Research On Super-resolution Image Reconstruction Algorithm Based On Improved Convolutional Neural Network With Image Texture Segmentation

Improving the resolution of image is very important in many image processing tasks.But some difficult technical problems may be faced with in hardware methods.Therefore,it is necessary to improve the resolution by software methods.Super-Resolution(SR)image reconstruction is a method to reconstruct a High Resolution(HR)image by using one or more Low Resolution(LR)images in the same scene.This method uses signal processing to deal with the LR image and efficiently improve the resolution of the image and satisfies people's requirements of visual.Convolution Neural Network(CNN)is a typical deep learning model.CNN can learn the extracted features repeatedly and get the complicated features because of its hidden layer.Coupled Deep Autoencoder(CDA)is a CNN model with deep structure.Using this model for SR reconstruction,CNN is regarded as encoder and decoder and LR image is encoded as ‘symbols' by Encode Network,then the LR ‘symbols' are input into Mapping Network and output the HR ‘symbols',the HR symbols are decoded as a HR image by Decode Network at last.As described,CDA has achieved outstanding results in Single Image Super Resolution(SISR).Research is mainly focused on the CDA and an improved CDA method is proposed in this thesis.Skip Connection(SC)structure is attached to the En-Decode Network to ensure the depth;Network in Network(NIN)structure is attached to the Mapping Network to enhance the mapping ability;To shorten the training time,the residual image is used as the output instead of the HR image.In this thesis,pre-training is adopted.Firstly,LR-HR image pairs are used to train the Low-resolution Autoencoder(LRAE)and Residual Autoencoder(ReAE),so that their input and output are both LR images and residual images respectively.Then the Mapping Network is trained by LR symbols and residual symbols so that it can map the former into the latter accurately.After that,Encode Network,Mapping Network and Decode Network are combined into a whole SR Network.Finally,the weights of SR Network are fine-tuned by back-propagation and bicubic interpolated image is added to the residual image to get the improved CNN for SR.Image texture reconstruction is the critical problem in the SR task.Therefore,an improved SR algorithm which based on improved CNN with image texture segmentation is proposed in this thesis.A better texture mask is got by an improved method which based on the Relative Total Variation(RTV).Then the texture region is extracted from an image efficiently,and this region is reconstructed by improved CNN method in order to reduce the reconstruction time and increase the efficiency.The experimental results show that the proposed algorithm can shorten the time effectively and the quality of image reconstruction is preserved.