The Improvement And Parallel Optimization Of Image Super-Resolution Using Convolutional Neural Network

High resolution(HR)images have important applications in monitoring,remote sensing,entertainment,medicine,and other aspects.Therefore there are often many problems for the hardware resolution in aspect of technology and cost.Super Resolution(SR)reconstruction technology could increase image resolution by means of software algorithms and has been widely studied due to its convenience and low cost.In recent years,deep learning neural network technology has developed rapidly,and super-resolution technology has also run into new opportunities,opening up hot research areas based on this point.At the same time,with the progress of the SR algorithms,the algorithm complexity is increasing and the processor has encountered higher requirements.As the bottleneck of Moore's Law approaches,the traditional serial architecture processors cannot meet the needs of such large-scale computing,making the parallel architecture processor received more extensive attention.Meanwhile,the application of graphics processing units(GPUs)for algorithms with high degree of parallelism has become increasingly widespread,and the way of development algorithms using parallel processors as platforms has become more popular.In this thesis,the research and engineering applications of the SR reconstruction algorithm based on Convolutional Neural Network(CNN)has been improved.By analyzing the parallelism of this algorithm,the GPU processor is used to accelerate and optimize the algorithm,in order to improve the practicability of the algorithm.The main work includes the following aspects.(1)Research and redesign of the SR reconstruction algorithm based on convolutional neural network.Both the network structure and the training set were researched and optimized to improve the reconstruction effect.Aiming at the insufficiency of partial information loss in the image preprocessing before the reconstruction step,a corresponding improvement method was proposed,in which a completely convolutional neural network SR reconstruction algorithm was designed,which could improve reconstruction quality on the basis of almost no increase in reconstruction time.Also the impact of the training set on the reconstruction quality was analyzed.When the data size of the training set is too small,putting the transposition of the training set into the training set again could improve the reconstruction quality.By combining two methods,the Peak Signal-to-Noise Ratio(PSNR)value between original images and reconstruction images is improved by 0.22 d B and 0.15 d B respectively in the test sets of Set5 and Set14 for the SR domain when the number of iterations is half of the original method.(2)Application-oriented reconstruction design by block segmentation.The method of block segmentation reconstruction was designed and implemented.The results show that the amount of information contained in each part of the image after block segmentation is very different.Therefore,it is not necessary to perform the same amount of calculation,so the recommendatory strategy is to perform high quality reconstruction only for the informationrich area,and to implement on low quality reconstruction for areas with insufficient information,which is helpful to provide precious computing resources to operations that can significantly improve reconstruction quality.(3)Accelerate actualization by CUDA(Compute Unified Device Architecture,CUDA)on the optimized algorithm and analyze reconstruction effect by experiment.Based on the GPU processor,the CUDA-C language was used to redesign the algorithm,then the algorithm parallelism and memory read and write methods were analyzed,so the shared memory and constant memory were used to further accelerate the algorithm.In this process,the nonlinear mapping layer convolution operation was converted to matrix multiplication,and the repeated operations and thread operation logic in the calculation process were analyzed,so it could fully tap the computing power of the GPU processor,and increase the speed of calculation.Experimental results showed that the CUDA acceleration method proposed in this thesis could achieve an acceleration ratio of around 100 compared to that from the CPU method,and CUDA acceleration does not affect the image reconstruction effect shown by PSNR values.