Adaptive Multi-column Deep Neural Networks with Application to Robust Image denoising
F.Agostinelli, M.R.Anderson,H.Lee, Advances in Neural Information, 2013, PP : 1-9
In this paper Adaptive Multi-column Stacked Sparse Denoising Autoencoder (AMC-SSDA )is used for image denoising. Denoising autoencoder(DA) is used to pre-train the network. The DAs are stacked to form Stacked DA (SDA) where each DA is trained by generating new noise. AMC- SSDA is the linear combination of several SSDAs and has 3 training phases.
(i) Training SSDAs :
Each SSDA is trained for a single type of noise, SSDA has the advantage of learning features which is generated by the activation of SSDA 's hidden layers .
(ii)Determining optimal weights for training images:
A new training is constructed to pair the features obtained from the hidden layers with the optimal column weights.The features from each column (f1,f2,...fc) are concatenated to form a whole feature vector (phi). The weights are constrained between 0 and 1 to avoid the problem of overfitting.
(iii) Training weight prediction module:
The feature vector (phi) is given as input to the weight prediction module to produce a weight vector(S) . The final denoised image is then produced by combining the columns using the weights.
This method is robust to various types of noise and provides better results for noise types that are present during training and noise types that are not seen by the denoiser during the training.
Software
(i) Training SSDAs :
Each SSDA is trained for a single type of noise, SSDA has the advantage of learning features which is generated by the activation of SSDA 's hidden layers .
(ii)Determining optimal weights for training images:
A new training is constructed to pair the features obtained from the hidden layers with the optimal column weights.The features from each column (f1,f2,...fc) are concatenated to form a whole feature vector (phi). The weights are constrained between 0 and 1 to avoid the problem of overfitting.
(iii) Training weight prediction module:
The feature vector (phi) is given as input to the weight prediction module to produce a weight vector(S) . The final denoised image is then produced by combining the columns using the weights.
This method is robust to various types of noise and provides better results for noise types that are present during training and noise types that are not seen by the denoiser during the training.
Software