A Review Study of the Visual Geometry Group Approaches for Image Classification

Abstract

In the realm of advanced machine learning for image classification, Convolutional Neural Networks (CNNs) stand as a pivotal tool, with the Visual Geometry Group-16 (VGG16) model standing out for its emphasis on deepening and expanding CNNs architecture to achieve better accuracy. However, the complex design of VGG16 presents challenges regarding computational efficiency and scalability. This study addresses these issues by refining the VGG16 architecture through strategic modifications, including reducing convolution blocks, integrating batch normalization (BN) layers, and incorporating a global average pooling (GAP) layer alongside additional dense and dropout layers. The proposed architecture's effectiveness was assessed through comprehensive experiments across ten benchmark datasets, comparing its performance against the standard VGG16 architecture. The proposed architecture sped up the execution time by 63.7% on average across all benchmark datasets, compared to the standard VGG16. Furthermore, the results showed that the proposed architecture outperformed VGG16 by improving the classification accuracy by up to 30.1% based on the overall datasets. In summary, the proposed architecture is made to be compact and accurate. By adjusting parameters, it processes information quickly and accurately. It also includes features to prevent overfitting and improve classification, resulting in a significant advancement in image classification.