Surface Defect Detection Based on Deep Convolutional Siamese Networks and Defect Saliency

  • Rixian Liu, Minghai Yao
Keywords: Defect Detection, Deep Convolutional Siamese Network, Deep Learning, Visual Saliency, Reverse Connection.


Compared with the general target, defect is sparse in the product surface image to be detected. There are many challenges in defect detection. Firstly, there are much fewer pixels available for small defective objects. Secondly, there is limited prior knowledge and experience in this area. The scale of defects changes considerable, and defective areas have greater visual saliency than other areas. This paper proposes a defect detection method using the deep convolutional Siamese neural network to automatically extract the features of defects and classify and identify defects based on the extracted features. Firstly, the defect image is manually labeled using Groundtruth, and a defect mask map image will be generated according to the Groundtruth annotation. The defect image and the defect mask map image will be respectively inputted into two sub-networks of the deep convolutional Siamese network to extract the saliency features and defect features. Finally, the enhanced convolution feature maps are sent to the RPN (Region Proposal Networks) for the initial assessment of the targets and background; the generated proposal boxes are sent to the detection network to be classified for specific categories; and the accurate target detection box is regressed by using the bounding box regression layer. The proposed method not only avoids the problem in traditional defect detection of manually selecting target features, but also solves the difficulty of detecting small defects. The method proposed in this paper has a high level of detection accuracy and is robust to defect types, defect sizes, defect scales and defect diversification. This paper puts forward a more efficient method for detecting defects in industrial product surface images.