Image information bidirectional guide fusion method, target detection method and system

By employing a bidirectional guided fusion method based on image information, and utilizing frequency domain decoupling and a bidirectional guided mechanism, the problems of detail loss and semantic ambiguity in small target detection in existing technologies are solved, achieving high-precision and robust detection of small targets.

CN122265791APending Publication Date: 2026-06-23SHANGHAI XINLIJI SEMICON CO LTD
5 Cites 0 Cited by

Patent Information

Application Number
CN202610729417.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-05-26
Publication Date
2026-06-23

AI Technical Summary

Technical Problem

Existing target detection networks struggle to balance detail preservation, noise suppression, and multi-scale adaptation for small targets in scenarios such as drone aerial photography and high-altitude surveillance, resulting in high false negative rates, inaccurate localization, and poor robustness for small targets.

Method used

A bidirectional guided fusion method based on image information is adopted. Through frequency domain decoupling and bidirectional guidance mechanism, high-frequency details in shallow features are separated from low-frequency semantics in deep features. Deep semantic information is used to guide the enhancement of shallow details, and feature fusion is performed in combination with an adaptive weight generation model.

Benefits of technology

It improves the accuracy and robustness of small target detection, reduces the false negative and false positive rates, and enhances the classification and localization accuracy of large and medium-scale targets.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention discloses an image information bidirectional guide fusion method, a target detection method and a system, and the method comprises the steps: inputting to-be-processed image information into a YOLO model, and obtaining a shallow feature image from a Backbone network and a deep feature image from a Neck network; performing characteristic frequency band decoupling on the shallow layer characteristic image and the deep layer characteristic image after up-sampling to respectively obtain a shallow layer high-frequency characteristic image and a deep layer low-frequency characteristic image; semantic information represented by the deep low-frequency feature map is analyzed to guide the shallow high-frequency feature map to enhance details, and a detail-enhanced image is obtained; evaluating the weight of each feature channel by using the shallow high-frequency feature map so as to guide the deep low-frequency feature map to calibrate semantics and obtain a calibrated image; and fusing the detail enhanced image with the calibration image. According to the invention, the problems of loss of detail information and fuzzy semantic information of small-scale targets in the existing target detection network are solved.
Need to check novelty before this filing date? Find Prior Art