KGC-YOLOv5: A YOLOv5-based Method for Steel Surface Defects Detection

To address the challenges posed by diverse defect types, significant scale variations, and potential missed detections of small objects on steel surfaces, this paper introduces the KGC-YOLOv5 model, a lightweight steel surface defect detection model based on YOLOv5. Firstly, the K-means++ algorithm is employed instead of the K-means algorithm to cluster anchor boxes, resulting in anchor boxes that are specifically tailored for steel surface defects and improving the quality of the target anchor boxes. Next, in the backbone network, the C3 layer is replaced with C3Ghost to enhance feature extraction capabilities and create a lightweight model. To enrich deep features and facilitate the detection of medium and small objects, we incorporate the lightweight upsampling method known as CARAFE. The experimental results demonstrate that the KGC-YOLOv5 model achieves a mAP of 76.9% on the NEU-DET dataset, showcasing a performance improvement of 3.1% over YOLOv5. Additionally, the model achieves a 4% increase in precision, reduces the number of parameters by 13%, and decreases the GFLOPs by 3.1.