Evaluating 2-Layer and 4-Layer CNN Architectures for Efficient Digital Image Classification

Abstract

This work focuses on the essential problem of designing the convolutional neural network (CNN) architecture for efficient and accurate digital Image classification. CNNs have been developed to address the issue of effective and accurate digital image classification. The impact of depth on model effectiveness is investigated by systematically comparing a simpler 2-layer CNN with a deep 4-layer CNN using the MNFST, Fashion-MNIST, CIFAR-10, CIFAR-100, and Tiny ImageNet datasets. Each architecture used consecutive convolutional layers with ReLU activation and maximal pooling to extract hierarchical features, followed by dense layers for classification. The models were carefully trained and tested using both CPU and GPU architectures to determine computational efficiency. The results show a considerable performance discrepancy, the 4-layer model achieved a higher test accuracy of 93.5% than the 2-layer model's 89.2%. Additionally, the deeper network showed improved metrics for F-score, precision, and recall, indicating predictions that were more reliable and certain. This work is innovative because it provides a clear architectural trade-off analysis and empirically quantifies that depth-related performance increases on standardized benchmarks. The study suggests that spending money on deeper networks is warranted for intricate multi- category classification tasks, providing a useful guideline for resource conscious model selection in embedded and large-scale systems.