In the imaging process of CMOS image sensors, several kinds of noise will be introduced into the image. Most image denoising algorithms are developed for additive white Gaussian noise (AWGN). But the noise in the real image does not completely conform to a Gaussian distribution. The noise in the real image is complex and difficult to be modeled analysis. In this paper, a three-channel convolution neural network (TC-CNN) denoising method for real RGB image is proposed. The TC-CNN denoising method separates the real image to three images of each RGB channel. The convolution neural network is used for denoising each channel image. A new loss function and a new network architecture are proposed, this work makes the convolution neural network more suitable for denoising work. Experiment on real image datasets shows that the TC-CNN denoising method has better denoising result than the common denoising method.
展开▼
