss curves (Fig. 14b) show a steady reduction as variance is reduced in 7-fold and 10-fold configuration, which is in favor of the plausibility of optimization. The error rates are lower in the higher folds implying that the model would still perform consistently even when it is divided over several folds which improves the chances of the lack of overfitting and enhances the trust to the predictive ability of the model. These findings confirm that the proposed architecture does not only provide good classification performances but also shows good performances under varying cross-validation conditions, which is fundamental in any real implementation of the culture dance recognition in the practice. Fig. 14Cross-validation performance analysis using proposed model based on (a) Accuracy (b) Loss. Cross-validation performance analysis using proposed model based on (a) Accuracy (b) Loss. The hyperparameter settings for the dance classification models were carefully selected to optimize performance and ensure effective training, display in Table 6. A batch size was selected for computational efficiency while remaining stable for convergence of the models. The architecture derived the model depth and based on this, CNNs used deep convolutional layers for feature extraction and ViT used transformer encoders. To make it consistent with models, the input image size was standardized. The non-linearity was applied with ReLU, and so was the activation, and SoftMax was utilized in the last classification layer. A tradeoff between the computational and feature granularity was made with the best inputs being a batch size of 32 and input image size of 224 × 224. The cosine decay-based learning rate of 0.00001 became smooth in convergence but a dropout rate of 0.1 became effective in countering overfitting. The weight decay parameter and optimizer Adam (Beta1 = 0.9, Beta2 = 0.999, epsilon = 1e-6) have been optimized to trade off the training and generalization. Data augmentation strategies, including width height shift and the change in brightness also helped to augment the resilience against image variability. The early stopping parameter was also 5 epochs, and the parameter did not overfit but maintained its efficiency. The hyperparameters were optimized by experimental trial and error over the proposed LeViT model, and the model was put into a converged point, and a stable performance. To improve model robustness, data augmentation techniques like width and height shift, horizontal flip, brightness range and contrast modification were used. Long range dependency was captured by attention mechanisms in ViT, and hierarchical feature learning was made by feature extraction layers in CNNs. Parameter sharing and skip connections were used to build models, for the purpose of making learning more efficient. Using the optimized kernel size, pooling type, number of filters, and stride convolutions, feature extraction efficiency was achieved with a parameter efficiency balanced between model complexity and performance. This hyperparameter sets the computational resource for model training, as LeViT model has good computational efficiency and balance in training and validation accuracy as shown in Fig. 15. Memory consumption is only 16%, which means LeViT uses GPUs very efficiently compared to standard ViT, consuming at 1.74GB and this is a very small amount. Table 6Hyperparameter settings of proposed model LeViT.ParameterRange valuesBatch Size32Model DepthLeViT-128Input Image Size224 × 224Learning Rate0.00001Epoch Learning Rate0.0001 (Cosine Decay)Dropout0.1Training Epochs100OptimizerAdamWLoss FunctionCross-Entropy LossEpsilon1.00E-06Beta 10.9Beta 20.999Weight Decay0.0001Width and Height Shift0.2–0.3Brightness Range0.2–0.4Patience5Kernel Size3 × 3Number of Filters128Number of Parameters40 M Table 6Hyperparameter settings of proposed model LeViT.ParameterRange valuesBatch Size32Model DepthLeViT-128Input Image Size224 × 224Learning Rate0.00001Epoch Learning Rate0.0001 (Cosine Decay)Dropout0.1Training Epochs100OptimizerAdamWLoss FunctionCross-Entropy LossEpsilon1.00E-06Beta 10.9Beta 20.999Weight Decay0.0001Width and Height Shift0.2–0.3Brightness Range0.2–0.4Patience5Kernel Size3 × 3Number of Filters128Number of Parameters40 M Hyperparameter settings of proposed model LeViT. Despite its lightweight architecture, LeViT shows near optimal training accuracy (around 99%) and validate accuracy (around 90%) that shows strong generalization. However, the validation accuracy also fluctuates after 30 epochs, which might be caused by the dataset complexity and low mild overfitting. Compared to popular deep learning models ViT and CNN, LeViT has a similar drop in accuracy but remains efficient in terms of computation which is better than ViT and makes it an optimal tradeoff between accuracy and computational efficiency suited for real time deployment, mobile implementation and edge computing where the performance of a high-performance AI model must be constrained by resource constraint. Running under merely 14 MB of memory and even with stable GPU usage, LeViT shows promise of being scalable with no sacrifice in classification accuracy. Fig. 15Computational resource during training of model progress. Computational resource during training of model progress. Table 7 represents the additive value of each element of the proposed framework through the ablation study. According to the baseline LeViT-128 model, data augmentation improved the model by enhancing resistance to visual variability in the photographs of the dance. Regularization (dropout and weight decay) of training and regularizing further yielded similar gains in precision and recall. In case of a cosine learning rate schedule, convergence became smoother and additional accuracy improvements were obtained. The best improvement occurred when knowledge distillation was added to it, with a ViT-B/16 teacher transferring not only soft labels but also feature representations to the LeViT student achieving 96% accuracy and, more generally, a boost in the F1-score, albeit at the cost of inference time a bit. The results confirm that all four elements (augmentation, regularization, improved scheduling, and distillation) combined with each other result in a rise of the stability and effectiveness of the suggested solution capable of solving the issue of the fine-grained dance classification. The performance also increases gradually between the LeViT baseline (93.1) to the complete pipeline (96.0). Distillation with the greatest single gain and augmentation, and regularization stabilizes F1. Like the efficiency claim, distillation also results in a small latency reduction. Table 7Analysis of results based on ablation study.#ComponentsAugmentationReg (Dropout)Cosine LRDistillationAccuracyPrecisionRecallF1-ScoreParams (M)Inference (ms/img)1LeViT-128 baseline✗✗✗✗93.192.893.092.9406.12Baseline + Data Augmentation (shift/brightness) ✓ ✗✗✗94.093.693.993.7406.13(2) + Regularization (dropout 0.1, weight decay 1e-4)✓✓✗✗94.894.394.794.5406.24(3) + Cosine LR schedule ✓ ✓ ✓ ✗95.394.995.195.0406.25(4) + Knowledge Distillation (teacher ViT-B/16 → student LeViT-128)✓✓✓✓96.095.196.095.5405.8 Table 7Analysis of results based on ablation study.#ComponentsAugmentationReg (Dropout)Cosine LRDistillationAccuracyPrecisionRecallF1-ScoreParams (M)Inference (ms/img)1LeViT-128 baseline✗✗✗✗93.192.893.092.9406.12Baseline + Data Augmentation (shift/brightness) ✓ ✗✗✗94.093.693.993.7406.13(2) + Regularization (dropout 0.1, weight decay 1e-4)✓✓✗✗94.894.394.794.5406.24(3) + Cosine LR schedule ✓ ✓ ✓ ✗95.394.995.195.0406.25(4) + Knowledge Distillation (teacher ViT-B/16 → student LeViT-128)✓✓✓✓96.095.196.095.5405.8 Analysis of results based on ablation study. The comparative analysis of the proposed LeViT shows the best validation accuracy and loss trends, having steady learning curve not overly fluctuated as compared. Comparing to CNN, ViT has 93% accuracy while fail in classification inconsistency of Sword Dance and variate validation accuracy and loss due to unstable learning. The CNN has a lowest accuracy of 89%, high misclassification rate on Sword Dance, and is highly overfit according to unstable validation accuracy and confused confidence distribution. This finding is further supported by the analysis of confidence, as LeViT continues to have high confidence across training, validation, test and unseen data, while ViT is more variable and the confidence of CNN, which is obviously generalizing less, is the least reliable since it is unreliable for unseen data. It is found that LeViT’s robustness leads to predictions that are more certain than ViT in most cases, and CNN predicts the most inconsistent. LeViT finally demonstrates itself to be the best model in terms of computational efficiency and generalization capacity while maintaining the balance of feature extraction, outperforming both traditional CNN and transformer based ViT models. Comparison with existing studies Deep learning models and the proposed LeViT model, which is much smaller and faster are presented in Table 8 in the classification of image-based dances. The massive success that the previous methods have demonstrated in the dance postures recognition compared to ResNet34 (85%), 3D CNN (89%) are effective yet fail to address the intricate spatial-temporal correlation in dance motions. The CNN + SVM model uses convolutional feature extraction, and the normal machine learning classifiers to combine the two resulting in a higher accuracy of 92% which provides improved generalization. However, the suggested LeViT model outperforms these algorithms with 96% accuracy and also employs lightweight mechanisms of ViT to capture fine-grained features allowing computational efficiency. The enhanced performance demonstrates the potential of the transformer architecture in accordance with the potential of improved performance of managing the complexities of postures and variations in styles more accurately. Table 8Comparison of proposed with existing models.Sr. noRefYearModelResult (Acc in %)132020ResNet348521320213D CNN893112024CNN + SVM924 Proposed - LeViT 96 Table 8Comparison of proposed with existing models.Sr. noRefYearModelResult (Acc in %)132020ResNet348521320213D CNN893112024CNN + SVM924 Proposed - LeViT 96 Comparison of proposed with existing models. Conclusion and future work This study proposed a hybrid neural architecture that integrates the LeViT transformer with knowledge distillation for classifying Chinese dance types from image data. The model achieved superior accuracy compared to ViT and CNN while maintaining greater stability, reduced overfitting, and computational efficiency. These results demonstrate that lightweight transformer models, when enhanced with distillation, provide a practical balance between performance and efficiency, offering meaningful implications for real-world applications such as cultural preservation, dance education, and performance analysis. Beyond reporting performance improvements, this work underscores critical insights. CNN models showed difficulty in distinguishing visually similar classes, while ViT, though effective, exhibited fluctuations that limit reliability. The proposed LeViT framework consistently captured fine-grained patterns in traditional dance, confirming its suitability for domains where interpretability and efficiency are equally important. This study is limited to static image-based recognition, which does not capture the temporal dynamics of dance movements. Since dance involves continuous sequences, rhythm, and transitions, relying solely on individual frames may overlook critical motion information. Future research can extend this work toward video-based approaches that integrate temporal modeling through 3D CNNs, recurrent or transformer-based sequence models, or pose trajectory analysis to more effectively represent the evolving structure of dance. Additionally, lightweight AI models optimized for real-time applications would enable deployment in mobile applications, virtual reality (VR), and interactive dance learning platforms, expanding the practical impact of AI in the field of dance recognition.