Accurate Ship Detection via Paired Semantic Segmentation
Xiaowu Xiao, Zhiqiang Zhou, Bo Wang, Zhe An · 2019
Ship detection in high-resolution optical satellite imagery is a challenging task due to complex backgrounds and docked ships side by side. In this paper, we propose a new way to generate ship proposals, and introduce an approach based on a novel deep encoding-decoding framework. According to the symmetry of the shape of the ship, we eliminate the need for designing a set of anchor boxes commonly used in prior ship detections and extract ship bounding boxes from the top-right and bottom-left segmentation parts of ship without location regression. By detecting ships as paired segmentation parts, we can detect docked ship side by side that previous semantic segmentation can not detect. The network is composed of multiple layers of convolution and de-convolution operators. We take the state-of-the-art convolutional neural network ResNet18 as the encoder network, which extracts the abstraction feature of image contents. The decoder network is responsible for recovering the image details. At the same time, we introduce skip-layer connections between convolutional and de-convolutional layers. Experiments have demonstrated the effectiveness of our approach both in qualitative and qualitative performance compared with state-of-the-art ship detection methods.