用于SAR图像小目标舰船检测的改进SSD算法

针对合成孔径雷达(synthetic aperture radar, SAR)图像小目标舰船检测中存在的检测率较低、虚警较高等缺点,提出了用于SAR图像小目标舰船检测的改进单步多框检测(single shot multibox detector，SSD)算法。首先,制作了一个专门用于SAR图像小目标舰船检测的数据集,在SSD目标检测算法的基础上,提出了迁移学习、浅层特征增强和数据增广3个方面的改进;利用性能更好的ResNet50作为特征提取结构,在浅层特征增强网络结构中采用了inception模块的分支结构,同时使用了空洞卷积扩大特征图的视觉感受野,增强了网络对小尺寸舰船目标的适应性;最后在数据集上进行了多组对比分析实验,实验结果表明所提方法相比于原始的SSD,平均准确率提高了5.4%,并且对SAR小目标舰船的漏检和误报明显减少。

Improved SSD algorithm for small-sized SAR ship detection

Aiming at the disadvantages of low detection rate and high false alarm in small-sized ship detection for synthetic aperture radar (SAR) images, this paper proposes a small-sized SAR ship detection algorithm based on convolutional neural network. Firstly, a data set specially designed for small-sized SAR ship detection is constructed. Secondly, based on the single shot multibox detector (SSD) detection algorithm, improvements in transfer learning, bottom feature enhancement, and data augmentation are proposed. Using ResNet50 with better performance as the feature extraction structure, according to the basic principle of the inception module and the dilated convolution to expand the visual receptive field in the bottom feature enhancement, the algorithm enhances the adaptability of the network to small-sized ship targets. In the dataset of this paper, several sets of comparative analysis experiments are carried out. The experimental results show that the proposed method improves the average accuracy by 5.4% compared with the original SSD, and the missing detection and false alarms of small-sized SAR ships are obviously decreased.