Target detection network training method and system, network, device and medium

Abstract

The invention discloses a target detection network training method and system, a network, a device and a medium, and relates to the field of target detection, including: marking the training data with an information regression frame; dividing the target mask into two parts partA and partB according to the grouping ratio K; In the partA part, combined with the target mask information corresponding to the target mask, the corresponding target mask area on the original image is masked to obtain the history image; for the partB part, all the label information corresponding to the partB part on the original image is retained to obtain the image to be input ; Input the history image and the image to be input into the target detection network to be trained, the original image passes through the first layer of the target detection network to obtain the first feature map, and the history image passes through the first layer of the target detection network to obtain the second feature map, and the first The feature map and the second feature map are added and input to the subsequent layer of the target detection network for correlation calculation to train the target detection network; the present invention enables the target detection network to still have a good detection effect under highly overlapping targets.

Description

technical field [0001] The present invention relates to the field of target detection, in particular to a target detection network training method and system, a target detection network or model, a target detection system, a target detection network training device, and a computer-readable storage medium. Background technique [0002] In the existing business scenarios of industrial defect detection, there are often a lot of overlapping defect data. The existing detection methods to solve overlapping defects mainly include the following two methods: the first is to make multiple predictions on a picture, and each A prediction excludes the result of the previous prediction, and then integrates the prediction results of each time to finally obtain the detection result including all targets. Although this method has a better detection effect on overlapping targets than the original detection method, the overhead of redundant computing resources is large. It is not conducive to ...

AI Technical Summary

Problems solved by technology

[0004] 1) When multiple defects are located at the overlapping position, the defects below the overlap are covered by the defects above, resulting in a large difference between the characteristics of the defects located below the overlap and the complete characteristics of such defects, so it is difficult to be detected

[0005] 2) Even if the defect described in the first point can be detected by the detector, its confidence level will often be low, and it will be suppressed by the NMS used to suppress overlapping frames in the post-processing stage

Images