Paper notes-Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric

Paper notes-Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric 

1. Main task

    They propose a very deep fully convolution encoding-deconding frame-work for image restoration such as denosing and super-resolution . The network is composed of multiple layers of convolution and de-convolution operators.The convolution layers act as the feature extractor,which capture the abstraction of image contents while eliminating noises.De-convolution layers are then used to recover the image details.They propose to symmetrically link convolutional and de-convolutial layers with skip-layer connections,its training converges much faster and attains a higher-quality local optimum.

    First,the skip connections allow the signal to be back-propagated to bottom layers directly,and tackles the problem of gradient of vanishing,making training deep networks easier and achieving restoration performance gains consequently.Second,these skip connections pass image details from convolution layers to de-convolutional layers,which is beneficial in reconvering the original image.They declare the model can handle different levels of noises using a single model.

2,background

    Recently,deep neural networks have shown their superior performance in image processing and computer vision tasks,They oberve that in order ....

    very deep network architecture,which consists of chain of symmetric convolutional and deconvolutional layers.The convolutional act as the feature extractor which encode the primary components of image contents while eliminating the corruption.The deconvolutional layers then decode the image abstraction to recover the image content details.

    skip connections.......

    can handle different level of noises

    experimental results hava a good work

3,Main solution

    RED-Net is a framework of fully convolutional and deconvolutional.

    The network contains layers of symmetric convolution and deconvolution.Skip shorcuts are connected every a few,in fact,it just use two shorcut to connect the input-output layer and convolution - deconvolution layer.The response from a convolutional layer is directly propagated to the corresponding mirrored deconvolutional layer,both forwardly and backwardly.

    The last output layer is the clean version of the input image,for low-level image restoration problems,they use neither pooling and unpooling in the network as usually pooling discards useful image details that are essential for these tasks.It set kernel size is 3*3 and performance well.The input and output of the network are image of the same shape,They using 64 feature maps for convolutional and deconvolutional layers satisfactory results.They propose two networks,which are 20-layer and 30-layer respectiverly.

4,Training

5,Testing

This can lead to slightly better denoising and super-resolution performance.

6, some of own my ideas

The RED-Net's idea is novel,and performance well.