Abstract: HPE(Human Pose Estimation) is a fundamental task in computer vision,aiming to accurately localize human keypoints and understand body structure,which is crucial for downstream tasks such as action recognition and detection.Although deep learning has driven significant progress in HPE,existing methods still struggle to effectively handle challenges like large scale variations,occlusion,and loss of details in complex scenarios such as dense crowds and dynamic movements with large pose changes.To address these issues,this paper proposes an improved architecture,EFW-HRNet,which fuses DWT(Discrete Wavelet Transform) with the HRNet(High-Resolution Network).It introduces DWT-based downsampling and feature fusion modules to capture and preserve multi-scale details.It designs a CBA(Cross Band Attention) module to enable adaptive interaction among DWT sub-band features and enhance robustness against occlusion.And it applies a FBCC(Frequency Band Channel Compression) strategy to compress high-frequency channels,significantly reducing computational redundancy and improving model efficiency.Experiments on the COCO dataset show that EFW-HRNet achieves a significant AP increase of 4.0 percentage points compared to the strong baseline UDP HRNet-W32.Ablation studies validate the effectiveness of the DWT,CBA,and FBCC strategies,where FBCC achieves a good trade-off between accuracy and efficiency,sacrificing only about 0.8 percentage points AP in exchange for a substantial reduction in parameters by about 66% and computational cost by about 51%.

Key words: Human pose estimation, High-resolution network, Discrete wavelet transform, Frequency band channel compression, Attention mechanism