Image recognition method and device

By constructing a test-time adaptation method for visual language models based on the complementary memory system of the biological brain, the problems of forgetting old adaptation results and semantic alignment of visual language models in dynamic environments are solved, and efficient recognition stability and robustness are improved.

CN122368628APending Publication Date: 2026-07-10TSINGHUA UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
TSINGHUA UNIVERSITY
Filing Date
2026-04-27
Publication Date
2026-07-10

AI Technical Summary

Technical Problem

Existing test-time adaptation (TTA) methods are prone to forgetting old adaptation results when dealing with dynamic and constantly changing test environments, and cannot achieve good semantic alignment between visual representations and text representations, resulting in limited adaptation performance and robustness of the model under complex distribution shifts.

Method used

By drawing on the complementary memory systems of the hippocampus and neocortex in the biological brain, we construct a precise memory cache and an abstract memory cache. Through cross-modal consistency, we jointly optimize the visual and text classifiers of the visual language model and dynamically adjust them to achieve a memory mechanism that combines long-term stable knowledge with short-term environmental experience. Finally, we combine the recognition results of visual and text branches to make the final decision.

Benefits of technology

It significantly improves the recognition stability and robustness of visual language models in dynamic and continuously changing scenarios. By taking into account both long-term stable knowledge and short-term environmental experience memory mechanisms, it breaks through the performance limit of a single modality and improves the model's adaptability and recognition accuracy in complex environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122368628A_ABST
    Figure CN122368628A_ABST
Patent Text Reader

Abstract

This application provides an image recognition method and apparatus, relating to the field of image processing. The method includes: acquiring a test image and its initial visual features; determining standard visual features corresponding to multiple candidate categories based on the test image; determining target visual features, a target visual classifier, and a target text classifier for the test image based on the initial visual features, standard text features corresponding to multiple candidate categories, and the standard visual features, wherein the accuracy of the target visual features is higher than that of the initial visual features; and determining the target recognition results of the test image on multiple candidate categories based on the recognition results of the target visual features on the target visual classifier and the target text classifier, respectively. This application addresses the shortcomings of existing TTA methods, such as the tendency to forget old adaptation results in new test data streams and the inability to achieve good semantic alignment between visual representations and text representations.
Need to check novelty before this filing date? Find Prior Art