An image detection method, device, equipment and medium

CN115937071BActive Publication Date: 2026-07-03QINGDAO HISENSE ELECTRONICS TECH CONSULTANCY CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
QINGDAO HISENSE ELECTRONICS TECH CONSULTANCY CO LTD
Filing Date
2022-05-13
Publication Date
2026-07-03

AI Technical Summary

Technical Problem

In existing technologies, unsupervised learning methods generate pseudo-labels with high noise levels, which reduces the recognition accuracy of the network model. Furthermore, image block sampling ignores spatial features, thus affecting recognition accuracy.

Method used

By pre-training the teacher network to construct pseudo-labels representing the feature differences between the original images, and using the recognition results of the target teacher network to train the student network, the feature recognition ability of the network model is gradually improved by combining the cross-entropy loss function and filtering.

Benefits of technology

It improves the recognition accuracy of the network model, makes full use of the spatial features between images, and enhances the feature recognition capability of the network model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115937071B_ABST
    Figure CN115937071B_ABST
Patent Text Reader

Abstract

This application discloses an image detection method, apparatus, device, and medium. The method includes: inputting a pair of images to be processed into a target student network to determine the image differences between the images based on the output of the network model. During the training phase of the target student network, a first pseudo-label representing the feature differences between each original image within the original image pair is constructed using a pre-trained teacher network. The student network is then trained using the original image pairs and the first pseudo-label. In this process, the teacher network is pre-trained using the original image pairs, enabling the trained target teacher network to recognize the feature differences between each original image within the original image pair. Furthermore, the student network is trained based on the target teacher network's recognition results of the original image pairs, further improving the student network's feature recognition ability on the basis of the target teacher network's recognition, thereby improving the recognition accuracy of the network model.
Need to check novelty before this filing date? Find Prior Art