APPLICATION AND ANALYSIS OF MACHINE LEARNING METHODS FOR IMAGE CLASSIFICATION

Abstract

Relevance. Image classification is a key task in computer vision, which has wide applications in medicine, transportation, industry, and security. The use of optimized CNN architectures allows high accuracy to be achieved with limited resources, which is relevant for mobile and embedded systems. Research subject: machine learning methods and neural network architectures for image classification. The purpose of the article is to develop and evaluate a modified convolutional neural network that provides a balance between classification accuracy and computational efficiency, as well as to compare its results with classical and modern models. Research results. The proposed CNN achieved 93.8% accuracy on the Fashion MNIST dataset, exceeding the performance of LeNet-5 (91.2%) and classical algorithms (KNN – 85.3%, Decision Tree – 82.7%, XGBoost – 90.4%). On the more complex CIFAR-10 dataset, the model showed an accuracy of 80.2%, exceeding LeNet-5 but falling short of ResNet and EfficientNet. This confirms the effectiveness of the model for tasks of medium complexity and systems with limited resources. Conclusions. Modified CNN is a compromise between simple classical methods and complex modern architectures. It provides an optimal balance between accuracy and learning speed, making it suitable for practical application in mobile and embedded systems. Further research may focus on the use of more complex datasets, automatic hyperparameter selection, and the integration of self-attention mechanisms. Scope of application of the results obtained: medium-complexity computer vision systems, mobile and embedded devices with limited resources, applied image classification tasks.