Adaptive and dynamic convolution network for accurate and lightweight detection of unsound wheat kernels

Wheat, a global staple, is prone to unsound kernels during storage and processing, reducing quality, safety, and value. Rapid and accurate detection of these defects is critical for grain grading, pricing, and ensuring food safety. This study proposes YOLO-CAD, a lightweight composite-attention detection model derived from YOLOv11, specifically optimized for small-object recognition in complex backgrounds. The model integrates three key modules, ODConv, ADown, and CGLU, to (i) capture multi-scale textures, (ii) retain fine details while downsampling, and (iii) emphasize informative channels. A high-quality dataset of 2762 RGB images was constructed, covering one sound and four types of unsound kernels. Experimental results show that YOLO-CAD achieves an mAP@0.5 of 92.6 % and mAP@0.5:0.95 of 50.0 %, improving by 3.2 % and 1.6 % over YOLOv11, respectively, while maintaining a low computational complexity of 7.2 GFLOPs. The model demonstrates enhanced robustness in recognizing kernels with blurred edges and subtle morphological differences, and achieves real-time inference at 20 FPS on Jetson Nano. Comparative analyses confirm YOLO-CAD's superior accuracy–efficiency balance and strong generalization under variable lighting conditions. This work provides an effective technical pathway for intelligent grain defect detection, contributing to the advancement of smart grain storage and food safety management.