Multi-scale defect detection on product surface in grayscale images using YOLOv8-Ms algorithm

Surface defect detection is crucial in engineering quality control. While existing single-stage and two-stage models have advanced the field, conventional object detection algorithms still face challenges in detecting multi-scale defects accurately and representing defect features in greyscale images due to their monotonicity. To address these issues, this paper proposes YOLOv8-MS, an enhanced YOLOv8 model for efficient and accurate multi-scale surface defect detection. We use multi-scale convolutions and design the MSC2f module to adapt to features at different scales. A novel GFPN is introduced for efficient multi-scale information fusion, and a multi-information attention module further boosts the network’s performance. To handle the single-channel nature of greyscale images, we develop an input preprocessing method. Experiments on three open-source datasets show that our model achieves a mean average precision (mAP) of 44.7% on NEU-DET, a 4.1% improvement over YOLOv8, and outperforms YOLOv8 and YOLOv5 on other datasets. Moreover, the inference time is not significantly increased, effectively balancing detection accuracy and efficiency.