Cross-domain few-sample camouflage target detection method based on semantic perception and information alignment

This cross-domain few-shot camouflage target detection method, which utilizes semantic awareness and information alignment, solves the problems of feature coupling and semantic guidance mechanism lack in existing methods by using bidirectional multi-head attention mechanism and global text feature alignment. This improves the model's performance in cross-class and cross-domain target detection and reduces the need for large-scale labeled data.

CN120976520APending Publication Date: 2025-11-18RES & DEV INST OF NORTHWESTERN POLYTECHNICAL UNIV IN SHENZHEN
View PDF 3 Cites 1 Cited by

Patent Information

Application Number
CN202511021408.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-24
Publication Date
2025-11-18

AI Technical Summary

Technical Problem

Existing methods for detecting camouflaged targets with few samples suffer from feature coupling issues and a lack of semantic guidance mechanisms, resulting in poor detection performance in cross-domain scenarios, and complete failure, especially in complex scenarios.

Method used

A cross-domain few-shot camouflage target detection method based on semantic awareness and information alignment is adopted. Features are decoupled through a bidirectional multi-head attention mechanism, and global text features and local visual features are combined for optimization. A pre-trained CLIP text encoder and a visual branch CLIP text encoder are used to establish cross-modal semantic alignment. A refined segmentation decoder is used for feature fusion and segmentation decoding.

Benefits of technology

It significantly reduces the coupling problem between target features and background noise, enhances feature decoupling efficiency in cross-domain scenarios, and improves the feature decoupling capability in cross-domain scenarios. By jointly optimizing the extracted global semantic features and visual local features, a unified semantic representation space is established, which improves the model's semantic understanding of cross-category and cross-domain targets and reduces the need for large-scale accurate labeled data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120976520A_ABST
    Figure CN120976520A_ABST
Patent Text Reader

Abstract

The invention discloses a cross-domain few-sample camouflage target detection method based on semantic perception and information alignment, and the method comprises the steps: extracting multi-scale features from a support image and a query image, and calculating an initial correlation matrix; performing splitting and feature enhancement on the initial correlation matrix to obtain an enhanced correlation matrix; performing convolution operation on the enhanced relevance matrix to obtain an initial query mask; performing feature enhancement and feature fusion to obtain fusion features, and according to the enhanced fusion features, querying shallow layer features of the image, supporting image features and a camouflage target mask to obtain a final prediction mask; global text features and local visual features are mapped to the same feature space, loss is calculated, and semantic alignment is carried out; and performing up-sampling operation and threshold processing on the final prediction mask to obtain a camouflage target detection result. According to the method, the coupling problem of target features and background noise is reduced, and the feature decoupling capability in a cross-domain scene is enhanced; and the demand on large-scale accurate annotation data is reduced.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Similarity matrix guided few-sample semantic segmentation method and system

    CN116468895A

  • Small sample camouflage target segmentation method

    CN118097150A

  • Counterfactual context-aware texture learning for camouflaged object detection

    WO2024187334A1

Cited By

  • Multi-modal industrial anomaly detection and segmentation method based on category perception

    CN121304656A