Mirage Net: Low Calculation Cost Network for Image Denoising

Abstract

Benefit from feature presentation with huge parameters and high GPU computing resources, deep convolution neural network has been widely studied in image denoising due to its considerable denoising performance. However, these parameters will consume quantities of memory and computing resources, meanwhile, lots of them are correlated and redundant. We propose a low calculation cost and fast denoising convolution neural network, namely Mirage Net, inspired by the natural phenomenon of mirage. Based on our refraction convolution, which is the combination of depth-wise and point-wise convolution, Mirage Net can reduce parameter redundancy and learn effective presentations from one-layer deeper feature maps by cheap cost linear transformations which will be concatenated with previous feature maps as input of the next convolution layer. We also use alternating training strategy with multi-loss which accelerate the training processing and convergence rate. Our experiments on public datasets show that Mirage Net can achieve higher quality denoised images than DnCNN, and furthermore, the calculation cost is only half of them.