An Image Denoising Algorithm Based On Image Quality Assessment

Abstract

The combination of image quality assessment (IQA) and conventional image processing algorithm or system can effectively improve the performance of the latter one. There are two ways of the combination. One is to use IQA as the pre-processing or post-processing part of the image processing algorithm or system. Another way is to embed IQA into it. In this paper, we investigate the latter one. As for the image processing algorithm or system, we use a denoising algorithm because image denoising is an indispensable step in many physical applications. We use a discriminative denoised algorithm which is called denoising convolutional neural networks (DnCNNs). We change its loss function, and add the IQA part. The original loss function of DnCNNs is 11 loss. It has clear physical meaning and simple calculation. However, it only represents differences in low-level features of images, ignoring human perception characteristics. The added IQA part can make up for the loss of human high-level perception, making the denoised result more in line with the human perception. The experiments show that changed DnCNNs has better ability to deal with overexposed image. In addition, the loss value of changed DnCNNs is steadier than origin DnCNNs. At the same time, the average PANR of changed DnCNNs is higher than origin DnCNNs.