Enhancing Fruit and Vegetable Image Classification with Attention Mechanisms in Convolutional Neural Networks

Classifying fruits and vegetables is a challenging task for traditional machine learning models, particularly convolutional neural networks (CNNs), which struggle to differentiate between similar-looking items. This differentiation is crucial in agriculture for efficient sorting and quality control. This study aimed to enhance the performance of classification models using a dataset of 3,115 images across 36 fruit and vegetable classes from Kaggle. The research explored four architectures: CNN, MobileNet, DenseNet, and Xception. Fine-tuning each model for the dataset, the CNN achieved a baseline accuracy of 96%. Further exploration with additional layers in MobileNet, DenseNet, and Xception yielded accuracies of 91%, 89%, and 89%, respectively. To address these limitations, squeeze-excitation blocks were integrated, significantly improving accuracies: MobileNet_SE and Xception_SE reached 97%, and DenseNet_SE achieved 98%. This study demonstrates the efficacy of advanced machine learning in fruit and vegetable classification and encourages further innovation for practical agricultural applications.