Recursive Dense Convolutional Neural Network And Its Applications

Super-resolution image reconstruction is a technique for recovering high-resolution images containing more information from the lowresolution images with less information.It has been successfully applied in medical imaging,image segmentation,target detection and so on.Based on deep learning and sparse representation theory,this dissertation discussed deep convolutional networks and their application and proposes the super-resolution reconstruction method based on bilayer deformable convolution network,recursive dense convolutional network,hyperspectral image super-resolution method using recursive densely convolutional neural network with spatial constraint strategy,and super-resolution image reconstruction method based on adaptive semi-coupled dictionary learning.The specific research contents are as follows:1.Most classical super-resolution image reconstruction methods based on deep learning do not consider the variations of scale and geometry.In order to solve this problem,this dissertation proposes a bilayer deformable convolutional network based super-resolution reconstruction.Firstly,this method replaces the standard convolutional layer with a proposed deformable convolutional layer to simulate the process of simple geometric deformation.Secondly,a bilayer deformable convolution unit is constructed using two different sizes of deformable convolution kernels to extract the features with different scales.Finally,residual connections are added among feature maps to alleviate the network training problem caused by the disappearance of the gradient.The experimental results show that the this method can extract the feature information of the image better than some existing reconstruction methods and improve the reconstruction effect of the image.2.In general,the performance of the network will increase if its depth increases.But the number of parameters will also increase rapidly,which requires a large amount of storage space and computing cost.This dissertation proposes an improved deep convolutional network,named recursive dense convolutional neural network to improve the parameter efficiency of deep networks.The recursive dense network still utilizes the dense block in the dense convolution network as a basic structure.But unlike the dense blocks connected in series in the dense network,it uses the output of the dense block as the input of the current block again.In other words,the recursive dense network converts the serial structure in the dense network into a recursive structure,which can reduce the number of parameters actually.Experiments show that the recursive dense network can still maintain or exceed the accuracy of some state-of-the-art deep networks with fewer parameters.3.In real applications,how to get high-resolution hyperspectral images is still a challenge.This dissertation proposes a hyperspectral image super-resolution method using recursive densely convolutional neural network with spatial constraint strategy.The method uses the proposed deep network to learn the mapping relationship from low-resolution hyperspectral image to the high-resolution hyperspectral image,and then applies the spatial constraint strategy to further improve the reconstruction effect of hyperspectral image.Due to the recursive structure,the recursive dense network can make the network deeper without increasing the number of parameters.In addition,the spatial constraint strategy can significantly improve the effect of reconstructed high-resolution images.The experiments show that our proposed method is superior to several state-of-the-art hyperspectral image super-resolution methods.4.Classical super resolution image reconstruction methods based on sparse representation usually depend on the suppose that the high and low resolution image patches have the same representation coefficient under their respective dictionaries,and consider the sparsity and cooperation of the representation coefficients separately.Aiming at these problems,this dissertation proposes a super resolution image reconstruction method based on adaptive semi-coupled dictionary learning.It is assumed that the high and low resolution image patches have linear mapping instead of the same representation coefficient,which relaxes the constraints and is more reasonable.In addition,using the kernel norm to construct a new regularized term to consider sparsity and synergy,it can adaptively balance the relationship between the sparsity and the cooperation according to the change of the dictionary,and generate a most suitable representation coefficient.The experimental results show that our method has better reconstruction effect than some existing dictionary-based reconstruction methods and has better robustness in the noisy environment.