Attention mechanism-based ultralightweight deep learning method for automated multi-fruit disease recognition system

Abstract

Automated disease recognition plays a pivotal role in advancing smart artificial intelligence (AI)-based agriculture and is crucial for achieving higher crop yields. Although substantial research has been conducted on deep learning-based automated plant disease recognition systems, these efforts have predominantly focused on leaf diseases while neglecting diseases affecting fruits. We propose an efficient architecture for effective fruit disease recognition with state-of-the-art performance to address this gap. Our method integrates advanced techniques, such as multi-head attention mechanisms and lightweight convolutions, to enhance both efficiency and performance. Its ultralightweight design emphasizes minimizing computational costs, ensuring compatibility with memory-constrained edge devices, and enhancing both accessibility and practical usability. Experimental evaluations were conducted on three diverse datasets containing multi-class images of disease-affected and healthy samples for sugar apple (Annona squamosa), pomegranate (Punica granatum), and guava (Psidium guajava). Our proposed model attained exceptional results with test set accuracies and weighted precision, recall, and f1-scores exceeding 99%, which have also outperformed state-of-the-art pretrain large-scale models. Combining high accuracy with a lightweight architecture represents a significant step forward in developing accessible AI solutions for smart agriculture, contributing to the advancement of sustainable and smart agriculture.