时序显著性增强的Transformer视频小目标检测方法OA

To address the problems of insufficient single-frame information,weak target texture,strong back-ground clutter and underutilized temporal information in small object detection from video streams,a Transform-er-based video small object detection method enhanced by temporal saliency is proposed.Taking the real-time detection Transformer(RT-DETR)as the baseline,a unified temporal enhancement mechanism is constructed from three aspects:backbone feature extraction,global semantic modeling,and multi-scale feature fusion.Spe-cifically,a high-order differential-accumulation temporal fusion module is designed in the backbone to jointly model multi-frame accumulation,first-order and second-order differential information,and generate a temporal saliency map to strengthen the persistent response of targets and the representation of motion trends.In the en-coder,a temporal saliency guided attention-based intra-scale feature interaction(AIFI)module is proposed to convert the temporal saliency map into attention bias,guiding high-level feature modeling to focus on motion-sensitive regions.In the neck,a lightweight cross-scale context fusion module and a temporal saliency guided FPN/PAN aggregation strategy are introduced to continuously propagate and utilize temporal information during multi-scale fusion.Experimental results on DAUB,DMIST and XS-VID datasets show that,when the intersec-tion to union ratio of the proposed method is 0.5,the average accuracy is improved by 1.6％,2.4％,and 1.6％respectively compared to the baseline modelwith only 5.53× 106 additional parameters,which validates the ef-fectiveness and generalization ability of the temporal saliency enhancement strategy.