Research On Image Super Resolution Algorithm And Application Based On Deep Learning

Super resolution reconstruction technology is to improve the quality of image by restoring the high-frequency details beyond the cut-off frequency in the imaging process,promoting the spatial resolution and visual clearance of the image.Traditional reconstruction algorithms often have many problems,such as low robustness and fixed size of input images,which cause obstacles for subsequent visual tasks such as image processing,recognition and analysis.With the rise of AI era,it is of great research significance and application value to study how to use deep learning algorithm to perform super resolution reconstruction technology.Therefore,this thesis deduces the convolutional neural network in image reconstruction in detail,which is based on the analysis of the process of mathematical modeling and theory of super resolution reconstruction from the perspective of mathematics.In addition,put forward super resolution reconstruction algorithms based on CNN model respectively for single frame and video frame image sequence.The main study works are as follows:1.Propose an improved convolutional neural network reconstruction algorithm based on channels combination.For the blur in the edge of the reconstructed image,put forward a channels combination algorithm that divides the network structure into two parts – convolution part and fusion part.Then make a fusion in pixel level for the features extracted from middle layer of convolutional neural network,which strengthens reconstruction details in the image and improves the network training convergence speed and image details of the definition.2.Furthermore,put forward an improved convolutional neural network model for video image super-resolution reconstruction.For the motion-blurred image,make full use of information inter video image frames and fusion of complementary information between frames.Propose an adaptive motion compensation algorithm,which further enhances the quality of reconstruction images and suppresses the influence of ringing phenomenon and noise.3.Propose a new algorithm of super resolution reconstruction based on generative adversial networks.For the problem of content loss in the process of reconstruction,introduce residual layers network to modify the model and make calculation and residual discrimination for the generated output,which can maintain the high spatial frequency image details well in the experiment.4.Complete the designing and building of super resolution reconstruction engineering system platform based on deep learning.Use the system platform to realize super resolution respectively for the high voltage image,medical image and security inspection image.The results show that the system is effective and can meet the demand in the practical appliance.