Positive sample expansion method for surface defect detection

Abstract

The invention discloses a positive sample expansion method for surface defect detection, which comprises the following steps: inputting position coordinates of a target frame and a detection frame generated by a detector into a non-maximum suppression unit, removing an overlapped detection frame through the non-maximum suppression unit, and reserving the detection frame containing target information; negative samples, positive samples and semi-positive samples are screened and judged through a sample judgment unit; sampling a negative sample, a positive sample and a semi-positive sample through a sample sampling unit according to a set positive and negative sample sampling proportion; inputting all sampling data into a loss regression unit, performing detection frame classification loss training and detection frame positioning regression training through the loss regression unit, and further expanding positive samples; according to the invention, the overlapping degree between the detection frame and the target frame can be measured more accurately; meanwhile, the imbalance problem of positive and negative samples is improved, the method can be directly migrated and applied to multiple target detection networks, pre-training is not needed, and the universality is good.

Description

technical field [0001] The present invention relates to the technical field of target detection, in particular to a positive sample expansion method for surface defect detection. Background technique [0002] In recent years, object detection technology based on deep neural network has developed rapidly and has been successfully applied in defect detection of many products. The two-stage R-CNN (Region-Convolutional Neural Networks) target detection algorithm is one of the current mainstream target detection methods. In the two-stage detector, the first stage uses a region proposal network (Region Proposal Network, RPN) to screen out several candidate boxes from dense pre-determined bounding boxes (anchor boxes), and then the second stage uses a region of interest subnetwork (RoI -subnet) performs target classification and regression positioning on these candidate boxes. [0003] In the two-stage R-CNN method, when the algorithm is trained, only a few anchor boxes that high...

AI Technical Summary

Problems solved by technology

[0004] When the two-stage target detection algorithm is used for product surface defect detection, due to the unique nature of defects, such as: incompatibility of defect categories, different defect sizes, etc., the number of anchor frames assigned to the target, that is, the number of positive samples, Even less, exacerbating the imbalance between positive and negative samples in the anchor box, reducing the network convergence speed and detection accuracy

Images