Soft matching network with application to defect inspection

Abstract Defect detection with templates is a major concern in manufacturing. Including subtraction and template matching, traditional methods based on images stumbled over the diverse disparities of input image pair. Yet, learning-based approaches have not been explored on this task. This paper proposed a learning-based soft template matching network for defect detection, using an innovative attention mechanism. Employing feature-pyramid-network-based atrous convolution enables our model to perceive multi-scale features. The proposed contrastive attention module enhances the query feature map. Experimental results demonstrate that our network can capture defects under the interference of disparities based on the correspondence of input image pair, showing practical value for industrial defect detection.