![]() ![]() Unless the network has a sufficiently large receptive field it will not be able to find pattern in the noise. Performing denoising task without prior information about noise requires learning noise from the image. A shallow network (like SRCNN) is not be able to perform this task because small number of layers (less number of trainable parameters) are insufficient to capture fluctuations due to noise level and type. The network has sufficient depth to learn the pattern of noise from the data itself. We can conclude (from the results) that the network proposed is equally good for denoising, super-resolution and combination of both tasks. The last layer is Conv with c filters of size 3x3圆4. The second to second last layers are Conv+BN+ReLU with 64 filters of size 3x3圆4. The network they used consists of 17 layers(in case of white Gaussian noise) or 20 layers(in the case of blind Gaussian noise).The first layer is Conv+ReLU with 64 filters of size 3x3xc, where c is the number of channels. The use of residual learning allows to train a single CNN model to tackle several tasks such as Gaussian denoising, single image super-resolution and JPEG image deblocking. Rather than being limited to additive white Gaussian noise (AWGN) at a certain noise level, their model is capable of performing Gaussian denoising with unknown noise level (i.e.,blind Gaussian denoising). They construct a feed-forward denoising convolutional neural network (DnCNN) to learn the residue and use batch normalization to speed up the training process as well as boost the denoising performance. propose a plain discriminative learning model to remove noise from an image. We trained our network on data set that not only contained patches from different images but also each patch had different (randomly chosen) noise level and noise type. It is worth mentioning that when we compare our model with the models proposed earlier we use same training data and same training time. Moreover, in some aspects the performance of our network is better than some other deep learning models available for similar task. This structure gives it capability to learn different noise types of varying level. We have intentionally kept small convolution kernels and repetitive layer design. The proposed CNN though deep enough for our task at hand, is quite simple in architecture. We name the proposed model super-resolution denoising convolutional neural network (SuRDCNN). The aim of this system is to remove noise and then enhance image resolution (i.e. This residual image is difference of input image and the original image. Instead of learning end-to-end mapping we train the network to output a residual image. In our work, we develop a single network capable of performing image super-resolution and denoising. ![]()
0 Comments
Leave a Reply. |