Research On Image Super-resolution Reconstruction Algorithm Based On Deep Learning

In many cases,the limitation of devices' hardware and transmission bandwidth leads that resolution of the image captured by terminal device is low,and it is impossible to extract enough image information to meet the application requirements.To address this issue,the researchers proposed a super-resolution reconstruction technique that using known low-resolution images reconstruct corresponding high-resolution images.The technology has been found highly in video surveillance,satellite remote sensing,and medical imaging and is worth for searching.For images with complex and rich texture structure,the dictionary training model in the traditional super-resolution reconstruction method brings poor quality and poor facticity of the reconstructed high-resolution image with a dictionary trained by artificial features.In addition to this,the objective function of the existing super-resolution reconstruction method generally based on the pixel domain results in the phenomenon that the reconstructed high-resolution image with the lack of high-frequency texture information makes its surface too smooth.Therefore,there still exists room to improve performance of existing super-resolution reconstruction algorithms.In view of the above problems,focus on image super-resolution reconstruction algorithm based on deep learning in this paper with deep learning.Firstly,this paper proposes an image super-resolution reconstruction algorithm based on PCANet and kernel method.It mainly includes: 1)Proposing a dictionary training model based on PCANet and kernel method.On the basis of PCANet network,adding a kernel method and redesigning the network structure for training low-and high-resolution dictionaries,it can extract more nonlinear features of the image and make low-and high-resolution dictionaries contain rich features,and at the same time improve the expression ability of the dictionary,which overcomes the defects caused by artificial features;2)Combining sparse representation and using the trained dictionaries,do super-resolution reconstruction of a single image.The experimental results show that the super-resolution reconstruction method proposed in this paper improves the detail definition of the image and makes the texture smoother and more natural.At the same time,there is a good objective quality.Then,this paper proposes an image super-resolution reconstruction algorithm based on GAN(Generation Countermeasure Network)and Res Net(Residual Network),redsigning the network structure and network function: 1)A network structure based on GAN and Res Net take advantage of the residual network with characteristics of information transmission deeper to add the residual network to the generator network of GAN.It allows many important features of the image to be continued.Besides,it improves the ability of reconstructing the image by network that a batch normalization layer is added to the discriminator and generator to speed up the training of the network;2)Design the new loss function by means of the least squares and add the feature loss function in the generator network so that the reconstructed image has more high-frequency information;3)Combining the IBP method,apply the reconstruction algorithm to the license plate super-resolution to improve the clarity of license plate images.The experimental results show that the proposed super-resolution reconstruction method improves the detail definition of the image,makes the texture smoother and more natural and has a good objective quality.The proposed image super-resolution reconstruction algorithms in this paper have good performance in both subjective quality and objective quality of the reconstructed image,so it can be applied in all kinds of actual scenes,such as video surveillance,satellite remote sensing image restoration and multimedia communications.