【作者】 刘恒；

【导师】 梁毅；

【作者基本信息】 西安电子科技大学 ， 信号与信息处理， 2022， 硕士

【摘要】 随着合成孔径雷达(Synthetic Aperture Radar,SAR)图像解译和深度学习技术的不断发展,学者们将深度学习算法引入SAR领域,充分利用高分辨SAR图像丰富的空间和结构视觉信息进行目标检测识别,检测性能优于传统方法。然而基于SAR图像的目标检测依然面临很多问题:首先面对复杂场景下SAR图像背景杂波强度相对较高、杂波统计非均匀、目标分布相对密集等问题,传统恒虚警率检测方法(Constant False Alarm Rate,CFAR)难以选取合适杂波模型,会产生大量的虚警和漏警,深度学习方法作为通用目标检测算法很少利用SAR图像的先验信息,复杂场景时检测性能有所下降;另一方面,考虑到便携性、隐秘性、网络传输带宽等因素,越来越多应用场景,如自动驾驶和弹载图像处理领域,均要求将目标检测模型直接部署在移动端或嵌入式设备上,在有限资源平台实现本地离线式实时目标检测。针对上述问题和应用需求,本文开展基于显著性增强的SAR目标检测与DSP加速研究,主要工作如下:1、分析介绍视觉显著性检测及深度学习检测理论:首先研究经典显著性检测算法,完成基于SAR舰船图像显著性提取实验;随后对深度学习中网络基础结构与训练方法进行介绍,并分析评价模型性能的相关指标,为后续研究改进算法提供理论支撑。2、针对港口岛岸等复杂背景下舰船检测虚警率高的问题,提出一种融合显著性信息的FCOS舰船检测方法。首先考虑到模型性能与计算复杂度的因素,采用深度可分离卷积对模型的特征提取网络进行轻量化改进;随后分析舰船目标的显著性,通过频率调谐方法提取显著性信息,对多尺度特征进行背景抑制和目标增强;最后采用无锚式回归框得到检测结果。实验结果证明显著性信息能有效抑制图像背景杂波,提升复杂岛岸背景下SAR舰船检测性能。3、针对移动端或嵌入式端等有限资源的计算平台下深度学习方法部署难的问题,研究改进的轻量化网络Nano Det在C6678 DSP平台上加速推理方法。首先,根据C6678提供的资源和实时检测系统需求,选择Nano Det作为轻量化检测模型,并且完成编程框架设计;其次,从DSP内存结构、向量编程、缓存机制以及编译优化等方面对深度学习模型中算子进行实现和加速优化;最后,实验结果证明C6678平台执行单幅图像的Nano Det前向推理耗时约152ms,满足实时检测要求并且性能良好。更多还原

【Abstract】 With the development of SAR(Synthetic Aperture Radar)image interpretation and deep learning technology,many scholars have introduced deep learning algorithm into SAR filed,making full use of the rich spatial and structural visual information of high-resolution SAR images for object detection and recognition.However,the object detection based on SAR images still faces many challenges.On the one hand,in complex scenes SAR images have many problems such as high background clutter intensity,non-uniform clutter statistics,and dense target distribution.Thus,it is difficult to select appropriate clutter statistical model for traditional CFAR(Constant False Alarm Rate)method,which lead to false alarms and missed alarms detection.Moreover,since SAR object detection methods based on deep learning algorithm rarely uses the prior information of SAR images,the detection performance has the potential to be improved.On the other hand,considering portability,secureness,network bandwidth and other factors,more and more application scenarios,such as automatic driving and missile-borne image processing,desire to deploy object detection model on embedded devices,and achieve local offline real-time object detection.In response to the above problems and application requirements,this paper carries out research on SAR object detection and DSP acceleration.The main works are as follows:1.The foundation of visual saliency detection and deep learning detection are analyzed.First,this paper studies the classical saliency detection methods and completes a saliency extraction experiment based on SAR ship image.Subsequently,the Convolutional Neural Network(CNN)structure and training methods of deep learning algorithm are introduced,which provides theoretical support for the subsequent improvement method.Moreover,the relevant indicators of the evaluation model are analyzed.2.Aiming at high alarm rate of SAR ship detection in complex background such as port and island,a saliency information fused FCOS ship detection method is proposed.First,considering the factors of model performance and computational complexity,depth separable convolution is used to achieve a lightweight model.Then,by analyzing the saliency of ship,background clutter in multi-scale feature is suppressed through the saliency information extracted by frequency-tuned method.Finally,an anchor-free detector is used to get the results of ship object detection.Experimental results prove that saliency information can effectively suppress background clutter,and improve the detection performance in complex background.3.Aiming at the model deployment of deep learning detection under resource limited embedded platform,a lightweight Nano Det model acceleration method on C6678 DSP platform is proposed.First,according to the resources of C6678 platform and real-time detection system requirements,this section designs a lightweight program frame of Nano Det.Then,accelerated CNN operators are completed considering DSP memory structure,vector programing,caching mechanism and compilation optimization.Finally,experimental results show that forward reasoning time cost is about 152 ms,which satisfies the real-time detection requirements.更多还原