Research And GPU Implementation Of Image Super-Resolution Algorithm Based On Mapping And Regression

With the development of semiconductor technology,more and more devices supporting ultra-high definition resolution have appeared on the market,and many mobile devices have begun to support 2K or even 4K resolution.However,high-end acquisition equipment is expensive,and storing and transmitting ultra-high-resolution resources also imposes a heavy burden on hardware systems and bandwidth.These factors have led to the lack of sufficient ultra-high definition image and video data sources on the market.Super-resolution technology uses software methods to improve the resolution of images and video at low cost,which can solve the above problems well.This thesis mainly studies the key techniques of super-resolution algorithm based on mapping and regression,and then completes the real-time implementation of super-resolution algorithm on GPU.The work of this thesis includes:(1)Firstly,an edge-orientation-based super-interpolation(SI)algorithm is studied.The SI algorithm classifies image blocks using edge amplitude features and direction features.However,the edge amplitude feature is only used to assist the classification,and the division of the edge direction is also rough.Aiming at the above problems,combined with EO and K-means clustering,the improvement of SI algorithm is proposed.The improved method can effectively utilize the amplitude and direction information of the edge.The simulation proves that the improved algorithm has a good performance and has a good balance between reconstruction quality and running speed.(2)The key techniques of local linear mapping and regression based superresolution algorithm(GLM-SI)are deeply studied.Local regression is introduced in the linear mapping stage of the algorithm.A new algorithm with two-layer regression is proposed: DRLM-SI.The improved algorithm uses knn to obtain a plurality of local candidate images,and obtains global candidate images by local regression.The algorithm further improves the nonlinear relationship between high and low resolution image blocks.By simulation,the DRLM-SI algorithm can further improve the quality of reconstructed images.(3)Research and introduce GPU parallel programming technology,and analyze the advantages of GPU and CUDA platform.The super-resolution algorithm is implemented by using the high parallelism of the GPU platform.In this thesis,the algorithm of fast bicubic interpolation and fast nearest neighbor search is used to optimize the running speed of key modules,and the GPU memory,thread and other resources are allocated and utilized reasonably,realizing the real-time super-resolution of video.