基于级联网络的复杂大场景SAR图像舰船目标快速检测

合成孔径雷达(synthetic aperture radar,SAR)图像舰船目标检测紧贴军事和民用需求,为海洋监视提供重要信息支撑。针对复杂大场景SAR图像,本文设计了一种基于级联网络的舰船目标检测框架,该网络框架主要由D-BiSeNet海陆分割、分块区域筛选和CP-FCOS目标检测三部分组成。通过改进双边网络(D-BiSeNet)进行SAR图像海陆分割,增强了图像空间位置信息及网络边缘损失,提高了分割性能。通过海域面积比参数设定进行分块区域筛选,可以有效选择网络处理图像块,提升算法整体检测效率。CP-FCOS网络将Category-Position特征优化模块应用于传统FCOS网络,强化网络特征提取能力,同时改进目标分类和边界框回归方式,提高舰船目标定位效果。基于Sentinel-1和高分三号大场景实测数据实验表明,相比于传统CFAR、Faster-RCNN和RetinaNet方法,本文方法综合检测性能提升25.7%,3.7%和9.9%,同时检测速度提升10.0%以上。 

Fast Detection of Ship Targets for Complex Large-scene SAR Images Based on a Cascade Network

Synthetic aperture radar (SAR) image ship target detection closely meets military and civilian needs, and provides important information support for marine surveillance. Aiming at complex large-scenes SAR Images, a fast detection framework for ship targets based on a cascade network is proposed in this paper. The detection framework is mainly composed of three parts: D-BiSeNet sea and land segmentation, block area screening and CP-FCOS target detection. D-BiSeNet is an improved bilateral network (BiSeNet) adapted to the sea and land segmentation of SAR images. It improved the segmentation performance by enhancing image spatial location information and network edge loss. Sea area ratio is set to select network processing image blocks effectively, which can improve the overall detection efficiency of the algorithm. A Category-Position feature optimization module is applied to the traditional FCOS network in CP-FCOS, which can strengthen the feature extraction capabilities of network. Meanwhile, the target classification and boundary box regression methods are redesigned to improve the effect of ship target positioning. The experimental results based on Sentinel-1 and GF-3 large-scene images show that compared with traditional CFAR, Faster-RCNN and RetinaNet methods, our method obtains 25.7%, 3.7% and 9.9% gains in the comprehensive detection performance, while the detection speed is improved more than 10.0%.