SURVEY OF DEEP NEURAL NETWORKS IN BLIND DENOISING USING DIFFERENT ARCHITECTURES AND DIFFERENT LABELS

Abstract

This thesis is aimed at providing a better understanding of using deep ConvNets in image denoising. Under a supervised setting, it is intuitive to use the ground truth as the label, and let the ConvNets learn to produce a latent image as the output. However we can also use the noise as its label, and subtract it from the noisy image to obtain the latent image. The first discussion would be the effectiveness and difference between using these two labels. Following that, we experiment between using the traditional ConvNets, and the recent ResNets. Since its introduction, ResNets has become popular due to its performance in the ImageNet competition. Its ability to reduce loss and improve accuracy of classification by simply adding more layers with skip-connections was the main reason for its growing popularity. With this hype, we wish to experiment its effectiveness in the denoising task, and similarly, we aim to examine its performance with using either truth or noise labels. Experimental results allow us to conclude that in terms of training loss, ResNets gives better performance regardless of the type of labels. However, in terms of PSNR for test data, the results are inconclusive, as ConvNets perform better.