HRT-AD Net: Hybrid Residual Transformer Attention-based Deep Network for Rice Disease Diagnosis

Abstract

One of the most significant staple crops that are consumed globally is rice (Oryaza sativa), and to avoid losses in its production and to maintain food security, it is necessary to timely identify diseases on its leaves. Nevertheless, disease detection in the real-field setting is not easy owing to the complication of the background, fluctuation of the light, inter-class similarity and the tiny size of the lesion areas. In this paper, the Hybrid CNN-Transformer Attention Network (HRT-ADNet) will be proposed to achieve robust detection and classification of rice disease. The suggested model combines a lightweight convolutional neural network backbone that localizes textures and a transformer encoder that models global features. A dual attention model that includes channel and spatial attention is proposed to improve lesion-conscious feature refinement. Besides, a multi-task branch of segmentation is optional and is used to enhance interpretability and localization performance. The model is trained and tested on a four-class rice disease dataset containing: Brown Spot, Bacterial Leaf Blight, Rice Blast and Leaf Smut. The model proves to have an experimental result of 96.8% test accuracy and a macro-average AUC of 0.975, which is lower than traditional CNN and standalone transformer models. The lightweight architecture guarantees a lower level of computational complexity and feasibility of deployment in real time. The suggested framework offers a solution that is reliable, interpretable, and effective in the diagnosis of intelligent rice disease in the context of precision agriculture.