Surface Defect Inspection Under a Small Training Set Condition

Abstract

The detection of surface defects in industrial production is an important technology for controlling product quality. Many researchers have applied deep learning methods to the field of surface defect detection. However, obtaining defect sample data in industrial production is difficult, and the number of samples available to train detection networks is not sufficient. Based on the you only look once (YOLO) detection system, we propose a lightweight small sample detection network (SSDN) to overcome the problem of fewer samples in surface defect detection. The SSDN is demonstrated to be a suitable network to represent defect image features as it is better at feature extraction and easier to train. We used only 10/type images to train the SSDN model without data enhancement techniques and achieved excellent results (average accuracy 99.72%) on defect detection benchmark data. Experimental results verify the robustness of the model.