Deep Residual Optimization for Stereoscopic Image Color Correction

The color correction algorithm is designed to eliminate color discrepancies between image pairs. Compared with the conventional algorithm, color correction for 3D stereoscopic images not only needs to achieve the color consistency of the resulting image and the reference image but also expected to ensure the structural consistency of the resulting image and the target image. For this problem, we propose a stereoscopic image color correction algorithm based on deep residual optimization. First, we get an initial result image by fusing a global color correction image and a dense matching image of the stereo image pair. Then, a residual image optimization scheme is used to improve the structural deformation and color inconsistency of the initial result caused by mismatching and fusion. By combining the target image with the optimized residual image, the structure and clarity of the target image can be retained to the maximum extent. In addition, we use the perceptual loss and per-pixel loss to improve the structural deformation and local color inconsistency while training the optimization network. Experimental results show the effectiveness of our method.