Enhancing Image Classification with Augmentation: Data Augmentation Techniques for Improved Image Classification

Saroj Kumar*, Ugochukwu Esiowu, Prince Asiamah, Oluwatoyin Jolaoso · 2025

Read Enhancing Image Classification with Augmentation: Data Augmentation Techniques for Improved Image Classification by Saroj Kumar*, Ugochukwu Esiowu, Prince Asiamah, Oluwatoyin Jolaoso (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-18691.txt.

Introduction

Kumar and co-authors published a computing paper in 2025 on data augmentation tricks to improve image classification. Image networks overfit when training sets are small. The team tested eleven augmentation sets, including three of their own design. One set copies color channels between images, one hides objects behind other objects, and one applies shaped masks.

Classic tricks such as rotation, flips, blur, and erasing were also tested. Each set fine-tuned the same base model, and results were compared on the Caltech-101 dataset. The combined ensemble of methods worked best. The authors present varied augmentation as a practical way to strengthen small datasets.

Catalog record

CREATORSaroj Kumar*, Ugochukwu Esiowu, Prince Asiamah, Oluwatoyin Jolaoso
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE32 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-18691

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Enhancing Image Classification with Augmentation: Data Augmentation Techniques for Improved Image Classification Saroj Kumar*, Ugochukwu Esiowu, Prince Asiamah, Oluwatoyin Jolaoso Abstract Convolutional Neural Networks (CNNs) serve as the workhorse of deep learning, finding applications in various fields that rely on images. Given sufficient data, they exhibit the capacity to learn a wide range of concepts across diverse settings. However, a notable limitation of CNNs is their susceptibility to overfitting when trained on small datasets. The augmentation of such datasets can significantly enhance CNN performance by introducing additional data points for learning. In this study, we explore the effectiveness of 11 different sets of data augmentation techniques, which include three novel sets proposed in this work. The first set of data augmentation employs pairwise channel transfer, transferring Red, Green, Blue, Hue, and Saturation values from randomly selected images in the database to all images in the dataset. The second set introduces a novel occlusion approach, where objects in the images are occluded by randomly selected objects from the dataset. The third set involves a novel masking approach, using vertical, horizontal, circular, and checkered masks to occlude portions of the images. In addition to these novel techniques, we investigate other existing augmentation methods, including rotation, horizontal and vertical flips, resizing, translation, blur, color jitter, and random erasing, and their effects on accuracy and overfitting. We fine-tune a base EfficientNet-B0 model for each augmentation method and conduct a comparative analysis to showcase their efficacy. For the evaluation and comparison of these augmentation techniques, we utilize the Caltech-101 dataset. The ensemble of image augmentation techniques proposed emerges as the most effective on the Caltech-101 dataset. The results demonstrate that diverse data augmentation techniques present a viable means of enhancing datasets for improved image classification. I Introduction Making machines learn requires data from which they can derive real-world representations. With recent advances in machine learning, which involve very large neural networks, training them has become a highly data-intensive task. This area of machine learning is known as Deep Learning, and it requires millions of data points. According to recent trends, the successful application of Deep Learning increasingly demands more data [ 17 ] . There are several methods to increase the size of a high-quality dataset; however, One of the simplest ways to enhance the quality of the dataset is through data augmentation. I-A Data augmentation Data augmentation is the technique to artificially increase the diversity in a training dataset by using various types of transformations. Data augmentation techniques are used to improve the model’s ability to generalize well to new and unseen data and thus improve its robustness. By using data augmentation techniques, the existing dataset becomes large and more diverse. This helps in avoiding the overfitting of the model and learning features that are invariant to such transformations. For visual datasets such as RGB images, there are several common data augmentation techniques such as rotation, flip, translation, color jitter, blur, random erasing, etc . As the name suggests, Rotation involves rotating the image by a certain angle which introduces rotation invariance in the model. Flip involves flipping the image horizontally or vertically which leads to orientation invariance in the model. Translation involves changing the position of the image horizontally or vertically which introduces location invariance in the model. In a similar fashion, color jitter, which involves changing the color of the image, causes lighting robustness in the model. Therefore, the model would be able to detect an object in an image even in difficult lighting conditions. Random erasing involves removing random patches of the image which introduces occlusion invariance in the model. All these image augmentation techniques lead to better performance of the model in terms of different metrics; however, if we can better learn the representation from the data with novel data augmentation techniques, it would lead to a significant improvement in different performance metrics. I-B Proposed Augmentation In this work we propose three novel sets of data augmentation techniques. First set of data augmentation is called Pairwise chanel transfer. Second set of augmentation technique are novel occlusion approach and the third set of augmentation technique are novel masking approach I-B 1 Pairwise chanel transfer The first set of data augmentation techniques involves pairwise channel transfer in images, where a channel from a random image in the dataset is transferred into other images in the dataset. This process facilitates the cross-pollination of information between images and introduces adulteration, which promotes irrelevant content invariance in the images. I-B 2 Novel occlusion Augmentation The second set of image augmentation techniques introduces a novel occlusion approach, where objects within an image are occluded by random images from the dataset. This augmentation is expected to result in a model that is better at handling occlusions. I-B 3 Novel masking augmentation The final and third set of image augmentation techniques involves a novel masking approach, where the images and their contents are masked by horizontal, vertical, checkered, and circular stripes. Such masking is expected to lead to better representation and help prevent overfitting of the model. In addition to these proposed augmentation techniques, we also investigate existing image augmentation methods, namely rotation, horizontal and vertical flips, resizing, translation, blur, color jitter, and random erasing. These techniques were tested on the Caltech-101 dataset [ 18 ] . I-C Dataset Used This dataset comprises 101 different categories of images, with each category containing between 40 and 800 images. Based on the proposed and existing augmentation techniques, we have developed three different variants of the dataset. The first variant is a vanilla dataset with no data augmentation, totaling 9,146 images. The second variant incorporates the original dataset enhanced by existing data augmentation techniques, resulting in a total of 54,864 images. The third variant includes the original dataset combined with newly proposed data augmentation techniques, leading to a total of 73,153 images. I-D Deep learning model used We train the base variant of the EfficientNet_B0 deep neural network [ 16 ] on these three different variants of the dataset to assess the performance of the proposed data augmentation techniques. Introduced by Tan et al., EfficientNet_B0 is a deep Convolutional Neural Network that is part of a study investigating the effects of width, depth, and resolution on the accuracy of neural networks. EfficientNet_B0, the smallest of the proposed models, has the least width, depth, and resolution, which results in the fewest number of parameters. Among all the proposed deep learning models, this variant has 4,138,210 trainable parameters. II Background Image augmentation and associated techniques have been foundational in the field of computer vision and machine learning. LeCun et al. [ 11 ] employed planar affine transformations, such as horizontal and vertical translations and scaling, to augment images for handwritten zip code recognition. This significantly improved test error rates in their seminal work, which introduced the Convolutional Neural Network (CNN) known as LeNet-5. Similarly, groundbreaking research known as AlexNet by Kirchofsky et al. [ 12 ] utilized techniques such as horizontal reflection and color jitter based on eigenvalues and eigenvectors for each channel, leading to a considerable increase in dataset size. These examples underscore the critical importance of image augmentation in training machine learning and deep learning models. Recent advances in computer vision

Read the complete text