DSK-YOLO: Feature-level Super Resolution Boosted Industrial Defect Detection

Despite the significant advancements made in industrial defect detection, accurately and timely identifying complex and small-sized defects remains a challenge. Most current lightweight defect detectors are unable to fully extract both global and local contextual information due to their simplified network architectures. To address the above issues, this paper introduces a novel real-time detector DSK-YOLO, which efficiently enhances global and local contextual information with a lower number of parameters. Specifically, DSK-YOLO comprises two key components: DSKblock and DSKSR. The DSKblock employs dilated separable kernels to expand the effective receptive fields (ERFs) without deep layer stacking, thereby identifying complex defects. For small-sized defects detection, we develop a feature-level super resolution (SR) auxiliary branch to enhance local contextual information in the training phase. Moreover, the train-only SR branch brings no extra computational overhead for inference, making it an impressive choice for real-time tasks. Experimental results demonstrate that, on the industrial datasets NEU-DET and ESD, DSK-YOLO achieves mAP of 45.7% and 64.8%, which are 1.4% and 1.0% higher than those of the baseline model YOLOv8n. Our proposed DSK-YOLO offers a favorable tradeoff between precision and parameters compared to state-of-the-art models.