A passive domain adaptation infrared ship semantic segmentation method based on brightness guidance and historical feature alignment

By employing brightness-guided and historical feature alignment methods, the problems of data scarcity and cross-modal recognition in infrared ship semantic segmentation were solved, achieving high-precision passive domain adaptive infrared ship semantic segmentation and improving segmentation accuracy.

CN122676178APending Publication Date: 2026-09-01BEIJING UNIV OF TECH
0 Cites 0 Cited by

Patent Information

Application Number
CN202610927932.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-25
Publication Date
2026-09-01

AI Technical Summary

Technical Problem

Existing technologies rely on a large amount of infrared pixel-level labeled data for semantic segmentation of ships, resulting in data scarcity and high labeling costs. Performance degrades during cross-modal recognition, and passive domain adaptation methods do not fully utilize brightness information and pseudo-label optimization, resulting in noise and limited segmentation accuracy.

Method used

Based on the brightness-guided and historical feature alignment method, this method utilizes prior brightness information and historical feature alignment technology to achieve high-precision segmentation without infrared labeled data, building upon a pre-trained model in the visible light domain. This includes brightness region division, historical feature extraction, and multiple forward propagation to generate structured pseudo-labels. The model is then updated using hierarchical supervision and feature alignment loss.

Benefits of technology

Without requiring infrared labeled data, it reduces boundary classification confusion, suppresses feature drift, improves the accuracy of infrared ship semantic segmentation, and overcomes cross-modal domain differences.

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

Abstract

This invention discloses a passive domain-adaptive infrared ship semantic segmentation method based on brightness guidance and historical feature alignment. The method includes: acquiring a pre-trained source segmentation model in the visible light domain and an unlabeled infrared target domain image; dividing the infrared target domain image into regions and selecting initial reliable regions, assigning them fixed category pseudo-labels; extracting current features, historical features, and category feature prototypes from the infrared target domain image; performing multiple forward propagations with perturbations on the infrared target domain image to obtain ensemble prediction results, combining the initial reliable regions to determine the consistency of the remaining regions, generating structured pseudo-labels and corresponding reliable masks; calculating the loss and jointly updating the current model based on the reliable mask, current features, historical features, and category feature prototypes; and implementing infrared ship semantic segmentation based on the updated model. This invention eliminates the need for infrared labeled data, effectively overcomes cross-modal domain differences, and improves the accuracy of infrared ship semantic segmentation.
Need to check novelty before this filing date? Find Prior Art