BS-HRNet: A Multi-Resolution Enhanced Bird Pose Estimation Method Based on Square High-Resolution Network

This paper addresses challenges in bird pose estimation—variable target scales, weak spatial perception of slender parts, and feather occlusionby proposing BS-HRNet (BirdSquareHRNet), an enhanced high-resolution network. Key innovations include: (1) Reconstructing HRNet's “inverted pyramid” into a square structure with uniform feature extraction layers across resolution branches, reducing parameters by 12% while enhancing multi-scale feature interaction; (2) Embedding a coordinate attention (CA) mechanism in the input layer to boost spatial localization of slender parts (e.g., beaks, tails) via orientation sensitivity; (3) Integrating multi-scale feature convolution (MFC) before the output layer to fuse local-global features through reiteration feature extraction; (4) Employing Online Hard Keypoint Mining with Mean Squared Error loss (OHKMSE) to prioritize occluded areas (e.g., feathered elbows/wrists). Experiments on the Animal Kingdom-Birds dataset demonstrate that BS-HRNet achieves improvements of 3.7% in PCK@O.05 and 3.4% in OKS, 12% parameter reduction, and 18% lower computational complexity, while significantly enhancing accuracy for slender and occluded keypoints.