In recent attempt in this direction of de-noising learning
DnCNN, FFDnet, CBDnet these three think it should be a series of very close contact, is the gradual generalization, consider gradually increasing the noise of a complex process, DnCNN mainly for de-noising Gaussian noise, emphasizing the role of the residual learning and BN, FFDnet generalized Gaussian noise is considered a more complex real noise, the noise level of the input part of the network of FIG, CBDnet is the noise level for starting the main portion of FIG FFDnet by FCN adaptation layer 5 is obtained from the noise level map, to achieve a certain blind on the degree of de-noising.
SRMD different from the previous three, largely from bicubic, consider the impact of nuclear blur and noise level, the LR, the blur kernel, unified network input noise levels to achieve the restoration of degraded for different models, I think this point there may still further space, SRMD needs of a given nuclear blur and noise level, it can not pass a similar form of adaptive CBD? Can be achieved on a large scale, any kind of blur kernel be restored? And then found DPSR essay seems to extend SRMD can be regarded as two days to read the following.
Specific paper notes when I learned the main reference following these three, I think written in great detail would not write in their own typesetting.