Development Of Model Compression And Inference Acceleration Algorithms Of Image Super Resolution Deep Neural Networks

Currently,while single image super resolution(SISR)networks based on deep learning technology achieve state-of-the-art reconstruction performance,which surpasses that of convolutional methods,the high computation complexity and memory usage has prevented their applications.It has become a hot academic topic in deep learning field how to best compress and accelerate networks,and it's possible to use such methods to reduce the SISR networks' demands for computation resources,thus enabling thier practical applications.Therefore,algorithms on compression and acceleration of SISR networks are developed in this thesis.Compression algorithms of SISR networks are developed first.SISR networks with limited theoretical computation complexity and good still PSNR and SSIM are given,by designing network architecture from various perspectives.Knowledge distillation algorithm is also designed to improve light weight SISR network's performance without changing its computation complexity.Then,acceleration algorithms of SISIR networks are developed.Pruning algorithm is designed based on channel pruning and different local pruning ratios,resulting in a network with little PSNR decrease and some FLOPs drop.Practical acceleration scheme is performed on light weight SISR network,based on GPU acceleration libraries,such as cudnn and TensorRT,and memory reuse strategy.Results show that the compressed and accelerated SISR network has good PSNR and SSIM performance,with 37.85dB,32,85dB,31.66dB and 32.06dB PSNR values for Sets,Set14,Urban100 and B100 dataset respectively,and could run at 56fps with 27.4%lower memory usage than non-acceleration methods.