Fast Detection Of Ships Based On GPU

The detection and identification of ships on the sea is of great significance to safeguarding the country’s maritime rights and interests.The synthetic aperture radar(Synthetic Aperture Radar,SAR)imaging technology is not limited by the shooting weather and shooting time,and can realize 24-hour uninterrupted real-time monitoring of the earth’s surface.The use of SAR images for ship detection has become a research focus.The accumulation of SAR images has become more and more abundant,and the research on SAR automatic target recognition(ATR)system has become a hot topic.The ATR system includes three parts:target detection,classification and recognition.The system extracts the region of interest in the detection stage,and the classification and recognition algorithms are all autonomously in the region of interest.Therefore,detection is the most important part of the entire ATR system,and an excellent target detection algorithm will save a lot of calculation and calculation time for the subsequent classification and recognition process.Due to the growth of deep learning algorithms,more and more researchers have proposed to use deep learning algorithms to achieve target detection in SAR images.However,as the size of SAR images becomes larger and larger,the image resolution becomes higher and higher,and the algorithm is real-time.It has become a very important indicator to measure the performance of an algorithm,but it is difficult for deep learning to meet this requirement.The Constant False Alarm Rate(CFAR)algorithm is widely used in SAR image ship detection.Its advantage is that the algorithm executes fast and can meet the real-time requirements of actual target detection projects.The constant false alarm algorithm is mainly divided into two parts:global CFAR and local CFAR.Global CFAR is fast in execution but highly dependent on background clutter,and local CFAR is low in dependence on background clutter distribution but slow in execution speed.This article proposes a series of optimization methods to further improve the execution speed of the CFAR algorithm.Firstly,the GPU-based parallel optimization methods are proposed for the global CFAR algorithm and the local CFAR algorithm.Use multi-level reduction optimization methods to optimize global CFAR,and use texture memory and reasonable allocation of thread tasks to optimize local CFAR algorithms.Secondly,for the global CFAR algorithm,its detection accuracy mainly depends on the distribution model of the background clutter and the accuracy of the threshold calculation.Choosing an appropriate probability distribution function and an efficient and accurate threshold calculation method can effectively improve the detection accuracy of the algorithm.Particle Swarm Optimization(PSO)is employed to increase the speed of threshold solution and the calculation accuracy,and the PSO-based CFAR algorithm is deployed on the GPU to further improve the efficiency of the algorithm.In order to prove that the method proposed in this paper is effective,this paper conducts experiments on the SAR image data taken by the Gaofen-3 satellite to verify the effectiveness of the proposed method from three aspects:quality factor,threshold calculation accuracy and algorithm execution speed.