A Multi-Disease Detection Method for Paddy Rice Based on Enhancing Detection Transformer With ConvNeXt-DAM-FFNet Refinement

Abstract

Global food security is seriously threatened by paddy rice diseases, which reduce annual yields in important growing regions. Real-world field circumstances with complex background interference provide significant obstacles for automated detection systems. Based on the Detection with Transformer methodology, this study offers a unique framework for the identification of plant diseases. Utilising the strong ConvNeXt architecture improves feature extraction, a suggested feature fusion network optimises cross-level contextual integration, and a deformable attention mechanism permits adaptive spatial localization. The Transformer architecture's structural changes improve the precision of detection. To improve generality, a new optimizer is used to update the model parameters. The Hard-Swish activation function is also included to improve the model's overall performance by fortifying its capacity to handle nonlinear features. Under varying illumination and occlusion conditions, the experimental evaluation shows superior detection performance with 80.0% precision, 83.2% recall and 81.6% F1-score with 61.5% mAP on a real field-collected dataset with 1200 images of four critical paddy rice diseases (bacterial panicle blight, blast, dead heart and hispa). Compared to the baseline model, it shows improvements of 9.3%, 11.9%, 10.6% and 5.5%, respectively. With potential uses in automating agricultural inspection procedures, this study provides a practical and efficient approach for identifying a variety of plant diseases in outdoor settings.