Super-resolution Reconstruction Methods Based On A Nested DID Structure And The Feedback DenseNet

The super-resolution reconstruction SR technology based on deep convolutional neural network has achieved excellent performance.However,the interpretability of SR convolutional networks for the selection of hierarchical features is still an open problem;at the same time,the construction method of the network that realizes the expression of feature diversity is not clear enough.Even though the dense network-based SR method utilizes the layered features of convolutional layers in dense blocks,the dense blocks in the network have a fixed depth and are still connected in a chain,without using the layered features of dense blocks.In addition,on the premise of ensuring the accuracy of image reconstruction,it is another open problem to speed up the reconstruction speed as much as possible.The research on the problem of feature diversity and the balance of accuracy and speed,the main contributions are as follows:1)Aiming at the feature diversity problem,we propose a nested dense network for image super-resolution reconstruction.First a residual dense block with variable depth(VRDB)construction method is proposed.The problem of overfitting of receptive fields and underfitting of large receptive fields.Second,construct a nested dense structure(Dense in Dense,DID),reuse layered features in a multi-scale sense,and obtain scale-scaling features to make up for the problem of inconsistent distribution of training data caused by the lack of BN.At the same time,the global dense connection strengthens the exchange of dense layered feature information between dense blocks,alleviates the gradient disappearance and gradient explosion in the case of a larger number of dense blocks,and accelerates the convergence.Experiments show that,compared with six advanced networks,the VRDB-based nested dense network proposed in this paper achieves better performance in the objective indicators PSNR/ SIMM and subjective visual perception.2)Aiming at the balance of speed and accuracy,we propose a feedback dense network for image super-resolution reconstruction.The network realizes the recursive timing of network nodes through feedback connection.There are three advantages: First,in each recursion,the network parameters are shared,the network parameters are few,the structure is simple,and the reconstruction speed is fast.Second,the feedback stream formed in the time dimension,which retains the high-level features of the previous moment and is used as feedback input,so this recursive structure of feedback is also powerful.Third,the intermediate SR images reconstructed in each iteration are used to calculate the network loss.The intermediate resultsof all iterations can guide the SR reconstruction process.Experiments show that,compared with four advanced networks with similar complexity,the dense feedback-based network proposed in this paper achieves better performance in objective indicators PSNR / SIMM and subjective visual perception.