基于级联CNN的SAR图像舰船目标检测算法

针对合成孔径雷达(SAR)图像中舰船目标稀疏的特点,提出一种基于级联卷积神经网络(CNN)的SAR图像舰船目标检测方法.将候选区域提取方法BING与目标检测方法Fast R-CNN相结合,并采用级联CNN设计,可同时兼顾舰船检测的准确率和速度.首先,针对SAR图像中相干斑噪声影响梯度检测的问题,在原有梯度算子的基础上增加平滑算子,并对图像尺寸个数和候选框个数进行适应性改进,使其提取到的候选窗口更快更准;然后,设计级联结构的Fast R-CNN检测框架,前端简单的CNN负责排除明显的非目标区域,后端复杂的CNN对高概率候选区域进行分类和位置回归,整个结构可以保证快速准确地对舰船这种稀疏目标进行检测;最后,设计一种联合优化方法对多任务的目标函数进行优化,使其更快更好地收敛.在SAR图像舰船检测数据集SSDD上的实验结果显示,所提出的方法相比于原始Fast R-CNN和Faster R-CNN检测方法,检测精度从65.2%和70.1%提高到73.5%,每张图像的处理时间从2235ms和198ms下降到113ms.

A ship detection method based on cascade CNN in SAR images

According to the sparse characteristic of ships in synthetic aperture radar(SAR) images, a ship detection method based on the cascaded convolutional neural network(CNN) is proposed, which combines BING with Fast R-CNN in a cascaded way while taking the accuracy and speed into account. Firstly, a smoothing operator is added on the original gradient operator to tackle the speckle noise of SAR images. And the number of image size and candidate proposals are reduced according the distribution of ships in SAR images. After these improvements, the region proposal method gets more accurate without additional computations. Then, a cascade Fast R-CNN detection framework is designed to detect ships fastly and accurately. Its front simple CNN is responsible for rejecting the obvious background regions, and the back complex CNN is responsible for conducting classification and regression for the high probability candidate regions. The whole architecture makes the detection of the sparse ships in SAR images fastly and accurately. Finally, a joint optimization method is proposed to optimize the multi-objective function. The experiments on the dataset SSDD verify the superiority of the proposed method. The accuracy and speed boost from 65.2% /70.1% and 2235ms/198ms of the Fast R-CNN and the Faster R-CNN to 73.5% and 113ms, respectively.