Robust training method for target detection network

Abstract

The invention relates to a robust training method for a target detection network, and the method comprises the steps: obtaining a training sample, wherein a part of detection targets on the training sample carries a manual labeling box; performing feature extraction on the training sample by using a target detection network, and generating a suggestion box on the training sample; marking an original sampling label on the suggestion box, wherein the original sampling label comprises a positive label and a negative label; carrying out pooling operation on the positive label by adopting a poolingbranch, and outputting a first region-of-interest feature; inputting the first region-of-interest feature into a mining network which is a fully connected neural network, and generating a new suggestion box label, namely a mining label, by the mining network; fusing the mining label and the original sampling label to generate a gold label; and applying the gold label to the training of the targetdetection network.

Description

technical field [0001] The invention relates to the technical fields of computer vision and target detection, in particular to a robust training method for a target detection network. Background technique [0002] In recent years, the object detection framework based on convolutional neural network (CNN) has become a powerful method for various computer vision tasks, and has been widely used in object localization and object statistics tasks. Meanwhile, object detection frameworks based on Convolutional Neural Networks (CNN) have been continuously improved and many excellent architectures have been proposed. Among them, region-based detection frameworks (e.g., Faster RCNN, FPN), which include a preprocessing step for region proposals, are widely used due to their more accurate detection performance. Meanwhile, many methods continue to improve the performance of feature extractors by optimizing their network architectures. However, how to enhance the robustness of training ...

AI Technical Summary

Problems solved by technology

[0005] Compared with the existing technology, a robust training method for the target detection network in the embodiment of the present application is adopted, and the process of mining the suggestion frame and label fusion is added in the training process of the target detection network, which effectively overcomes the problem caused by manual labeling. If the frame is missing or the set threshold (the first threshold and the second threshold) is too high or too low, the suggested frame labeling is wrong or the phenomenon of too many false positives occurs in the sample, which improves the anti-interference ability of the network training process

Images