A fast button surface defect detection method based on Siamese network with imbalanced samples

Abstract

Surface defect detection for button is a tough task because of complex surface texture, variety of defects, and limited defect samples which often leads to an imbalanced issue. Aiming at solving these problems, a similarity metric method based on Siamese network is proposed for detecting defects of button and applied in a practical machine-vision-based system. In our system, the Siamese network with a specifically designed loss function is used for automatic feature extraction and similarity metric of samples. The learning process minimizes the specific loss function, which drives intra-class distance among positives to be smaller and inter-class distance to be larger in the feature space, so that after training, defect-free samples are clustered while defect samples are mapped to outliers. The proposed method is evaluated on button datasets of multiple kinds of defects including dent, crack, stain, hole, uneven etc. Experimental results show 98% detection precision for the proposed method, and 95% detection precision when dealing with imbalance issue, indicating its advantage over conventional methods. Comparison experiments show that for our task, the proposed loss function outperforms other recent published loss functions in face recognition or ReID field. Moreover, we optimize our method with different strategies. Our method reaches 6 fps detection speed on an embedded DSP platform, indicating its potential in providing an effective approach for online detection on production.