The Algorithm Research Of 3D Hand Mesh Reconstruction From A Single Low-Resolution Image

3D human body reconstruction has been a research hotspot in the field of computer vision,with extensive applications in virtual reality,human-computer interaction,and motion capture.As an important component of human reconstruction,hand reconstruction has attracted increasing attention.However,in practical applications,hand images often have low resolution,taking up only a small fraction of the entire human body image.Most existing hand reconstruction studies are conducted independently,with little consideration of this situation.To address this issue,the deep learning method is adopted in this thesis,focusing on hand reconstruction from a single low-resolution image.The main research contents are as follows:(1)Analysis of hand reconstruction under low-resolution conditions; exploring the influence of image resolution and background information on hand reconstruction and demonstrating through experiments how they affect hand reconstruction; investigating the parametric and non-parametric models for hand representation to provide insights for subsequent algorithms.(2)Introducing a multi-task hand reconstruction algorithm based on the MANO model.By utilizing the MANO hand parameter model and combining it with multi-task learning,this method achieves the reconstruction of hand meshes and 3D joints.The proposed approach extends low-resolution images to multiple resolutions,employs hand contours to remove background interference,and converts them into high-dimensional features.Through the integration of features using a Transformer Encoder,the designed CrossAttention module ensures the preservation of resolution-specific features.The algorithm simultaneously predicts 3D joint coordinates and MANO model parameters to accomplish hand reconstruction.Comparable results to existing methods have been achieved on the low-resolution Frei HAND dataset,and hand reconstruction has also been successfully performed on the low-resolution RHD and HO3 D datasets.(3)Introducing a low-resolution hand reconstruction algorithm based on UV-Map.By utilizing the UV-Map representation of hand shape,this method compensates for the loss of reconstruction details caused by the use of parametric models,thereby improving the accuracy of the reconstruction results.The proposed approach generates hand images at different resolutions using a super-resolution network and segmentation algorithm and generates multiple UV-Maps using the UV-Net network.The UV-Maps at different resolutions are gradually fused using learnable weights.The fused UV-Map is refined to obtain a smooth 3D hand shape.Compared to existing methods such as METRO and Mob Recon,this algorithm achieves superior performance and higher reconstruction accuracy on the low-resolution Frei HAND dataset.Both algorithms proposed in this thesis achieve 3D hand reconstruction in real-world low-resolution scenarios.