LEAF DISEASE IDENTIFICATION USING AN ATTENTION MECHANISM AND A LIGHTWEIGHT DEEP RESIDUAL NETWORK

Abstract

The need for very accurate identification in the field of leaf disease detection is higher than ever before. This project presents SE-VRNet, a lightweight model developed specifically for the purpose of detecting leaf diseases on mobile devices. Its target audience is the agriculture industry, which heavily uses smartphones and tablets. Utilising state-of-the-art methods such as the deep variant residual network (VRNet) and an attention mechanism in the squeeze-and-excitation (SE) module, SE-VRNet extracts disease-related characteristics from leaf pictures with great skill, guaranteeing accurate diagnosis and classification of different leaf diseases.By accurately recognising diseases in leaf pictures, SE-VRNet overcomes obstacles such as scattered lesion sites and varying area widths. Lightweight models are crucial for mobile devices, and this experiment highlights the necessity of computing economy without sacrificing accuracy, which SE-VRNet does. The model's practicality and efficacy are emphasised, providing a hopeful answer for farmers regarding the prompt control of diseases and the preservation of crops.The team is aiming for a detection threshold of 0.85mPA or higher and is investigating other approaches like YoloV5 and YoloV8 to further improve performance. With a 99.80% training accuracy and a 96% test accuracy on self-generated data, the base paper's results with SE-VRNet provide a solid basis for this investigation into diverse deep learning models. These models promise improved efficiency in detecting leaf diseases and enhancing agricultural resilience.