The paper introduces a novel denoiser network CFMNet which has novelty for utilizing noise level map with only shifting modulation when denoising on U-Net. CFMNet reflects that fact the introducing multiple layer of CFM modules with residual network improves the network performance with direct estimation. Unfortunately, the improvement is not guaranteed to kernel prediction.
FFDNet addresses a limitation where the noise level map is modulated and only concatenated once before the first layer as an input.
CFMNet also has its own U-net for noise level map. The noisy image and the noise map propagates to each U-net. However, for every downsampling & upsampling two pass of RS-CFM modules exists.
RS-CFM modules is consisted with 1. residual connection and 2. generating shifting map which both leads to residual shifting operation. For noisy map feature, the noisy image feature is concatenated and pass through convolutional layers. Then one layer outputs the shifting map, which will shift the noisy image feature with residual connection.
The reason why the paper only chose to modulate with shifting operator instead of scaling is because shiting plays a huge role finding balance between preserving details and denoising. When concatenating noise level map, the same variance is fed to the 2D feature. When large variance means large noise, the intention of only using shifting is to let the network learn the balance by reducing the effect of noise.
On gray scale image, the CFMNet outperformed other SOTA denoising networks with preserving textures and stalks better. Also on color images, CFMNet shown better denoised results compared to FFDNet. Also by Ablation studies the paper shown each CFM, residual connection, and convolution activation contribute to the performance of CFMNet. However, the fact was that all three combined increased the running time. Due to two UNet, CFMNet requires more parameters, but with quite reasonable running time.
