End-to-end visual matching method and device based on dynamic computation graph, equipment and medium

CN122391678APending Publication Date: 2026-07-14NAT INNOVATION INST OF DEFENSE TECH PLA ACAD OF MILITARY SCI
View PDF 0 Cites -1 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NAT INNOVATION INST OF DEFENSE TECH PLA ACAD OF MILITARY SCI
Filing Date
2026-04-20
Publication Date
2026-07-14

AI Technical Summary

Technical Problem

Existing feature matching methods suffer from information fragmentation, error accumulation, and low utilization of computational resources, resulting in insufficient stability and efficiency in matching under complex conditions.

Method used

An end-to-end visual matching method based on dynamic computation graphs is adopted. Multi-scale features are extracted through a neural network backbone with shared parameters, and feature enhancement is performed using confidence maps and attention maps as feedback. Feature recomputation is triggered in regions with high uncertainty, and a lightweight controller is used to adaptively adjust the computation depth to achieve feature fusion and matching.

Benefits of technology

It improves the accuracy and efficiency of visual matching, especially maintaining stability under complex conditions such as large changes in viewing angle, low texture or uneven lighting, adapting to different task scenarios, and supporting real-time performance and scalability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122391678A_ABST
    Figure CN122391678A_ABST
Patent Text Reader

Abstract

The application relates to an end-to-end visual matching method, device and equipment based on a dynamic calculation graph and a medium. The method extracts multi-scale features by inputting two visual images into a shared parameter neural network trunk. A normalized similarity matrix of corresponding pixels / feature blocks is calculated based on the features, a confidence map and an attention map are generated from the matrix, and the confidence map and the attention map are fed back to the trunk to enhance the features, so that enhanced multi-scale features are obtained. The matching probability is calculated by using a Softmax algorithm based on the similarity matrix, and the matching uncertainty is calculated by using an entropy index. If the uncertainty is greater than a threshold, a feature recalculation branch is triggered, refined features are obtained by performing deep calculation on the enhanced features through a feature recalculation network, and the enhanced features and the refined features are fused through a dynamic calculation mask to obtain the final features of the two images. The feature corresponding relationship is calculated based on the final features, and a visual feature matching result is output. The method can effectively improve the visual matching precision and optimize the calculation efficiency.
Need to check novelty before this filing date? Find Prior Art