在基于合成孔径雷达(SAR)图像的舰船目标检测中,针对图像背景复杂、舰船尺寸大小不一等问题,提出了一种改进的YOLOv3深度卷积神经网络(CNN),用于SAR图像中的舰船目标检测。该方法基于训练数据集中的尺寸标签信息,使用交并比作为距离度量,利用k-means聚类方法为舰船目标提取了九组先验锚点框作为后续候选框边框参数优化的初始值;引入rGIOU来代替交并比rIOU,用于更新框回归损失和置信度损失,从而得到更加合理的损失函数,能将候选框与标注框之间的相对位置信息引入候选框的边框参数优化。为了验证改进版YOLOv3网络的性能,文中基于高分辨SAR舰船检测数据集AIR-SARShip-2.0,利用平移、翻转、调整亮度等方法进行数据集扩充,得到训练数据集和测试数据集,并进行舰船目标检测实验。实验结果表明:相较于常规YOLOv3网络和Faster R-CNN网络,改进YOLOv3网络在舰船目标检测上的总体效果更好,具有更高的准确率和更少的虚警,提高了平均精度指标,且需要的计算时间更少。

In the detection of ship targets based on synthetic aperture radar (SAR) images, according to the complex background in the SAR images and different sizes of ships, an improved YOLOv3 convolutional neural network (CNN) is proposed for the detection of the ship targets in SAR images. In this method, based on the dimension information labelled in the training images, with ratio of intersection over union as distance measurement, the k-means clustering method is applied to extract nine sets of prior anchor frames for the ship target; rGIOU(generalized intersection over union) is used to calculate the loss of box regression and confidence score instead of rIOU(intersection over union), so as to get more reasonable loss function and optimize the box parameters by taking better advantage of position information of the bounding boxes. In order to verify the performance of the improved YOLOv3 network, experiments are finished to carry out the ship target detection test. Based on the high resolution SAR ship detection data set AIR-SARShip-2. 0, the data set is expanded with translation, flip and brightness adjustment to obtain the training data set and test data set. The experimental results show that compared with the conventional YOLOv3 network and Faster R-CNN, the improved YOLOv3 network has better global effect on ship target detection in that higher accuracy ratio and lower false alarm ratio are achieved, average accuracy index is increased, and less calculation time is needed.

TN957.52

国家自然科学基金资助项目