Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Skip Connections

Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections

Papers:
Nips:https://papers.nips.cc/paper/6172-image-restoration-using-very-deep-convolutional-encoder-decoder-networks-with-symmetric-skip-connections.pdf
arXiv:https://arxiv.org/pdf/1606.08921v3.pdf

1.Questions

Can a deeper network in general achieve better performance?
Can we design a single deep model which is capable to handle different levels of corruptions?
When the network goes deeper or using operations such as max pooling, too much image detail is already lost.

2.Solutions

3.Details

The size of input image can be arbitrary.
Deconvolution is better than convolution in up-sampling layers.
Use down-sampling in convolutional layers to reduce the size of the feature maps.
Add skip connections between two corresponding convolutional and deconvolutional layers.
Element-wise skip connections is better than sequential blocks.
Corrupted images on multiple orientations are proposed to achieve smoother results.

4.Conclusions

Deep convolution denoising autoencoder with skip connections.