"基于多尺度注意力机制的三维人脸建模网络"

The article "Three-dimensional Face Modeling Based on Multi-Scale Attention Mechanism Phase Unwrapping" introduces a novel approach to improving the accuracy of 3D modeling through phase unwrapping in complex scenes. The precision of phase unwrapping directly impacts the accuracy of 3D modeling, and traditional spatial phase unwrapping methods often struggle with issues such as undersampling and discontinuous phase, while temporal phase unwrapping requires additional information. In response to these challenges, the article proposes a phase unwrapping network based on a multi-scale attention mechanism. The proposed network utilizes an encoder-decoder structure to integrate multi-scale features and embeds an attention sub-network in the decoder to obtain context information. The authors also constructed two datasets for the training and testing of the phase unwrapping network: a FACE dataset containing 5000 sample groups and a MASK dataset containing 100 sample groups, each with truncated phase and continuous phase ground truths. The results of the phase unwrapping network on the FACE dataset and MASK dataset demonstrated promising performance, with root mean square errors of 0.0387 rad and 0.0273 rad, and structural similarity indices of 0.9850 and 0.9793, respectively. Furthermore, the network proved to be capable of quickly and accurately extracting phase features in undersampled and discontinuous regions, ensuring the correctness of phase unwrapping. In conclusion, the effectiveness and feasibility of the proposed network were validated through comparative experiments. The findings of the article hold significant implications for advancing the field of 3D modeling and could potentially contribute to improvements in various applications such as 3D scanning, biometrics, and computer vision.