Unsupervised domain adaptation object detection method based on riemannian statistical alignment and related device

CN121921599BActive Publication Date: 2026-06-09SHENZHEN UNIV

Patent Information

Authority / Receiving Office: CN · China
Patent Type: Patents(China)
Current Assignee / Owner: SHENZHEN UNIV
Filing Date: 2026-03-26
Publication Date: 2026-06-09

AI Technical Summary

Technical Problem

Existing unsupervised adaptive object detection methods struggle to adequately characterize high-order channel correlations in deep feature representations within Euclidean space. This results in the destruction of crucial structural information that distinguishes semantic content from environmental noise in cross-domain object detection. Furthermore, the model training process is complex and inference costs are high.

Method used

A Riemann statistical alignment-based method is adopted, which maps feature maps onto a symmetric positive definite Riemannian manifold, uses the covariance matrix to characterize the higher-order statistical properties of the features, and introduces a manifold domain discriminator and manifold statistical alignment loss for iterative optimization to construct an unsupervised adaptive target detection model based on Riemann statistics.

Benefits of technology

It effectively overcomes the geometric distortion problem in Euclidean alignment, improves the accuracy and efficiency of cross-domain feature alignment, achieves real-time detection with zero computational overhead, and maintains the original detection speed.

✦ Generated by Eureka AI based on patent content.

Smart Images

Figure CN121921599B_ABST

Patent Text Reader

Abstract

Embodiments of the present application relate to the technical field of artificial intelligence, and disclose a kind of unsupervised domain adaptive target detection method based on Riemann statistics alignment and related device, the method comprises: obtaining image to be detected;The image to be detected is input into target detection model, and target detection result is obtained;Source domain sample set, target domain sample set and style migration sample set are input into target detection model, and first prediction loss and second prediction loss are calculated;The multiscale feature map corresponding to source domain sample set and the target domain sample set is mapped into Riemann manifold statistical representation;Statistical representation is input into manifold domain discriminator, and the loss of confrontation is calculated;According to first prediction result and second prediction result, alignment loss is calculated;According to first prediction loss, second prediction loss, alignment loss and the loss of confrontation, iterative optimization is carried out, and target detection model is obtained;Output target detection result.The present application embodiment realizes the model training of high performance, zero inference overhead.

Need to check novelty before this filing date? Find Prior Art