基于深度学习和自注意力的光度立体方法

Photometric Stereo Method Based on Deep Learning and Self-Attention Mechanism

针对基于深度学习的非标定光度立体方法,设计了一种基于自注意力和多重特征融合的网络模型。该模型在光照估计网络中引入了自注意力机制,用于帮助网络理解图像长距离像素间的依赖关系,提升网络对图像深层特征的感知能力。同时,为了提升在多图像输入时的特征融合效果,设计了一种基于多重最大池化和残差模块的法线恢复网络。该方法在DiLiGenT光度立体数据集上测试的光源方向和法向的平均角度误差分别为3.2和8.5。

In response to deep learning-based uncalibrated photometric stereo method,this paper proposes a network model based on self-attention and multi-feature fusion.This model introduces a self-attention mechanism into the illumination estimation network to help the network understand the long-range pixel dependencies in images,enhancing the network's perception of deep image features.Additionally,to improve the feature fusion effect when multiple images are input,a normal recovery network based on multiple max-pooling and residual modules is designed.The proposed method achieves average angular errors of 3.2 for light source direction and 8.5 for surface normal on the DiLiGenT photometric stereo dataset.