Research On Image Resolution Enhancement Method Based On Deep Learning

In practical applications,people often cannot obtain high-resolution images due to limitations of imaging equipment.Improving the resolution of images by improving hardware devices is costly.Image resolution enhancement of an image refers to the reconstruction of a high-resolution image by supplementing the details of the image from a low-resolution image.Resolution enhancement method can improve image quality without modifying the image hardware by using software algorithms.The cost of this technology is low,and it has broad prospects for development in many fields such as satellite remote sensing,medical image,video surveillance,etc.In this thesis,we first introduce the traditional image resolution enhancement algorithms and deep learning based image resolution enhancement methods.Besides,we analyze their advantages and disadvantages.According to the advantages and disadvantages of the resolution enhancement methods,this thesis first proposes a deep cascaded structure for DC T domain image up-sampling.Then,for the image interpolation problem,this thesis proposes an image interpolation method based on deep recursive network.Although both image up-sampling and image interpolation aim for image resolution enhancement,there are some differences between them.In the task of image up-sampling,the observed low-resolution images are obtained by down-sampling high-resolution images.During the down-sampling process,all the pixels in the high-resolution images were changed.However,in the task of image interpolation,the low-resolution images are obtained by extracting low-resolution pixels from the original high-resolution images.The low-resolution details of the original high resolution image are exactly the same as those of the low-resolution image.For the deep learning based DC T domain up-sampling method,due to the de-correlation characteristics of DCT coefficients,the high-frequency DCT coefficients are first estimated by using the residual fully connected network,followed by a fully connected conversion network that converts the processed DCT coefficients into spatial domain images.Post-processing convolution network is then used to refine the obtained image and output the high-resolution image by enhancing the resolution.The experimenta l results show that,compared to the existing methods,our method can have better performance in terms of PSNR and SSIM with the same level of complexity(model parameters).Moreover,the enhanced high-resolution image has better visual effects.The image interpolation method based on recursive learning uses neural network to complete the image interpolation task.Firstly,the features are extracted by using a convolutional layer,and then the nonlinear mapping is performed by recursive learning.Finally,the reconstruction of the high resolution image is completed by reconstruction net.By using local residual,each module of the network has a direct path with other modules.The recursive learning method is adopted to re-use the shared weights of convolutio n layers,in order to reduce the number of parameters of the network.Dense connection facilitates gradient transmission in the network.Because the input low-resolution image is similar to the output high-resolution image in the image interpolation task,the global residual learning is adopted,which greatly improves the quality of output image.The experimental results shows that,compared with existing methods,the proposed image interpolation method not only requires fewer network parameters,but also acheives higher image reconstruction quality in terms of PSNR and SSIM values.