Radon single-pixel flying target classification via texture-fused lightweight differentiable operators Authors: Hubin Ling, Bingzhang Hu, Dongfeng Shi, Yinbo Huang, Yingjian Wang Published: 2025 ABSTRACT Due to its rapid sampling speed and long-range imaging capabilities, Radon single-pixel imaging (SPI) has unique advantages in birds monitoring and counter-drone applications, such as airport security protection. Additionally, the Radon SPI algorithm can further enhance imaging speed by reducing the sampling rate, enabling better tracking and capturing of fast-moving aerial objects. However, low sampling rates also significantly degrade imaging quality, posing challenges for content recognition. For computational simplicity, most current SPI classification algorithms rely on shallow and small networks with limited discriminative power. However, with the rapid development of GPU hardware and deep learning techniques, using advanced deep models for ultra-low-sampling-rate Radon SPI classification has become a promising solution. This paper builds upon state-of-the-art (SOTA) lightweight classification models and fully leverage prior knowledge of Radon SPI characteristics. By integrating traditional texture operators and line-filtering operators into differentiable modules, an efficient classification model is specifically designed and optimized for Radon SPI. Experiments on a built Radon SPI flying target classification dataset demonstrate that the proposed model achieves the highest Top-1 accuracy compared to SOTA lightweight classification models. The source code will be made public. Introduction Single-pixel imaging (SPI) enables the use of specialized sensors (e.g., infrared, terahertz, and other non-visible spectral bands) to achieve imaging in scenarios where conventional visible-light imaging is infeasible1. Compared with traditional array-based imaging, single-pixel imaging reduces sensor array costs while exhibiting superior adaptability to extreme conditions and robustness against interference2. In recent years, single-pixel imaging has been developed rapidly3–5, demonstrating significant application value in deep-space exploration, medical imaging, industrial inspection, and long-range imaging. Common single-pixel imaging approaches include computational ghost imaging6, compressed sensing SPI7, Fourier SPI8, and Hadamard SPI9–11. Shi et al.12 proposed Radon single-pixel imaging, which employs the Radon transform to achieve rapid imaging of moving objects. With the explosive growth of the unmanned aerial vehicle (UAV) industry, drones, birds, and other flying targets have posed severe security threats to critical infrastructures such as airports. Radon SPI can further enhance real-time performance by reducing the sampling rate13, making it highly promising for flying target imaging. However, lower sampling rates inevitably degrade imaging quality and introduce artifacts. Therefore, target recognition under extremely low sampling rates in Radon single-pixel imaging holds substantial research significance for related applications. Currently, research on SPI primarily focuses on the imaging task, leveraging deep neural networks for low-sampling-rate reconstruction14, such as deep auto-encoder15, attention mechanism16, and deep generation model17. In the field of SPI target recognition, existing studies predominantly concentrate on the imaging module, while the classification module often employs simple convolution neural network (CNN)18–20, shallow UNet21 or multi-layer perceptron (MLP)22–26 architectures evaluated on small-scale synthetic datasets such as modified national institute of standards and technology database (MNIST)19–27 . These state-of-the-art (SOTA) SPI classification algorithms strike a reasonable balance between computational complexity and classification accuracy. However, constrained by the learning and generalization capabilities of early deep models, their classification performance under extremely low sampling rates remains suboptimal. Recent advancements in GPU technology and lightweight recognition models have provided powerful computational device and architectural support for SPI target recognition at low/ultra-low sampling rates. Therefore, this paper focuses on designing an effective recognition model to further improve the classification accuracy of Radon SPI flying targets under such difficult conditions. The evolution of image classification networks has driven the design of vision backbone models28–30. Representative CNN-based architectures include LeNet, VGG31, GoogleNet32, ResNet33, DenseNet34 and so on. In recent years, with the intersection of language and vision tasks, Transformer and Mamba architectures have also successfully applied in the field of image classification, such as ViT35, SwinT36, and Vision-Mamba37. Furthermore, a series of SOTA lightweight models, such as MobileNet38,39, EfficientNet40,41, DeiT42, and MambaOut43, have been proposed, achieving real-time classification through lightweight structural design, pruning, distillation, and quantization. However, these advanced classification models are primarily designed for natural images captured by visible-light array cameras, lacking tailored optimization and validation for single-pixel imaging. Notably, the absence of Radon SPI classification models also serves as an important motivation for this paper. Under low sampling rates, Radon single-pixel imaging exhibits significant difference from natural images, resembling coarse texture patterns instead. Based on this observation, this paper draw inspiration from traditional pattern recognition operators and propose differentiable units based on the texture operator local binary pattern (LBP)44 and linear gradient operators. Building upon advanced lightweight classification frameworks (e.g., MambaOut43), a dual-space fusion encoder is designed that jointly processes image and texture features to generate discriminative tokens. In addition, a lightweight differentiable linear-operator-attention unit is proposed within the basic computational blocks. The proposed method is validated on a dedicated Radon single-pixel flying target dataset. The main contributions of this paper are summarized as follows: This paper proposes a classification model specifically for Radon single-pixel flying targets, offering potential applications in aviation security and aerial surveillance. Unlike existing SPI classification algorithms that typically employ shallow CNNs or MLPs, this paper systematically evaluates and adapts the SOTA lightweight classification models.Leveraging the unique characteristics of low-sampling-rate Radon SPI, this paper first designs deep differentiable units based on traditional local operators. Then, an image-texture dual-space fusion encoder and a linear-operator-attention gated-block are proposed to extract Radon SPI features by incorporating domain-specific priors.A Radon SPI flying target dataset is constructed with three categories (i.e., aircraft, bird, and UAV), with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$61,000+$$\end{document} training samples and 975 test samples. As illustrated in Fig. 1, the proposed method achieves the highest Top-1 classification accuracy while maintaining an average inference speed of 5.5ms per image (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$224\times 224$$\end{document}) on RTX4090 GPU.Figure 1Top-1 accuracy (%) of different methods on Radon SPI flying target classification dataset. The proposed method outperforms some SOTA lightweight models. In addition, some Radon SPI examples of different sampling rates are provided. This paper proposes a classification model specifically for Radon single-pixel flying targets, offering potential applications in aviation security and aerial surveillance. Unlike existing SPI classification algorithms that typically employ shallow CNNs or MLPs, this paper systematically evaluates and adapts the SOTA lightweight classification models. This paper proposes a classification model specifically for Radon single-pixel flying targets, offering potential applications in aviation security and aerial surveillance. Unlike existing SPI classification algorithms that typically employ shallow CNNs or MLPs, this paper systematically evaluates and adapts the SOTA lightweight classification models. Leveraging the unique characteristics of low-sampling-rate Radon SPI, this paper first designs deep differentiable units based on traditional local operators. Then, an image-texture dual-space fusion encoder and a linear-operator-attention gated-block are proposed to extract Radon SPI features by incorporating domain-specific priors. Leveraging the unique characteristics of low-sampling-rate Radon SPI, this paper first designs deep differentiable units based on traditional local operators. Then, an image-texture dual-space fusion encoder and a linear-operator-attention gated-block are proposed to extract Radon SPI features by incorporating domain-specific priors. A Radon SPI flying target dataset is constructed with three categories (i.e., aircraft, bird, and UAV), with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$61,000+$$\end{document} training samples and 975 test samples. As illustrated in Fig. 1, the proposed method achieves the highest Top-1 classification accuracy while maintaining an average inference speed of 5.5ms per image (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$224\times 224$$\end{document}) on RTX4090 GPU. A Radon SPI flying target dataset is constructed with three categories (i.e., aircraft, bird, and UAV), with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$61,000+$$\end{document} training samples and 975 test samples. As illustrated in Fig. 1, the proposed method achieves the highest Top-1 classification accuracy while maintaining an average inference speed of 5.5ms per image (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$224\times 224$$\end{document}) on RTX4090 GPU. Top-1 accuracy (%) of different methods on Radon SPI flying target classification dataset. The proposed method outperforms some SOTA lightweight models. In addition, some Radon SPI examples of different sampling rates are provided. Related work Single-pixel image classification As mentioned before, research on single-pixel image classification mostly relies on simple task with small simulated datasets, such as MNIST19–27, Fashion-MNIST45 and CIFAR-10 dataset19. Zhu et al.46 employed traditional machine learning classifiers like support vector machine (SVM) to achieve image-free classification with short-duration SPI. With the development of deep neural networks, multilayer perceptrons (MLPs) have often been used for SPI classification22–26. To minimize computational cost, these MLPs typically consist of only a few fully connected (FC) layers (e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2-4$$\end{document} layers), limiting their learning capacity for complex real-world tasks. Convolutional neural networks (CNNs) have achieved remarkable success in visible-light image classification, and thus early CNN models like LeNet have also been applied to SPI classification19,20 or dual-optical classification18. Recently, UNet architecture is also applied for SPI classification21,27, and number of layers is also slightly increased to achieve better performance. Additionally, some recent work still focuses on designing efficient imaging computations to achieve rapid localization and classification of moving objects47. Overall, current research on SPI classification remains insufficient, with limited exploration and application of SOTA deep models. For Radon SPI as a specific category, related classification techniques are still lacking. Lightweight visible-light image classification As the accuracy of deep models continues to improve, the number of learnable parameters and the scale of networks have also expanded. To meet the computational constraints in practical tasks, many advanced classification algorithms have focused on lightweight and low-computation improvements. For example, MobileNet38 introduced depthwise separable convolutions to minimize computational cost, while ShuffleNet48 utilizes channel shuffling and group convolutions to enhance efficiency. Subsequent work has further concentrated on designing lightweight network modules or employing neural architecture search (NAS)49 to discover efficient architectures, such as SqueezeNet50, GhostNet51, MobileNetV239, EfficientNetV241, PatchConvNet52, InceptionNext53, MogaNet54, and ResMLP55. Recently, some studies have focused on lightweight designs or knowledge distillation56–58 for Transformer-based architectures. For instance, the DeiT method42 introduced a distillation token parallel to the class token in the input sequence, achieving efficient image classification through distillation and small-scale training datasets. Pei et al.59 further improved the computational efficiency of vision Mamba. In MambaOut43, Hao et al. found that removing the state space model (SSM) from the Mamba architecture surprisingly enhanced performance in lightweight recognition scenarios. These SOTA lightweight classification models for visible-light images provide an excellent foundation for SPI classification. With the rapid development of computational hardware such as GPUs, designing high-precision SPI classification methods based on advanced models has become a valuable research direction. Method Overall framework The architecture of the proposed Radon SPI classification model is illustrated in Fig. 2. For an input Radon single-pixel image, a dual-space fusion encoder (DSFE) is first designed to extract and fuse image and texture features. Subsequently, the dual-domain image features undergo feature extraction and processing through a series of stacked linear-operator-attention gated blocks (LOA-Blocks). Within these stacked blocks, four downsampling modules, which composed of a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} convolution with stride 2 and a layer normalization (LN), are employed to progressively reduce feature resolution. Finally, a classification head is used to generate final classification predictions. From an architectural perspective, the overall architecture inherits mature designs from recent SOTA lightweight classification models such as MambaOut43 and DeiT42,56. The following sections will provide detailed descriptions of each proposed module.Figure 2Overall framework of the proposed lightweight Radon SPI classification network.Figure 3Differentiable design of traditional texture and line-gradient operators. (a) structure of DeepLBP unit, (b) structure of DeepLine unit. Overall framework of the proposed lightweight Radon SPI classification network. Differentiable design of traditional texture and line-gradient operators. (a) structure of DeepLBP unit, (b) structure of DeepLine unit. Deep differentiable design of traditional local pattern operators As illustrated in Fig. 1, low-sampling-rate Radon single-pixel images exhibits significant differences in visual features compared to natural images. Under extreme under-sampling conditions, the images approximate coarse texture patterns, making it difficult to subjective recognition. Therefore, incorporating texture priors can effectively reduce the learning difficulty for the deep models. The local binary pattern (LBP)44 is a classical operator for traditional texture recognition. However, directly applying LBP to deep networks faces the issue of non-differentiability. Inspired by binarized CNN60 and texture restoration models61, this work reformulates LBP as a differentiable texture feature processing unit, namely deepLBP, from the perspective of differential convolution, as shown in Fig. 3a. To minimize computational overhead, DeepLBP unit employs depthwise convolution. For a single-channel input feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{X}$$\end{document}, eight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} differential kernels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{w}_p^{3\times 3} (i=0,1,...,7)$$\end{document} with a center value of 1 and a neighbor value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-1$$\end{document} are used to perform separate convolutions, simulating the numerical comparison operation of traditional LBP. The texture encoding value is then computed via weighted summation with predefined coefficients. Finally, the texture feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{Y}_{LBP}$$\end{document} is obtained through instance normalization. The computation process is designed as follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \varvec{Y}_{LBP} = \text {norm}\left( \sum _{p=0}^{P-1} \frac{2^p}{2^P - 1} (\varvec{w}_p^{3 \times 3} * \varvec{X}) \right) , \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {norm}(\cdot )$$\end{document} denotes normalization operation, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P=8$$\end{document} represents the total number of LBP neighboring pixels in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} local region, which is also the number of differential convolution kernels. Similarly, we can observe that low-sampling-rate Radon SPI exhibits distinct radial patterns with linear characteristics, making it suitable for feature processing using linear filters. To fully leverage this prior, this paper also constructs a differentiable linear filtering operator unit, namely DeepLine, as shown in Fig. 3b. To further reduce the computational cost of this unit, only four linear convolution kernels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{w}_i^{line} (i=0,1,2,3)$$\end{document} are employed for the primary directions (i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0^\circ , 45^\circ , 90^\circ , and 135^\circ$$\end{document}). For a single-channel feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{X}$$\end{document}, the DeepLine linear texture feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{Y}_{Line}$$\end{document} is calculated as follows:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \varvec{Y}_{Line} = \text {norm}\left( \sum _{i=0}^{K-1} \frac{2^i}{2^K - 1} (\varvec{w}_i^{line} * \varvec{X}) \right) , \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K=4$$\end{document} represents the total number of linear convolution kernels.Figure 4Structure of proposed encoder module and basic blocks. (a) Spatial-texture dual-space fusion encoder, (b) linear-operator-attention gated-block. The number of channels is also noted in the figure. Structure of proposed encoder module and basic blocks. (a) Spatial-texture dual-space fusion encoder, (b) linear-operator-attention gated-block. The number of channels is also noted in the figure. Dual-space fusion encoder for image and texture features For natural image classification tasks, deep models typically employ convolution layers or encoding modules to transform the original image into a latent feature domain. Based on the designed DeepLBP unit, this paper proposes a dual-space fusion encoder module for image and texture features, as illustrated in Fig. 4a. For the input image, image features are firstly extracted and the resolution is halved using a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} convolution with a stride of 2, a layer normalization (LN) and a GELU activation. Subsequently, the corresponding texture features are extracted using the DeepLBP unit, effectively leveraging the texture prior of low-sampling-rate Radon SPI. Notably, after concatenating the image-domain features and texture-domain features, a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times 1$$\end{document} convolution is used to compresses the number of channels by half. In other words, this module fuses and refines the features from dual domains without expanding the output channels, thereby avoiding additional computational cost in subsequent modules. Finally, the fused features are obtained as tokens for further computation through another 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}3 convolution with a stride of 2 and a normalization. Linear-operator-attention gated-block The stacked linear-operator-attention gated-blocks are the primary feature processing module in the proposed classification model. In Transformer-based architectures, various forms of self-attention are commonly used for global attention perception before feature processing, but they incur high computational costs. In Mamba-based models, the state-space-model (SSM) modules are employed for global information perception. Interestingly, in recent lightweight model MambaOut43, removing the SSM module and retaining only the feedforward unit has been shown to achieve better discriminative performance. Inspired by recent SOTA deep models and the linear characteristics in Radon single-pixel imaging, this paper designs a lightweight linear texture attention unit based on the proposed DeepLine. The attention map is computed using a Deepline and then multiplied with the original features. The detailed structure of this block is illustrated in Fig. 4b. The input feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{F}_{in}$$\end{document} is first passed through a fully-connected (FC) layer to expand the channel dimension. The feature is then split along the channel dimension into three components, i.e., the gating feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{F}_{Gate}$$\end{document}, the identity feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{F}_{Identity}$$\end{document}, and the convolutional feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{F}_{Conv}$$\end{document}. To reduce computational cost, only the convolutional feature undergoes processing by the linear operator attention (LOA) unit followed by a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} convolution. The convolution result is concatenated with the identity feature and then multiplied by the gating feature. Finally, another FC layer is applied to produce the output \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{F}_{out}$$\end{document}, which matches the input channel dimension. Additionally, the block incorporates an end-to-end residual connection. The overall computation of the block is expressed as follows:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \varvec{F}_{out} = \text {FC}\left( \varvec{F}_{Gate} \odot \left[ \varvec{F}_{Identity}, \text {Conv}\left( \text {LOA}(\varvec{F}_{Conv})\right) \right] \right) + \varvec{F}_{in}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\odot$$\end{document} element-wise multiplication, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Conv(\cdot )$$\end{document} denotes the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} convolution operation, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$LOA(\cdot )$$\end{document} represents the linear operator attention with DeepLine unit. In this paper, the DeepLBP operator is computed only once during the texture feature extraction and encoding process, so we designed 8-directional filtering calculations following the traditional LBP operator. However, the DeepLine operator is part of the basic LOA-Block and will be stacked and repeatedly computed in the network. Therefore, to reduce computational overhead, DeepLine is designed with as few directions as possible while still covering most linear orientations—selecting 4 directions for a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} kernel. Increasing the number of directions would require enlarging the kernel size, significantly increasing the convolutional computation cost. Additionally, the number of kernels would also increase, leading to higher memory and computation cost. Reducing the number of directions may reduce the accuracy of directional description, but it also enhances robustness to line orientations. This approach of quantization to achieve high robustness is inspired by the design of traditional operators like LBP, which uses extreme binary quantization to achieve high robustness against grayscale variations. The results of the following ablation study also show that the 4-direction linear kernels perform even better than 8-direction kernels. Other details The secondary modules include downsampling and the classification head, which adopt the commonly used structures as in other classification models42,43. As shown in former Fig. 2, the downsampling module consists of a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} convolution with a stride of 2 and layer normalization. The classification head sequentially includes an FC layer, a GELU activation function, layer normalization, and a second FC layer that outputs the final prediction. Note that the proposed method focuses on Radon SPI classification of flying objects, including aircraft, bird, and UAV, so the second FC layer has an output dimension of 3. The proposed model is trained in a fully supervised end-to-end manner, using the most commonly adopted classification loss, consistent with the approaches42,43. Radon SPI and dataset Radon single-pixel imaging The Radon SPI computation in this paper follows the approach described in method12. Its core idea involves acquiring projection information of the target object at different angles through the Radon transform and then reconstructing the image of the object using filtered back projection (FBP). The main steps are as follows. Projection Pattern Generation Phase. In Radon SPI, the r-th row vector of the Hadamard matrix is back-projected to generate a 2D projection pattern:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p_r(x, y, c, \theta ) = H_r(c) \delta (x \cos \theta + y \sin \theta - c) , \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_r(c)$$\end{document} denotes the value at the r-th row and c-th column of the Hadamard matrix, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(x \cos \theta + y \sin \theta - c)$$\end{document} represents the position of the linear stripe, c is the distance from the line \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(x \cos \theta + y \sin \theta = c)$$\end{document} to the origin, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta (t)$$\end{document} denotes the Dirac delta function (1 when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t = 0$$\end{document}, 0 otherwise). Subsequently, the 2D projection pattern is split into positive (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_r^+$$\end{document}) and negative (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_r^-$$\end{document}) components:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{aligned} p_r^+(x, y, c, \theta ) = \frac{1}{2} \left( 1 + p_r(x, y, c, \theta )\right) , \\ p_r^-(x, y, c, \theta ) = \frac{1}{2} \left( 1 - p_r(x, y, c, \theta )\right) , \end{aligned} \end{aligned}$$\end{document}Projection and Modulation Phase. The generated positive and negative projection patterns are sequentially projected onto the target object. The modulated optical signals are then captured by a single-pixel sensor and aggregated into total intensity values (\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} and \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}). The differential measurement \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I(r, \theta )$$\end{document} is obtained by subtracting the negative-pattern intensity from the positive-pattern intensity:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} I(r, \theta ) = I^+(r, \theta ) - I^-(r, \theta ) = \sum _{x, y, c} H_r(c) f(x, y) \delta (x \cos \theta + y \sin \theta - c) = \sum _c H_r(c) F(c, \theta ) , \end{aligned}$$\end{document}where f(x, y) represents the 2D discrete distribution function of the target object, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F(c, \theta )$$\end{document} denotes the Radon transform value of the object at 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}. Radon Value Calculation Phase. Leveraging the orthogonality and completeness of the Hadamard matrix, the full projection data for a given angle in the Radon transform domain is computed using the differential measurements corresponding to the Hadamard patterns:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} F(c, \theta ) = \sum _r h_r(c) I(r, \theta ) . \end{aligned}$$\end{document}By repeating this process at multiple angles (ranging from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0^\circ$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$180^\circ$$\end{document}), a sequence of Radon values is obtained for each angle and arranged into a 2D data matrix, forming a sinogram. Finally, the sinogram is reconstructed using the filtered back projection algorithm to obtain the reconstructed image, as follows:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} f(x, y) \approx \frac{1}{2} \mathcal {B} \left( \mathcal {F}_t^{-1} S * F(c, \theta ) \right) (x, y) , \end{aligned}$$\end{document}where S denotes a low-pass filter (i.e., the Ram-Lak filter) in the frequency domain, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {F}_t^{-1}$$\end{document} represents the inverse Fourier transform, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$*$$\end{document} denotes the convolution operation, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {B}$$\end{document} denotes the back-projection operation. More computational details of Radon SPI can be found in literature12. The artifacts in low-sampling-rate Radon SPI are mainly caused by insufficient sampling density in the Radon domain, leading to information aliasing. This results in distortions such as linear artifacts and imaging blurring. Due to the low sampling rate, critical frequency components in the Radon domain may be skipped, making it impossible to reconstruct fine details during the imaging process, thereby causing noticeable information loss. In terms of classification, these artifacts can be misinterpreted as inherent linear features of the target, potentially disrupting the continuity of genuine features. This leads to the classifier learning pseudo features, which poses significant challenges to the generalization capability of the learned classifier.Figure 5Examples with different sampling rates in the RaSPIFly training set. Examples with different sampling rates in the RaSPIFly training set. Radon SPI flying target classification dataset (RaSPIFly) To establish the Radon SPI flying target classification dataset (RaSPIFly), this paper collected 5,000 images each of three types of flying objects, i.e., aircraft, bird, and UAV, from the internet. These images feature sky backgrounds, and the target regions were cropped using object detection to center the flying objects. Note that, in real-world application scenarios, dynamic target tracking is typically employed to rapidly locate flying objects and then perform single-pixel imaging centered on the target62. Therefore, the detection and cropping preprocessing better simulates the imaging results of Radon SPI. Specifically, a pre-trained YOLOv8 model63 is utilized to detect and extract target regions from the original training images. To facilitate comparisons with state-of-the-art (SOTA) classification models, all images were uniformly resized to the commonly used resolution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$224\times 224$$\end{document} as in ImageNet. After data cleaning and manual removal of flawed samples, the RaSPIFly training set consists of 4, 466 aircraft images, 4, 645 bird images, and 3, 200 drone images, totaling 12, 311 training images. Subsequently, aforementioned Radon SPI was applied to simulate low-sampling-rate degradation, generating images at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\%, 1.5\%, 2\%$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\%$$\end{document} sampling rates, along with fully sampled (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$100\%$$\end{document}) images. Figure 5 showcases some training examples and their degraded versions under different sampling rates. It is evident that subjective classification becomes challenging at extremely low sampling rates. Through FBP-based degradation, the training set comprises 61, 555 images across five sampling levels. To thoroughly evaluate the generalization capability of the trained deep models, the test set consists of 195 manually selected images (65 per class) featuring diverse shooting angles, ensuring variability and differentiation from the training scenarios. Some test images are displayed in Fig. 6. To further enhance the distinction from the training set, different manual cropping operation was applied, and the resolution was standardized to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$224\times 224$$\end{document}. The test set covers the same five sampling levels, resulting in 975 test images in total.Figure 6Examples of the RaSPIFly test set. The test set includes images with significant variations to validate the generalization capability of trained models. Examples of the RaSPIFly test set. The test set includes images with significant variations to validate the generalization capability of trained models. Results Implementational details The experimental implementation follows DeiT42 and MambaOut43. Due to the limited scale of the SPI dataset, knowledge distillation was not employed in the experiments. For data augmentation, since the primary application scenario involves aerial targets such as airplanes and drones, which rarely appear upside down, horizontal flipping augmentation was adopted. All models utilized the AdamW optimizer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\beta _1=0.9,\beta _2=0.999)$$\end{document}, with an initial learning rate set to 0.001 (Cosine learning rate decay) and a batch size of 32. All models were trained on the same dataset for 300 epochs, which were set according to literature43. The experiments were conducted on an RTX 4090 GPU using the PyTorch platform. Experimental results To validate the effectiveness of the proposed algorithm, we compared it with several SOTA lightweight classification models, including MobileNetV239, EfficientNet40, EfficientNetV241, ResMLP55, PatchConvNet52, InceptionNext53, MogaNet54, DeiT42, CaiT57, and MambaOut43. For a fair comparison, these algorithms are all trained and tested on the same Radon SPI flying target classification dataset. Experimental results of each algorithm are listed in Table 1, from which the following observations can be made. First, subjective visual recognizability differs from model classification performance. Under extreme conditions (e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\%$$\end{document} sampling rate), the imaging results are no longer subjectively recognizable, yet some advanced classification models can achieve over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$70\%$$\end{document} Top-1 accuracy, while the proposed method even reaches \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$79.48\%$$\end{document}. Second, SOTA ResMLP55 and Transformer-based models of DeiT-T42 and CaiT-S57 underperform compared to convolutional architectures on the Radon SPI dataset. This may be because Radon SPI focuses on imaging flying targets with relatively simple content and clean backgrounds, and the dataset size is also limited. Some models optimized for complex natural image classification may not perform optimally in this scenario. Third, among the compared algorithms, EfficientNetV2-S41 and MambaOut-femto43 achieve relatively better performance. Note that MambaOut-femto has much fewer parameters than EfficientNetV2-S, demonstrating the superiority of the MambaOut architecture. However, performance of MambaOut-femto declines abnormally at higher sampling rates, suggesting that removing attention-based modules may weaken intrinsic feature learning in certain cases. Lastly, the proposed algorithm achieves the highest recognition accuracy across all sampling rates, fully demonstrating its superior performance in Radon SPI image classification task. The parameter counts, FLOPs, and average single-image inference times of each model are also listed in Table 1. The results show that current lightweight advanced classification models can achieve strong learning capabilities with relatively low parameter counts and FLOPs. Notably, the inference time for each Radon SPI sample is mostly under 10 ms (on an RTX 4090 GPU), indicating that with the advancement and widespread adoption of GPU devices, using SOTA classification models for high-speed Radon SPI classification has become a promising solution. Furthermore, with the widespread application of deep learning algorithms, the computational power of edge computing cards, such as Nvidia Jetson Nano, MediaTek APU, and Google Edge TPU, has been continuously improving. This advancement also provides the possibility for future model deployment and inference on embedded devices.Table 1Classification Top-1 accuracy on RaSPIFly test set.ModelParam.Time (ms)FLOPsAcc. 1%Acc. 1.5%Acc. 2%Acc. 5%Acc. 100%MobileNetV2 392.2M8.90.31G71.2871.7978.9775.3878.94EfficientNet-S 4019.6M11.81.74G73.3375.3877.9482.5684.61EfficientNetV2-S 4120.3M10.62.89G74.3676.9279.4883.5987.69ResMLP 5516.7M4.63.01G64.6167.6969.7473.3374.35PatchConvNet-S 5225.2M6.93.99G72.3172.8275.3878.9778.97InceptionNext 533.2M5.60.50G69.2370.7779.4880.0073.33MogaNet-T 544.9M10.81.09G68.2074.3680.5178.9783.07DeiT-T 425.5M4.41.26G67.6967.1772.3076.4180.00CaiT-S 5746.5M6.39.34G68.2168.7270.2574.3577.95MambaOut-femto 437.3M4.81.15G74.3678.9480.5183.5983.07Ours7.3M5.51.21G79.4882.5683.0784.6197.43 Table 1Classification Top-1 accuracy on RaSPIFly test set.ModelParam.Time (ms)FLOPsAcc. 1%Acc. 1.5%Acc. 2%Acc. 5%Acc. 100%MobileNetV2 392.2M8.90.31G71.2871.7978.9775.3878.94EfficientNet-S 4019.6M11.81.74G73.3375.3877.9482.5684.61EfficientNetV2-S 4120.3M10.62.89G74.3676.9279.4883.5987.69ResMLP 5516.7M4.63.01G64.6167.6969.7473.3374.35PatchConvNet-S 5225.2M6.93.99G72.3172.8275.3878.9778.97InceptionNext 533.2M5.60.50G69.2370.7779.4880.0073.33MogaNet-T 544.9M10.81.09G68.2074.3680.5178.9783.07DeiT-T 425.5M4.41.26G67.6967.1772.3076.4180.00CaiT-S 5746.5M6.39.34G68.2168.7270.2574.3577.95MambaOut-femto 437.3M4.81.15G74.3678.9480.5183.5983.07Ours7.3M5.51.21G79.4882.5683.0784.6197.43 Classification Top-1 accuracy on RaSPIFly test set. Extended experiments In order to further verify the generalization of these models, a AIGC test set is added as extended test. We generated 990 test images (330 per class) using a AIGC model (Stable Diffusion V3). Example images are shown in Fig. 7a. After low-sampling-rate degradation, this AIGC test set totally contains 4950 samples. Since these images are generated, the trained models have never seen them. The trained models are directly tested on this AIGC test set without any tuning stage. Experimental results of different models are presented in Table 2. Due to the high quality, clarity, and large scale of the AIGC-generated images, all trained models achieved much better classification performance on this dataset. Moreover, the proposed method still outperforms others across all sampling rates. These results not only confirm the generalization capability of the trained algorithms (achieving high performance on unseen synthetic data) but also validate that the original RaSPIFly test set construction was indeed more challenging and suitable for assessing generalization differences.Figure 7Examples of the extended test sets, (a) generated test samples in AIGC test set, (b) noisy samples with different noise levels in noisy test set.Table 2Classification Top-1 accuracy on AIGC test set.Acc.1%Acc.1.5%Acc.2%Acc.5%Acc.100%MobileNetV2 3992.2294.3490.3087.5797.47EfficientNet-S 4095.4596.9797.2795.0598.68EfficientNetV2-S 4196.7097.4796.7696.8698.48ResMLP 5591.8193.3390.4083.1380.40PatchConvNet-S 5293.2395.1594.3491.4194.74InceptionNext 5390.5094.8488.8884.8493.33MogaNet-T 5493.7395.3594.3485.3597.87DeiT-T 4292.1288.7889.3980.6094.94CaiT-S 5789.3988.3882.8279.0992.12MambaOut-femto 4396.9697.3797.4798.3893.03Ours98.0898.8898.8898.98100.00 Examples of the extended test sets, (a) generated test samples in AIGC test set, (b) noisy samples with different noise levels in noisy test set. Table 2Classification Top-1 accuracy on AIGC test set.Acc.1%Acc.1.5%Acc.2%Acc.5%Acc.100%MobileNetV2 3992.2294.3490.3087.5797.47EfficientNet-S 4095.4596.9797.2795.0598.68EfficientNetV2-S 4196.7097.4796.7696.8698.48ResMLP 5591.8193.3390.4083.1380.40PatchConvNet-S 5293.2395.1594.3491.4194.74InceptionNext 5390.5094.8488.8884.8493.33MogaNet-T 5493.7395.3594.3485.3597.87DeiT-T 4292.1288.7889.3980.6094.94CaiT-S 5789.3988.3882.8279.0992.12MambaOut-femto 4396.9697.3797.4798.3893.03Ours98.0898.8898.8898.98100.00 Classification Top-1 accuracy on AIGC test set. In addition, experiments are conducted with noisy test sets to evaluate the robustness of trained models. We applied random noise augmentation with random levels, including Gaussian noise (randomly sampled with intensity between 0.1 and 0.2) and salt-and-pepper noise (randomly sampled with scale between 0.01 and 0.2). Each original image was used to generate 5 noisy variants, resulting in a total of 975 noisy test images (325 per class). After low-sampling-rate degradation, this noisy test set contain 4875 samples in total. Some noisy examples are shown in Fig. 7b, and the results of different models (without fine-tuning) are presented in Table 3. The results indicate that due to the low sampling rate and strong noise interference, the classification performance of all models significantly declined. However, the proposed model still achieved better classification accuracy and better robustness than these compared SOTA models. Additionally, comparing to completed sampling, many models exhibit a slight improvement in noise resistance when the sampling rate is reduced to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\%$$\end{document} in noisy environments.Table 3Classification Top-1 accuracy on noisy test set.Acc.1%Acc.1.5%Acc.2%Acc.5%Acc.100%MobileNetV2 3954.5656.4161.3362.5660.51EfficientNet-S 4054.1557.5357.8466.1566.15EfficientNetV2-S 4156.5159.6964.8272.0070.46ResMLP 5550.3552.5154.1566.7969.23PatchConvNet-S 5258.1554.7663.3859.4859.79InceptionNext 5353.9453.5367.1763.5962.15MogaNet-T 5453.5356.3062.1559.7960.20DeiT-T 4250.5652.4154.9763.8951.89CaiT-S 5751.5953.7458.2562.3555.59MambaOut-femto 4356.4161.8465.2366.2562.46Ours59.6963.2868.3075.6973.23 Table 3Classification Top-1 accuracy on noisy test set.Acc.1%Acc.1.5%Acc.2%Acc.5%Acc.100%MobileNetV2 3954.5656.4161.3362.5660.51EfficientNet-S 4054.1557.5357.8466.1566.15EfficientNetV2-S 4156.5159.6964.8272.0070.46ResMLP 5550.3552.5154.1566.7969.23PatchConvNet-S 5258.1554.7663.3859.4859.79InceptionNext 5353.9453.5367.1763.5962.15MogaNet-T 5453.5356.3062.1559.7960.20DeiT-T 4250.5652.4154.9763.8951.89CaiT-S 5751.5953.7458.2562.3555.59MambaOut-femto 4356.4161.8465.2366.2562.46Ours59.6963.2868.3075.6973.23 Classification Top-1 accuracy on noisy test set. Ablation test The ablation test results of the proposed modules are listed in Table 4. Here, Backbone represents the basic architecture without the proposed dual-space fusion encoder (DSFE) and the linear-operator-attention (LOA) unit in the LOA gated-blocks. Note that w/o DSFE means removing the DeepLBP unit and only using a regular input layer, while w/o LOA means removing the DeepLine-based LOA from the basic Blocks. From the comparison of the results in the first three rows, it can be seen that the differentiable texture feature fusion through the DSFE module significantly improves recognition accuracy. After introducing LOA into the basic blocks, the discriminative performance of the model is further enhanced. These results validate the effectiveness of the DSFE and LOA modules designed for Radon SPI classification. To further verify the efficiency of the LOA designed based on linear sampling priors, Table 4 also lists the results of different attention units in the basic block. For instance, we replaced LOA with the lightweight spatial-attention (LSA) commonly used in convolutional models64, the lightweight linear attention (LLA) often seen in Transformer models65, and replaced the DeepLine unit in LOA with the DeepLBP unit. It can be observed that the proposed LOA significantly outperforms traditional LSA and LLA, making it more suitable for extracting discriminative information from Radon SPI data. Additionally, although the DeepLBP utilizes more differential kernels, the LOA unit using DeepLine incurs lower computational cost while delivering slightly better performance. Figure 8 further visualizes the processing effects of the proposed DeepLBP and DeepLine units. It can be seen that these differentiable operators focus on linear textures and effectively highlight important edges and lines, making them highly suitable for edge feature perception of Radon SPI flying targets. Table 5 presents the performance of the proposed algorithm with different sizes. It can be seen that increasing the number of basic blocks from 9 to 18 effectively enhances discriminative capability. Note that the increase in parameters only improves network capacity, while performance is also related to dataset size and optimization processes. When the model size is further doubled, a slight improvement in discriminative ability is observed in most sampling scenarios, but performance even slightly declines in the extreme case of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\%$$\end{document}. By balancing the classification accuracy and computational cost, the proposed method adopts a moderate scale (i.e., 18 blocks). It should also be noted that if larger datasets can be collected in future work, increasing the number of learnable parameters under sufficient computational resources still holds the potential to further enhance the discriminative performance of the proposed method.Table 4Ablation test results of different module on RaSPIFly test set.MethodParam. (M)Acc. 1%Acc. 2%Acc. 5%Acc. 100%Backbone (w/o DSFE & LOA)7.374.8780.5183.0784.10Backbone+DSFE7.376.9282.0583.5993.33Backbone+DSFE+LOA7.379.4883.0784.6197.43LOA \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rightarrow$$\end{document}LSA7.374.3580.5183.0784.10LOA \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rightarrow$$\end{document}LLA8.075.3881.5383.5984.10LOA (DeepLine\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rightarrow$$\end{document}DeepLBP)7.376.4182.0584.1092.30Table 5Ablation test results of different model size on RaSPIFly test set.BlocksParam. (M)Acc. 1%Acc. 2%Acc. 5%Acc. 100%Size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 2364.677.9584.1085.1297.95Size / 2912.476.9282.0582.5691.79Ours187.379.4883.0784.6197.43 Table 4Ablation test results of different module on RaSPIFly test set.MethodParam. (M)Acc. 1%Acc. 2%Acc. 5%Acc. 100%Backbone (w/o DSFE & LOA)7.374.8780.5183.0784.10Backbone+DSFE7.376.9282.0583.5993.33Backbone+DSFE+LOA7.379.4883.0784.6197.43LOA \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rightarrow$$\end{document}LSA7.374.3580.5183.0784.10LOA \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rightarrow$$\end{document}LLA8.075.3881.5383.5984.10LOA (DeepLine\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rightarrow$$\end{document}DeepLBP)7.376.4182.0584.1092.30 Ablation test results of different module on RaSPIFly test set. Table 5Ablation test results of different model size on RaSPIFly test set.BlocksParam. (M)Acc. 1%Acc. 2%Acc. 5%Acc. 100%Size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 2364.677.9584.1085.1297.95Size / 2912.476.9282.0582.5691.79Ours187.379.4883.0784.6197.43 Ablation test results of different model size on RaSPIFly test set. Figure 8Visualization of the results of DeepLBP and DeepLine. (a) input images, (b) results of DeepLBP, (c) results of DeepLine. Note that these modules actually operate on hidden features, and these images are only used to visually demonstrate the processing effects of these modules. Visualization of the results of DeepLBP and DeepLine. (a) input images, (b) results of DeepLBP, (c) results of DeepLine. Note that these modules actually operate on hidden features, and these images are only used to visually demonstrate the processing effects of these modules. Weakness and future work Compared to current SPI algorithms that employ shallow models and small-scale synthetic datasets like MNIST, the proposed method demonstrates its superiority on a larger synthetic flying object dataset. However, this dataset remains synthetic set. Images acquired by real-world non-visible-light single-pixel sensors may differ from such simulated data. This paper validates the effectiveness of advanced models for low-sampling-rate Radon SPI classification. In future work, we will capture real data for model fine-tuning, enabling Radon SPI classification in practical scenarios. Exploring more application scenarios for Radon SPI represents another important future direction. As the range of applications expands, more classification types can be incorporated based on specific application scenarios. Notably, reducing the sampling rate further enhances the privacy-preserving capabilities of Radon SPI, while deep models can still perform classification effectively. This advantage also facilitates the new potential applications. In addition, at low sampling rates, the trained models face challenges from various potential noises present in real-world scenarios. Therefore, enhancing the model’s noise resistance and robustness is also a direction for future research. Conclusion This paper proposed a Radon single-pixel imaging (SPI) classification network based on texture-fused lightweight differentiable operators. Leveraging prior knowledge of texture and linear features in Radon SPI, this paper constructs differentiable operator units derived from traditional local texture operators and linear gradient operators. Subsequently, a feature extraction encoder is proposed based on spatial texture dual-feature space fusion using the differentiable texture units. Additionally, in the basic blocks of the network, differentiable linear operator unit is employed to calculate linear operator attention, guiding the network to focus on linear features and important edges. Unlike most existing SPI classification algorithms, which rely on small-scale simulated datasets and simple networks, this work trains and validates recent SOTA lightweight classification models on a larger Radon SPI flying target dataset. Experiments demonstrate that the proposed algorithm achieves superior performance in Radon SPI classification.