A High-Resolution Network Based on Feature Redundancy Reduction and Attention Mechanism

Model lightweighting is an essential aspect of computer vision tasks. We found that HRNet achieves superior performance by maintaining high resolution through multiple parallel subnetworks, but it also introduces many unnecessary redundant features, resulting in high complexity. Many methods replace modules in the backbone network with lightweight ones, often based on MobileNet, Sandglass modules, or ShuffleNet. However, these methods often suffer from significant performance degradation. This paper proposes GA-HRNet, a lightweight high-resolution human pose estimation network with an integrated attention mechanism, built upon the HRNet framework. The network structure of HRNet is restructured with reference to the G-Ghost Stage, introducing GPU-efficient cross-layer cheap operations to reduce inter-block feature redundancy, significantly lowering complexity while preserving high accuracy. Moreover, to compensate for the accuracy loss due to reconstruction, an attention module is designed and introduced to emphasize more important information in the channel and spatial dimensions, thereby improving network precision. Experimental results on the COCO and COCO-WholeBody datasets demonstrate that the proposed GA-HRNet is lighter and yet more accurate than HRNet. Moreover, it outperforms several state-of-the-art methods in terms of performance.