Archery Human Pose Estimation System Based on DFPDeblur GAN Algorithm

An Archery human pose estimation system based on Deep Feature Prior Deblurring Generative Adversarial Network (DFPDeblur GAN) is proposed. The system incorporates dynamic convolution, a technique that can adaptively adjust the parameters of the convolution kernel to extract multidimensional features, and a bidirectional Feature Pyramid Network (FPN), which effectively improves the image deblurring effect and multiscale feature fusion capability. Subsequently, the Improved High-Resolution Network (IHRNet), combined with the Coordinate Attention (CA)-mechanism, a mechanism that improves the accuracy of key point detection by focusing on the spatial location, is used to realize the precise localization of the key joints of the Archery action. The experimental results show that the proposed model achieves 94.3% key point localization accuracy, 9.1% Peak Signal-to-Noise Ratio (PSNR) improvement compared with the traditional method, up to 0.92 Structural Similarity Index (SSIM), and less than 2 seconds running time, which exhibits good real-time performance and robustness. The results show that the model performs well in a variety of lighting conditions and action phases, providing effective technical support for action analysis and training in Archery