基于高阶奇异值分解和深度学习的光场图像拼接方法

光场图像拼接旨在提高光场图像的视场角。考虑到光场数据包含较多冗余，且传统拼接方法对于低纹理场景的光场图像鲁棒性不足，本文提出一种基于高阶奇异值分解(high-order singular value decomposition, HOSVD)和深度学习的光场图像拼接方法。首先，通过光流估计和HOSVD对光场图像进行降维，得到所有视角下一致空间信息的主基和不同视角下高频信息的其他基带。其次，提出注意力增强的无监督单应性估计网络来提高图像的配准精度。最后，将扭曲后的参考基带和目标基带进行光场重建与图像融合，得到最终的拼接光场。实验结果表明，该方法在拼接光场的主客观质量和角度一致性方面表现出较好的性能。

Light field image stitching method based on high-order singular value decomposition and deep learning

The purpose of light field image stitching is to improve the field of view of the light field images.Considering that the light field data contains more redundancy,and the traditional stitching method is not robust enough for the light field image of low-texture scenes,a light field image stitching method based on high-order singular value decomposition (HOSVD) and deep learning is proposed in this paper.Firstly,optical flow estimation and HOSVD are used to reduce the dimension of the light field images to obtain the principal basis containing consistent spatial information from all viewpoints and other basis containing high frequency information from different viewpoints.Secondly,an attention-enhanced unsupervised homography estimation network is proposed to improve the accuracy of image registration.Finally,the warped reference basis and target basis are used for light field reconstruction and image fusion.The final stitched light field images are obtained.Experimental results show that the proposed method performs well in subjective quality,objective quality and angular consistency.