A Defect-Sensitive Loss Function based on Siamese Network to Defect Detection with Imbalanced Samples

Abstract

Surface defect detection for industrial production is tough for variety of defects and limited defect samples which make it difficult to extract expected effective defect features. To solve this problem, a defect-sensitive loss function based on Siamese Network is proposed for detecting defects of industrial production surface. The learning process minimizes the designed loss to drive the intra-class distance of the defect-free images to be smaller and enlarge the distance between hardest defect image and the defect-free images. The proposed method is evaluated on a real-word dataset. Experimental results show 100% accuracy for proposed method with imbalanced rate 10:1 and 50:1, indicating its advantage over classification CNN method. Comparison experiments show that proposed loss function outperforms other recent published loss function, the proposed loss function can be more sensitive to defect samples.