論文閱讀筆記之——《DN-ResNet: Efficient Deep Residual Network for Image Denoising》
本文提出的DN-ResNet,就是a deep convolutional neural network (CNN) consisting of several residual blocks (ResBlocks).感覺有點類似於SRResNet的思路。並且對於訓練這個作者所提出的網路,作者還採用了edge-aware loss function
文章focused on denoising either Gaussian or Poisson corrupted images,
在影象去噪中,最常用的是高斯退化模型。the ith observed pixel is
低光散射噪聲導致的降級取決於訊號,並且通常使用泊松噪聲建模
當前的許多去噪網路都是建立LR和HR之間的mapping。而作者認為他們都不適用於實際的圖片,由於網路的size。為此作者This training strategy not only allows the resulting DN-ResNet to converge faster, but also allows it to be more computationally efficient than prior art denoising networks.
本文的貢獻有三:
1、證明了ResNet is effective for image denoising, and using edge-aware loss function signicantly improves the perceptive quality.
2、introduce the depthwise separable ResBlock (DS-ResBlock) to construct DS-DN-ResNet.
3、the proposed DN-ResNet works well for all types of noises, even without knowing the noise level.
寫本部落格之時~其實本人並不在乎作者用什麼方法,只是想知道作者是怎麼估算noise level的,或者他有沒有寫到關於noise level的描述