Micro-defect Detection Based on Multi-scale Feature Backtracking Structure

Target detection has a wide application prospect in many scenarios, but the current surface defect detection generally has some problems such as low precision of small target detection or excessive resource cost. In this paper, aiming at the characteristics of industrial micro-defects and the idea that the most important component of Feature Pyramid Networks(FPN) is divide and conquer based on different situation, we propose a feature backtracking network structure. In this structure, negative weight based on background judgment and base judgment are introduced to enhance feature sensibility field and achieve an accurate target location prediction. The feature backtracking network structure ensures the accuracy of detection effectively by optimizing multi-scale feature fusion in FPN and using first-level feature to detect multi-layer feature processing. Surface pictures of mobile phones on a production line of XX factory were collected for labeling, and PSDD data set was made. The algorithm was tested and compared with the public data set VOC2012 to verify that the performance and speed of this method can meet the requirements.