Research On Application Of Embedded GPU In SAR Image Change Detection And Video Frame Rate Transformation

Traditional desktop GPU has the disadvantages of high power consumption and large device size,which is not conducive to carrying in the embedded mobile device.In recent years,with the rapid development of high performance computing,the performance of system level chips has been improved,and the computing power of embedded platform devices has been significantly improved.Embedded GPU,with low power consumption and high performance,is challenging the traditional desktop GPU devices.This thesis focuses on the research of NVIDIA embedded GPU in two algorithms: the first algorithm is the parallel CUDA-FLICM clustering algorithm which is used in SAR image change detection,and the second algorithm is parallel video frame rate conversion algorithm.The main contents of this thesis are as follows:(1)In this thesis,the SAR image detection algorithm based on FLICM clustering is analyzed,and a parallel CUDA-FLICM clustering algorithm is proposed.The parallel CUDA-FLICM algorithm is divided into two parts: In the first part,the difference graph is solved by parallel method.The parallel method is to map the log-ratio calculation of each pixel in SAR image into a thread of GPU.The second part is the parallel FLICM clustering operation,which is the computation of the penalty factor of each pixel and the calculation of the membership degree,which is mapped in a thread of GPU.In this thesis,We test the speedup ratio of the parallel CUDA-FLICM algorithm on the embedded Tegra K1 GPU and the embedded Tegra X1 GPU.The results show that CUDA-FLICM parallel algorithm can achieve the highest speedup of 130×.(2)In this thesis,the parallel conversion algorithm of video frame rate is parallelized by CUDA parallel framework.The parallel video frame rate up conversion algorithm consists of 3DRS motion estimation algorithm and motion compensation algorithm based on median-filtering.Firstly,we analysis the 3DRS motion estimation algorithm.and parallel motion estimation CUDA-ME algorithm is proposed.In the CUDA-ME algorithm,as the space candidate motion vectors are dependent on each row block,Therefore,the computation of 10 candidate vectors for each row block is mapped on the GPU and each threads of GPU computes a candidate vector.In addition,we analysis the motion compensation algorithm,and the parallel motion compensation algorithm CUDA-MC is proposed.In the CUDA-MC algorithm,as the operations between each pixel are independent of each other,the interpolation operations of each pixel in the motion compensation algorithm can be accomplished by the GPU kernel.The parallel frame rate up conversion algorithm is implemented on the NVIDIA embedded Tegra K1 GPU platform,and we test the speedup ratio of CUDA-ME parallel algorithm and CUDA-MC parallel algorithm respectively.Experiments show that the CUDA-ME parallel algorithm can achieve the highest speedup of 19×,and the CUDA-MC parallel algorithm can achieve the highest speedup of 56×.After that,the visual system of parallel video frame rate conversion algorithm was realized by Open GL.The application foreground of NVIDIA embedded GPU in engineering field was demonstrated.