부산대학교 도서관

Human pose estimation (HPE) depends on the inter-dependence among body parts for consistency. Existing methods use images of different resolutions to extract features at various scales, which may cause context information conflicts and long-range dependency issues, leading to inaccuracies. To address this, we propose SHGNet, an improved algorithm based on a single-layer hourglass structure. SHGNet extracts local and global features using a pyramid structure and introduces a feature enhancement module for context enhancement and long-range dependency resolution. By fusing features from multiscale convolutions, improving context information representation, and employing a sparse multi-layer perceptron (sMLP) to capture long-range dependencies. Finally, by fusing the convolutional layer and the output layer with attention mechanism in a weighted manner, we achieve a better balance between local and global information, obtaining more refined global image features. Additionally, we introduce a channel purification module for more accurate feature capturing. When applied to the YOLOPOSE pose estimation model and tested on the MSCOCO keypoints validation dataset, our method achieves a 1.4% AP increase and a 1.7% APL increase compared to the original YOLOPOSE baseline, demonstrating enhanced accuracy and performance.