A Densely Connected Neural Network Based on SSD for Multiscale SAR Ship Detection

Abstract

Synthetic Aperture Radar (SAR) ship detection plays an increasingly important role in marine applications. With the development of deep learning methods, many object detection algorithms based on deep neural networks have emerged. However, these methods only obtain detection scores from a high-level convolutional feature map, ignoring the multiscale characteristic of objects. To solve the above problem, this paper proposes a densely connected neural network based on SSD. Theoretically, the high-level convolutional feature maps contain more semantic information, while the low-level convolutional feature maps collect the detailed features of the image. Thus, we combine the high- and low- level feature maps by adding dense connections to SSD, achieving the multiscale and multiscene SAR ship detection. In addition, considering the characteristics of SAR images, we discard the image augmentation strategy and use K-means clustering to design the multiscale candidate bounding boxes, improving the object detection accuracy. Finally, experiments carried out on SAR Ship Detection Dataset (SSDD) verify that the proposed method achieves satisfactory performance on multiscale and multiscene SAR ship detection.