LeafFocusAI: An ROI-Aware Deep Learning Framework for Automatic Leaf Disease Detection Using FocusNet-LDD

Plant disease is a silent threat to global food security and agriculture; the only way to manage plant diseases effectively is through accurate and timely diagnosis. State-of-the-art manual inspection methods are Labor-Intensive and prone to error. In contrast, the few existing automated deep-learning methods struggle with reduced accuracy due to the automatic extraction of superfluous background information and a loss of interpretability. Most state-of-the-art models studied complete leaf images, ignoring localized disease regions, making them less robust and practical. In this paper, we present FocusNet-LDD, a ROI-aware deep learning framework designed to narrow the gaps with state-of-the-art Region of Interest (ROI) extraction, attention mechanisms, and sequential feature modelling, thereby improving plant disease detection. The proposed methodology utilizes YOLOv8 and U-Net architectures to target the areas occupied by diseased leaves, thereby minimizing background noise and concentrating on features associated with the symptoms. The classification model utilizes CBAM to enable spatial and channel-wise attention features, followed by a Transformer encoder that learns contextual representations to support the classification of discrete disease classes across various image acquisition settings. FocusNet-LDD leverages a dual-direction dilated convolution framework to enhance the representation ability of image features, incurring only a minor increase in time and space costs. This is demonstrated through extensive experiments on benchmark datasets, which show that it achieves the best overall accuracy (98.79%) compared to the baseline and more recent state-of-the-art models. Ablation studies validate the contribution of each module,  and Grad-CAM visualizations also provide explainability by highlighting which disease-relevant regions drive predictions. The high accuracy,  interpretability, and robustness of the proposed framework might pave the way for it to become a real-world tool for timely disease diagnosis of crops and appropriate decision-making. Its modular architecture also enhances its integration within precision agriculture systems and mobile platforms that can operate under low-resource conditions, promoting sustainable crop management and mitigating yield losses.