Monitoring and precisely localizing rice diseases is essential for agricultural productivity and food security. Existing detection methods face challenges such as high computational complexity, semantic information loss, difficulty detecting small targets, and limited robustness. To address these issues, this study proposes ECL-RTDETR, an enhanced RT-DETR–based rice disease detection model. First, a lightweight EfficientViT backbone is employed for feature extraction, incorporating a streamlined multi-head self-attention module to improve inference speed, reduce computational cost, and strengthen local feature extraction. Second, the CARAFE upsampling operator is introduced to better preserve detailed feature information without added computational burden, enhancing fine-grained representation. Finally, standard convolution in the neck network is replaced with LDConv (lightweight dynamic convolution) to enable adaptive feature learning under complex conditions, addressing variations caused by illumination, occlusion, and disease diversity. Experimental results show that ECL-RTDETR improves mAP@0.5 by 0.7%, increases detection speed by 22.2 FPS, and reduces computational cost by 81.8 GFLOPs and parameters by 22.12M compared with the baseline RT-DETR. Overall, ECL-RTDETR delivers superior accuracy, speed, and efficiency, offering a robust solution for intelligent rice disease detection and localization, and advancing smart agriculture and sustainable food security.
Zhang et al. (Wed,) studied this question.