Paper notes-Wide Inference Network for Image Denoising viaLearning Pixel-distribution Prior

Wide Inference Network for Image Denoising viaLearning Pixel-distribution Prior

1. Main task：
       In this paper,the author propose an innovative strategy for image denoising by using convolutional neural networks to learn similar pixel-distribution features from noisy images. The key of approach is a discovery that wider CNNs with more convolutions tend to learn the similar pixel-distribution features,which reveals a new strategy to solve low-level vision problems effectively that the inference mapping primarily relies on the priors behind the noise property instead of the deeper CNNs with more stacked nonlinear layers.The network named Wide Inference Networks(WIN),WIN based network consistently achieves significantly better performance than current state of the art deep CNN-based methods in both quantitative and visual evaluations
2. background：
       They provide background on distributions of image noise and sptial feature,they show that CNNs can be modeled and constructed with stronger capability to extract the spatial feature distribution,so the deep CNN-based image denoising methods have a function of regularization techniques,and learning strategies for generalization. In the sequel,they show that CNNs can be modeled and constructed with stronger capability to extract the spatial feature distribution,which can then be used as a prior for image denoising.
   
   they introduce some CNNs structure.the deepCNN-based state of the art denoising models:DnCNN and RED-Net
   2.1  DnCNN

1. DnCNN adopts a 20 layers deep architecture,a learning strategy of residual learning,and a regularization method of batch Normalization
   2.2 RED-Net
   RED-Net employs a 30 layers deeper structure with skip connections added to connect corresponding layers of convolution to the deconvolution,and it is justified by the residual network.In deep structures,learning strategies and regularization methods also work for accelrating the learning proess and boosting performance.
   

1. Main solution：

1. The Wid Inference NetWork(WIN) is based on the plain convolutional neural network architecture with an exploration of the impact of the "width" of the network on the low-level vision tasks such as image denoising.
   3.1 WIN's structure
   3.1.1 the number of Filters
       WIN with the number of layers L=5 outperforms both shallower or deeper networks.
   3.1.2 the number of Filters(K)
       WIN with K=128 achieves remarkable performance gains than the network with K=64,but K=128 is founded to be the optimal value for high performance in this denoising task.
   3.1.3 the size of Filters
       F = 7*7 is able to improve performance remarkably compared to smaller F.In general,large F leads to better performance.larger F also dramatically increases both training time and computation complexity
   
   
   

   3.2 WIN5-R:WIN5+Residual Learning
   

   
   

   3.3 WIN5-RB:WIN5+Residual Learning+Bath Normalization
   

   
   

   3.3 The performance
   
   

   4. Some of my own ideas

In fact,I think this theory is not a lot of work,just adjust the parameters,and then I think it is wrong.