Towards The Automatic Parallelization For Remote Sensing Image Enhancement Algorithms Based On Many-core Platform

Unmanned aerial vehicle(UAV)remote sensing system has the desirable features such as high resolution,high flexibility,and low cost.The system has been used in major natural disaster rescues and it played important roles in emergency management nowadays.However,the operation of UAVs are susceptible to natural disasters such as earthquakes and haze where the acquired images have poor visual effects and unqualified target characteristics.Remote sensing(RS)image enhancement algorithms can improve the quality and visual effect of the RS images.The computing time to process large amount of UAV RS images with image enhancement algorithms is often too long to meet the real time requirements of rapid acquisition of the information from the disaster area in time.In addressing such challenges,some scholars began to use the many-core platforms such as graphics processing units(GPUs),many integrated cores(MICs)to develop corresponding parallel image enhancement algorithms.However,the studies adopting these methods have one or more of the following problems:(1)With the UAV RS system,different disaster scenarios need different RS image enhancement algorithms.Since the image enhancement algorithm has many variants,when many image enhancement algorithms have to be considered,a single treatment will not work.(2)For different image enhancement algorithms,their parallelization development cycle will be elongated due to different hardware used as GPUs and MICs have different programming models and hardware characteristics,which results different parallelization strategies in different image enhancement algorithms.(3)Some automatic parallelization tools can solve the problems mentioned above and improve the development efficiency,but most automatic parallelization tools are often not freely available or inefficient,moreover,they are lack of specific optimizations for RS image enhancement algorithms.In addressing the problems,this thesis builds an automatic parallel model for UAV RS image enhancement algorithms based on the open source automatic parallel package,Par4 All.The main research topics focus on:(1)Investigate the principle of typical UAV RS image enhancement algorithms,then design and implement the corresponding parallel algorithms by with the automatic parallel software,Par4 All and in the traditional way,i.e.,manual parallelization.After the comparison of the structural characteristics and test results of the parallel algorithms with the form of OpenMP and OpenCL in those two ways,we explore the hotspots of the parallel image enhancement algorithms with the way of automatic parallelization,namely,the Par4 All package.(2)Research on structure of the modules and the automatic parallelization working mechanisms of Par4 All.Based on that,the thesis introduces a new artificial-intelligence(AI)-based searching algorithm,iterative hill climbing algorithm,into the module of Par4 All.Thus,the elapsed time of the automatic parallelization procedure is shortened significantly,and the performance is improved.(3)After carefully examine the structure of the generated parallel image enhancement algorithms with Par4 All,we found the parallel algorithms have problems such as,high threads overhead,out-off-balance of the number of threads and that of the cores.In addressing the problems,we construct an automatic parallelization module that dynamically set the threads for remote sensing image enhancement algorithms.(4)In the generated code from Para4 All,it shows high computation and less logic computation in the high parallel area.Moreover,CPU end of the parallel algorithm is always idle when the many-cores end processes the intensive computation.In addressing these challenges,we introduced the cooperative computing strategy to improve the automatic parallelization module for remote sensing image enhancement algorithms based on(3).In the experiments,the RS images acquired by UAV at different stages have been processed with the corresponding parallel image enhancement algorithms.After we carefully compared the experiments results,we conclude that:(1)The parallel image enhancement algorithms with the manual parallelization have higher computing cost but have higher speedup;(2)The automatic parallelized algorithms using Para4 All have shorter development cycle,however,the speedup is much lower;(3)The paralle l image enhancement algorithms developed by the proposed automatic parallelization module have the advantages of both(1)and(2).Furthermore,we use another image enhancement algorithm,Gaussian filter algorithm to validate and verify the proposed automatic parallelization module.The results show that the automatic parallel model for remote sensing image enhancement is effective and has high efficiency for other remote sensing image enhancement algorithms.Obviously,the research has certain academic value to explore the automatic parallelizat ion model in other field of remote sensin g image processing.