利用超像素级上下文特征进行靠岸集装箱船检测

Detection Containers Based on Superpixel-Level Contextual Feature

高分辨率光学遥感影像中靠岸集装箱船受到岸边建筑、阴影和背景环境的干扰严重,且其船身模式与相邻陆地上集装箱非常相似,较难实现自动化检测。针对这一难题,提出了一种利用超像素级上下文特征进行靠岸集装箱船检测的方法。首先,对图像进行过分割生成超像素,在超像素区域提取颜色、纹理特征并级联邻域超像素特征形成超像素级上下文特征;然后,将目标超像素作为正样本,并自适应地选择较难区分的背景超像素作为负样本来训练分类器,实现对目标、背景超像素的分类;最后,利用全连接条件随机场对分类结果优化,实现对靠岸集装箱船的检测。实验结果表明,该方法能够较为可靠地检测靠岸集装箱船,具有一定的应用前景。

Inshore containers in high resolution optical imagery are under severe interference, such as structure, shadow, and environment, and the ship bodies are very similar to the container structures on nearby land. These situations make the automatic detection of inshore containers a very challenging task. In order to address this problem, this paper proposes a detection method for inshore containers based on the superpixel-level contextual feature. Firstly, the image is segmented into superpixels, and the features of the superpixel and its neighboring superpixels are concatenated into the superpixel-level contextual feature. Then, based on the positive samples and the actively selected negative samples, the target and the background superpixels are classified via machine learning. Finally, the fully connected conditional random field is employed to refine the classification result and realize the detection. The experimental result verifies the applicability of the proposed method.