Improving unsupervised image denoising via ordinal regression

Abstract

In recent years, deep neural network (DNN) have achieved great success for various computer vision fields including image denoising. Unsupervised image denoising models that does not require a clean image outperforms the classical denoising methods. However, Current unsupervised denoiser still suffers from a serious problem that the denoised image looks blurry because denoisers cannot restore the clean edge of the images. This problem caused by the receptive field of the model takes less information from the noisy image. In this paper, we suggest a novel unsupervised denoising model by improving the learning algorithm. We adopt the ordinal regression task to the network which can supply more information of the input image. The proposed model outperforms the Noise2Self [3] on the benchmark datasets, such as Set12 and BSD68 [5] and Urban100.