ND-YOLO: An enhanced framework integrating omni-dimensional dynamic convolution and occlusion-aware mechanisms for small defect detection in noodle cakes

Automated visual inspection of instant noodle cakes faces significant challenges stemming from minute defect sizes, complex morphological variations, and severe partial occlusion caused by noodle texture. To overcome these limitations, we propose ND-YOLO, an enhanced YOLOv8-based framework specifically designed for high-precision, real-time small defect detection in manufacturing environments. The architecture is strategically improved through: 1) a dedicated small-scale feature layer to prevent fine-grained feature loss; 2) the novel C2f-OD module, leveraging Omni-Dimensional Dynamic Convolution (ODConv) for morphology-aware feature representation; and 3) the Separated and Enhancement Attention Module (SEAM), which reconstructs the detection head to robustly mitigate false detections induced by texture occlusion. Additionally, we integrate a SpA-Former generative adversarial network for shadow removal preprocessing, significantly improving environmental robustness. ND-YOLO achieved a mean average precision (mAP50 of 94.9% with preprocessing, representing a substantial improvement over the YOLOv8s baseline. Crucially, real-time deployment verification on an industrial line, accelerated by TensorRT, confirms a stable throughput of 142 items/minute, validating the algorithm's superior practical deployability for critical quality control applications. • Proposed ND-YOLO framework for efficient detection of small defects in noodle cakes. • Integrated C2f-OD module to boost feature extraction for complex morphological defects. • Introduced SEAM head to reduce false detections caused by partial occlusion. • Employed SpA-Former in preprocessing to reduce shadow interference and enhance detection accuracy. • ND-YOLO achieves 94.9% mAP50, outperforming YOLOv8s and mainstream models in detection accuracy.