PCA METHOD IMPACT ANALYSIS ON DEEP NEURAL NETWORKS ACCURACY AND ARCHITECTURE FOR IMAGE CLASSIFICATION

Abstract

The growth of visual data in modern information systems creates a need to reduce the computational complexity of classification methods. This paper presents the impact study of data preprocessing by the Principal Component Analysis (PCA) method on the effectiveness of training and classification ability of deep neural networks. The focus is on a comparative analysis of two architectures — the Multilayer Perceptron (MLP) and the Convolutional Neural Network (CNN). The work is aimed at solving the urgent scientific and practical problem of optimizing computing costs in computer vision systems operating in conditions of limited hardware resources. Based on the analysis of experimental data obtained on the MNIST image set, a comparative analysis of four architectural approaches was performed: classical MLP, classic CNN, hybrid PCA+MLP, and hybrid PCA+CNN. The effect of loss of spatial locality during image transformation by the PCA method, which leads to deterioration of the results of CNN-based models, as opposed to reduction of computational complexity of MLP-based models while maintaining classification accuracy, is studied. The paper provides a combined image classification method, an analysis of the obtained accuracy metrics and loss function, as well as a justification of the observed phenomena. The conducted study allows us to draw conclusions about the feasibility of using the PCA method in image classification problems in combination with MLP. The results show the importance of aligning data preprocessing methods with the architectural features of machine learning models. In both hybrid models, the number of parameters was reduced by 4 times, while the PCA+MLP training time was reduced by 2 times with an accuracy of 97.87%, and the PCA+CNN training time was reduced by 8 times with an accuracy of 95.02%. The PCA+MLP hybrid model is found to have 4 times lower computational complexity and 12 times less training time than CNN, with a small accuracy loss of 1.5%.