A Study On Several Algorithms For Image Reconstruction

Image super-resolution reconstruction(SR)is an important branch of image processing.Its research and development have wide application value.Image super-resolution reconstruction can be divided into two directions: single image super-resolution reconstruction(SISR)and multi-frame image super-resolution reconstruction(MISR).This dissertation mainly studies SISR algorithm based on sparse representation and deep learning.Image super-resolution reconstruction based on sparse representation is the representative of traditional methods,which has been improved by quite a number of papers,but it still has some room for improvement.Deep learning is a method developed with the improvement of computer hardware computing ability.Its application effect in the field of super-resolution reconstruction is very outstanding,and some algorithms have reached a high level.We carry out some of them to improve the reconstructed effect.At the same time,we try to reduce the capacity of the network to adapt to the application of mobile devices.Based on sparse representation theory and deep learning theory,three research works are carried out in this dissertation: super-resolution reconstruction based on cascade of convolution sparse coding,neighborhood regression and self-examples,super-resolution reconstruction based on deep dilation convolution neural network and super-resolution reconstruction based on lightweight MobileNet.The details are as follows:1.In recent years,a series of cascade based SISR methods have been proposed.However,the existing methods just attempt to connect a single method in series and improve the performance of super-resolution layer by layer,so that only the advantages of a single method can be used.In order to utilize synthetically the advantages of different methods,this paper proposes a new cascade method for SISR called different traditional methods cascade(DMC)and its improved version(DMCself)added self-examples layer.In the first layer,convolutional sparse coding(CSC)is used for reconstructing the whole low-resolution(LR)image directly.The second layer adopts adjusted anchored neighborhood regression(A+)to reconstruct the LR image patch by patch,and forms the whole SR image by averaging the pixels in the overlapped area.The last layer is a self-examples layer,which aims to take advantages of both internal and external statistics information.Contrast experiment shows that different methods can really improve more than a single method.2.Recently,the 20-layer convolutional network was adopted as the basic framework,and the very deep convolution network was used to reconstruct the high-resolution(HR)image well.In order to expand the perception field while the filter parameters remain unchanged,we introduce dilated convolution layers to it.Firstly,we analyze the perception field of dilated convolution blocks with different expansion coefficients combination,and a better structure is selected.Then,stacking the convolution blocks to form a deep convolution neural network.Finally,the network is retrained by some training techniques.The results show that the effect of network reconstruction could be improved by 0.1-0.2dB for the larger scaling factors of “Set5” and there are also clear visual advantages.In addition,in order to make the network converge better,we add the residual connection in the network and the reconstruction effect is further improved.3.In order to adapt to the application of mobile and vehicle-mounted equipment,miniaturization of network model has been paid more and more attention.Based on ESPCN,we attempt to introduce MobileNet network structure into image super-resolution reconstruction.By decomposing the standard convolution network into depthwise convolution and pointwise convolution operation,the network reduces the number of parameters and calculation to about 1/4 of the original.The results show that the reconstruction effect is better at all scale factors except the scale factor of ×2.We also attempts to use MobileNet v2 network structure to improve the network,the effect is further improved and exceeds all the comparison methods with small increasing the number of parameters and computation.When the scale factor in the data set “Set5” is ×4,the PSNR can be increased by 0.23 dB.In addition,the contrast experiments show that using the mean absolute value error(MAE)as the loss function can get better results than the mean square error(MSE)in this network.These two networks not only have better effect on qualitative indexes,but also have certain advantages in vision.