Balancing Accuracy and Efficiency: A Comparative Study of CNN Models on Versatile Image Datasets

In this work, we explore a varieity of Convolutional Neural Network architectures, including both existing architectures and proposing some variations, and evaluate their performance across publicly available image datasets focused on real world use cases in Bangladesh. At first, we examine different choices in the design of our architecture, such as convolutional depth, pooling strategies, bottleneck layers, normalization and classifier head configuration in order to propose a compact and stable architecture for image classification. Based on our findings, we propose a compact CNN built around a shallow convolutional structure with max pooling layers and a global average pooling layer at the end. Then, we compare the proposed architecture against VGG16 and ResNet18, both pre-trained and with transfer learning, in order to evaluate its performance. Our results conclude that the proposed architecture can achieve competitive performance in various datasets with different dataset sizes while offering an advantage in model size and training efficiency.