Vegetable Classification with CBAM-integrated VGG-16

Abstract

This study investigates the efficacy of integrating the Convolutional Block Attention Module (CBAM) into the VGG-16 architecture for fine-grained vegetable classification. The standard VGG-16 model, while a powerful feature extractor, treats all spatial regions and channel-wise features equally, which can be suboptimal for discriminating between visually similar vegetable species. To address this limitation, we sequentially integrated CBAM into the convolutional layers of VGG-16. CBAM's dual-attention mechanism—comprising a channel attention module to recalibrate feature channel weights and a spatial attention module to highlight informative regions—allows the model to focus adaptively on discriminative parts of the input image, such as unique textural patterns on a vegetable's skin or its distinctive shape. Experimental results on the Kaggle Vegetable Image Dataset demonstrate that the proposed CBAM-integrated VGG-16 model achieves a top-1 accuracy of 94.7%, outperforming the baseline VGG-16 model by a significant 4%. This performance gain underscores the critical role of attention mechanisms in enhancing feature representation for agricultural image analysis.