Research And Implementation Of GPU-based Large Scene SAR Target Detection Method

Due to the all-day and all-weather characteristics of Synthetic Aperture Radar(SAR),SAR target detection is widely used in military and civilian fields.With the increase of SAR image data and the complexity of detection algorithm,the traditional large scene SAR target detection algorithm based on CPU serial calculation can not meet the requirements of real-time detection and detection accuracy.Thanks to the rapid development of GPU general-purpose computing technology,it provides a feasible solution for real-time target detection of large scene SAR images.Therefore,the research of GPU-based large scene SAR target detection method is of great significance for real time target detection of SAR images.In this thesis,the principle of classical CFAR algorithm and the AdaBoost algorithm based on Haar-like feature is briefly analyzed,and the related theory of CUDA programming of GPU platform is introduced.Then,the two algorithms are analyzed in parallel,and the factors affecting the real time detection and detection accuracy of the algorithm are explored.Finally,in view of the existing deficiencies,the traditional algorithm is improved,and the improved algorithm is implemented in parallel on the GPU platform.Through experimental comparison,the accuracy of the improved algorithm in SAR image target detection and the real time performance of GPU parallel implementation are verified.The specific work arrangement of this thesis is as follows:(1)The CFAR target detection algorithm and the AdaBoost target detection algorithm based on Haar-like feature are analyzed.The related theories of CUDA programming model and dynamic parallelism of CUDA parallel acceleration are introduced.(2)The CFAR algorithm is analyzed and implemented in parallel,and its shortcomings in detection accuracy and real-time detection are explored.The CFAR is improved according to the existing deficiencies.The CP-CFAR target detection algorithm is proposed,and the CP-CFAR algorithm is analyzed and implemented in parallel.By comparing with the CFAR algorithm,the CP-CFAR algorithm achieves the real-time target detection requirement while ensuring the detection accuracy.(3)The AdaBoost target detection algorithm based on Haar-like feature is analyzed to explore the reason why it is difficult to achieve real-time performance under the condition of large scene SAR image target detection.CUDA parallel acceleration is performed separately for its training process and detection process.For the process of calculating Haar-like eigenvalues,an improved algorithm which is more suitable for parallelism on the GPU platform is proposed.The iterative and cyclic operations of the training process are accelerated by the dynamic parallelism of CUDA,and the detection process is implemented by CUDA in parallel.After the acceleration,the training process reach a speedup of 30.In the case of large scene SAR images,the detection process also meets the real time detection requirements while achieving a high detection accuracy.