Cascaded Convolution Neural Network for Color Image Recovery
Neng Li, Yuanyuan Deng · 2022
Reconstructing the high-quality image from its degraded version has attracted more interest in recent years. This data recovery problem can be first defined as an ℓ2norm minimization problem and then solved by deep learning techniques. In the paper, the task of color image recovery from partly observed gray scale data is tackle. It is assumed that some blocks or rectangular area of the gray scale is not observed, making the problem more complicated. The baseline convolutional auto-encoder network is first described. By dividing it into tasks of completion of missing values and image coloring, two sub-networks are proposed with similar architectures to solve the two sub-problems, and they are combined to get the final satisfying results. Experimental results shows that the proposed cascaded network can recover the image with higher PSNR and SSIM performance comparing to the baseline model.