Implementation And Optimization Of Threshold Segmentation Algorithm For Feiteng Platform

Digital signal processor(DSP)is widely used in image processing because of its timeliness and efficiency.In recent years,the research of high-performance digital signal processor in China has made certain achievements.FT series multi-core processor and vector processor independently developed by the University of national defense science and technology have emerged.However,in the rapid development of domestic high-performance DSP,there is no high-performance image processing algorithm which can give full play to its architecture advantages.In view of this problem,the text has carried out a deep research on the threshold segmentation algorithm which is widely used in FT-M7002 platform.The main work of this paper is as follows(1)The overall implementation process of two threshold segmentation algorithms for FT-M7002 platform is designed.By studying Otsu threshold segmentation algorithm,adaptive threshold segmentation algorithm and FT-M7002 hardware structure,including the overall structure of FT-M7002,the kernel structure of FT-MT2 and the memory hierarchy,the implementation process of two threshold segmentation algorithms combined with the characteristics of FT-M7002 architecture is designed.By testing the hot program segment in the threshold segmentation algorithm,the specific optimization methods available in the two threshold segmentation algorithms are determined.(2)Combined with the characteristics of ft-m7002 platform,Otsu threshold segmentation algorithm and adaptive threshold segmentation algorithm are optimized in parallel.For the ultra long vector register provided by the platform,the algorithm vectorization rewriting using FT vector instruction set is realized,which reduces the number of data memory access and effectively excavates the data level parallelism contained in the threshold segmentation algorithm;for the DSP pipeline and vilw structure provided by the platform,the number of instructions executed at the same beat is improved through optimization methods such as loop expansion and pipelining,The instruction level parallelism contained in the threshold segmentation algorithm is effectively mined;according to the structural characteristics of DMA hardware and vector memory in FT-MT2 kernel,DMA array transposition is used to improve the data layout,realize the continuous access of data in the threshold segmentation algorithm,and ping-pong cache is used to realize the parallel of data transmission and kernel computing,so as to effectively hide the data transmission time.The results show that the optimized threshold segmentation algorithm can achieve the correct segmentation effect.For Otsu thresholding segmentation algorithm,the optimized algorithm has 1.77 ? 1.87 times performance improvement;for adaptive thresholding segmentation algorithm,the optimized adaptive thresholding segmentation algorithm can obtain 1.66 ? 1.71 times and 1.80 ? 2.62 times acceleration effect respectively under various neighborhood sizes and different threshold calculation methods.