Few-Shot Steel Defect Detection Method Based on Improved Meta-Learning Model

To address the problem of scarcity of certain defect samples that unable to meet the training requirements of deep learning models, a few-shot steel surface defect detection method based on an improved meta-learning model is proposed in this paper. Meta-learning algorithm MAML is combined with YOLOv7 network to enhance its ability to recognize small target defects in steel surfaces through the SPP module. To fit the scene variations in different steel defect images, domain adaptation optimizer is introduced to endow the model adapt to different scenes for defect detection quickly. In addition, to deal with the issue of high similarity between background and target in certain defect images, label smoothing is incorporated into the YOLOv7 algorithm to prevent model overconfidence, thereby improving its generalization and detection performance. Experimental results on the publicly available NEU-DET dataset demonstrate that compared with other methods, the proposed approach significantly improves the detection accuracy of steel surface defects under few-shot conditions and effectively solves the problem of few sample of surface defects for steel products.