Modern approaches to improving image recognition efficiency with limited datasets

Abstract

In modern computer vision systems, the accuracy of deep learning models largely depends on the volume and diversity of training data. However, in many application areas, collecting large labeled datasets is a difficult, expensive, or sometimes even unattainable task. This necessitates the use of approaches that allow improving the results of models even under limited sample conditions. One of the most promising solutions is data augmentation, which involves creating additional training examples by transforming existing images. In this work, a practical experiment was conducted using the CIFAR-10 dataset, where only 10,000 examples out of 50,000 available were used to simulate resource-constrained conditions. A single convolutional neural network was used for training, and the results were compared between the model that was trained without any transformations and the model that used basic augmentation. The list of applied methods included horizontal mirroring, random cropping with the addition of fields, as well as changes in brightness, contrast and color saturation. The results obtained showed that the use of even basic augmentation techniques allows to significantly increase the model's resistance to variations in the input data. If the model without additional transformations demonstrated a tendency to overtraining and lower accuracy on the test sample, then the addition of augmentation gave a noticeable increase in the generalization ability indicators. In particular, the training graphs showed a decrease in the difference between the training and test accuracy, which indicates a more effective balance between adjusting to the data and the ability to work with new examples. An important difference of the conducted study is the emphasis on the conditions of limited samples, which makes it relevant for practical tasks where access to large volumes of labeled data is difficult. The results obtained not only confirm the effectiveness of classical augmentation, but also emphasize its potential as a basic tool that can be further combined with other methods, for example, semi-supervised learning or synthetic data generation. Thus, the work demonstrates not only the theoretical, but also the applied value of augmentation for improving the accuracy of computer vision models. This study is a starting point for further research into the impact of augmentation on neural networks in image recognition tasks.