Deep Convolutional Architecture for Natural Image Denoising
X.Wang,Q.Tao,L.Wang,D.Li,M.Shan, International conference on Wireless Communications & Signal Processing, 2015
Image denoising is done by using deep convolutional architecture with a modification of setting the sampling rate of the pooling layers. The denoised gray image is obtained by corrupting the clean gray image with AWGN. Network takes noisy image as input and the features are extracted by convolving with a set of small convolution kernels. After a full connected convolution with final hidden layer the corresponding denoised image is obtained.
ReLu is used as activation in the hidden layers and tanh is used as activation function in the output layer to get a normalized output. The learning rate is set to 0.001 for the final layers and 0.1 for other layers.
ReLu is used as activation in the hidden layers and tanh is used as activation function in the output layer to get a normalized output. The learning rate is set to 0.001 for the final layers and 0.1 for other layers.