基于原型增强与时序误差学习的红外可见光融合视频目标分割方法OA

RGB-thermal fused video object segmentation(RGB-T VOS)can effectively address the segmentation challenges in complex environments by combining the fine-grained texture information from visible images with the target saliency cues provided by thermal imagery.However,most existing RGB-T VOS fusion paradigms adopt an open-loop strategy,in which the fusion decisions for the current frame rely solely on the instantaneous feature inputs,while ignoring the temporal continuity and the model ability of the contribution variations across different modalities over time.To address this limitation,in this paper,a closed-loop RGB-T fusion framework based on prototype enhancement and temporal error learning is proposed,which is built upon an explicit attention-enhanced fusion architecture.Specifically,a temporal prototype enhancement mechanism is introduced to leverage the target feature statistics from the previous frame,thereby strengthening the feature representation of the target regions in the current frame.In addition,a learnable error encoder is introduced to compensate for the prediction deviations arising during the fusion process,enabling the adaptive correction of channel-wise gating and spatial attention decisions.The experimental results show that the proposed method achieves the average values of the Jaccard index(J)and F-score(F)of 87.7 and 88.6 on the VisT300 and VT-UAV datasets,respectively,outperforming existing methods.These results validate the effectiveness of the proposed closed-loop fusion strategy in improving temporal stability and segmentation accuracy.