Adversarial elimination weak supervision target detection method based on channel selection

Abstract

The invention relates to a feature channel selection-based adversarial elimination weak supervision target detection method, which is used for solving the problem of weak supervision target detectionpositioning errors. The method comprises the following steps: firstly, taking weak supervision depth target detection as a bottom layer framework, generating candidate boxes on training set data by adopting a selective search method, and taking the candidate boxes, training set images and corresponding image tags as inputs of a weak supervision network; secondly, constructing a feature extractionnetwork model by taking VGG16 as a basic network, performing channel weighted selection on the obtained feature image in a feature channel compression mode, and exciting an image feature layer beneficial to classification to suppress a feature layer having interference on classification; then, adopting an adversarial elimination method to obtain complete feature expression capable of expressing animage target as input of a prediction network; and finally, training a prediction network according to the multi-task cross entropy loss to realize target detection. According to the invention, the position of the target object can be positioned more accurately, and the object identification precision can be improved.

Description

technical field [0001] The invention belongs to the technical field of target detection and introduces a weakly supervised target detection method based on confrontation elimination. Background technique [0002] With the development of science and technology and the improvement of the intelligent level of human life, mobile robots have gradually entered human production and life, and have been widely used in various industries. Object detection based on mobile robots has been widely used in inspection, security, video surveillance and search and other fields. The current deep learning target detection algorithm generally requires a large number of manually labeled data sets, which not only wastes a lot of manpower, material resources, and financial resources, but also incorrectly labeled data will also affect the robustness of the model and the detection accuracy of the model. The target detection algorithm based on weak supervision only needs image-level labels to complet...

AI Technical Summary

Problems solved by technology

In the feature extraction network, the feature channel selection strategy is used to increase the weight of the feature channel that positively promotes the classification, and then the feature confrontation elimination is used to extract more comprehensive features of each detected object, thereby solving the problem of weak supervision target detection positioning error

Images