deep learning Deep Belief networks for image denoising

http://arxiv.org/pdf/1312.6158v2.pdf

Deep belief network (DBN), 

          hierarchical generative models  for feature representation

          Features (extracted by DBNs) are presented as the values of the last layer nodes. 

Image denoise:

In the last later of DBNs, Distinguish between nodes presents noise and nodes presenting image content. 

        The nodes in the last layer are divided in two distinct groups of nodes. 

Detect nodes, then make the noise nodes inactive and reconstruct a noise-less image. 

        Use a criterion called relative activity to detect noise nodes.

Relative activity of each node is defined as the difference between two values of a particular node resulted from feeding the network using a noiseless image and its corresponding noisy image. 

传统的一些降噪方法：

         Fourier analysis, Spatial filtering, Wavelet transform, Sparse coding and dictionary learning, 

与DL相关的，另外两篇关于降噪的paper：

Image denoising and inpainting with deep neural networks

Convolutional networks and applications in vision