Manuscript received August 31, 2025; revised October 9, 2025; accepted March 10, 2026; published September 10, 2026.

Abstract—To ensure an adequate supply of food while preventing loss of yield, timely identification of rice leaf diseases is vital. Traditional deep learning models often exhibit poor discrimination among visually challenging diseases, low generalization for images with natural background, and fail to handle datasets with class imbalance. Convolutional Neural Networks (CNN)-based models perform poorly in a variety of field settings due to their significant emphasis on local feature extraction and difficulty capturing long-range spatial patterns. Attention mechanisms that are vital for emphasizing fine-grained lesion points are often ignored, which lowers localization accuracy and makes them more susceptible to background noise. This paper proposes a dual-branch attention-enhanced robust CNN model with transformer architecture named as CATNet to address these problems by combining transformer-based global context modelling with CNN-driven local texture learning. To enhance lesion-focused feature refinement and suppress unimportant background information, the model uses an attention fusion technique. On public as well as self-generated comprehensive real-world datasets, CATNet achieves 98% test accuracy with 99% category-wise recognition, effectively handling class imbalance and complex backgrounds where many state-of-the-art models fall short. The key design principle is to maintain it lightweight with only 3.9 M parameters and utilize the least amount of memory of 2287 MB as compared to state-of-the-art models, thereby rendering it ideal for real-world, resource-constrained precision farming. CATNet beats popular models like EfficientNetB0, DenseNet169, and MobileNetV3Small in terms of tradeoff between accuracy as well as efficiency, reaching 98% test accuracy with just 332 MFLOPs, thereby rendering it as a lightweight, efficient, and robust model for detecting rice diseases in real time systems.

Keywords—deep learning, rice disease detection, transformer, depth-wise convolution, convolutional neural network, attention mechanism

Cite: Sanam Salman Kazi, Bhakti Palkar, Vishwayogita Savalkar, and Supriya Khaitan, "CATNet: Attention-enhanced Robust CNN Architecture with Transformer for Rice Crop Disease Detection," Journal of Image and Graphics, Vol. 14, No. 5, pp. 733-746, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).