Biomechanical analysis and application of an anti-fuzzy decomposition method for sports dance movement images based on multi-attention

Abstract

In the realm of biomechanics, accurate analysis of sports dance movements plays a crucial role in understanding human motion patterns and optimizing athletic performance. To address the challenge of analyzing sports dance error movement images that often suffer from high segmentation difficulty due to blurriness, we propose an anti-fuzzy decomposition method based on multi-head attention. Firstly, the Multi-scale Retinex automatic color enhancement algorithm is employed to correct the color of sports dance error action images, as color correction can enhance the visual clarity which is essential for subsequent biomechanical feature extraction. Subsequently, an anti-fuzzy decomposition model of sports dance error action images based on Diffusion Model-U-shaped network (DM-Unet) is constructed. The corrected image is divided into image blocks by a block embedding layer and then input into the encoder which is constructed by the confrontation generation network. The encoder selects the residual network not only to extract image features but also to capture biomechanically relevant details such as joint positions, limb orientations, and body postures. The multi-head attention module is utilized to suppress the dynamic blur of the image, which helps in precisely identifying the key movement elements during sports dance. Moreover, through downsampling operations, the dimension of image features is reduced while retaining the essential biomechanical information. The decoder then uses the up-sampling module to restore the encoder output results to the original size. Through the global residual connection module, the features of each layer of the encoder and decoder are connected, enabling the retention of shallow features of the image that are significant for analyzing the fine-grained biomechanical aspects of sports dance movements. The comprehensive loss function is used to train the model, and the anti-fuzzy decomposition results of sports dance error action images are outputted. The experimental results show that this method can effectively decompose the wrong action image of sports dance into structure and texture parts. Importantly, the peak signal-to-noise ratio of the decomposed image being higher than 26dB indicates enhanced clarity for further biomechanical analysis. For example, these decomposed images can be used to study the impact of different movement errors on joint torques, muscle activations, and overall body balance in sports dance, providing valuable insights for coaches and biomechanics researchers to improve training programs and understand injury mechanisms.