基于旋转候选区域的船舶检测

Ship detection based on rotating region proposal

利用图像进行船舶检测可以为海上侦察监视提供重要信息。以往的检测方法对水平或近水平的目标都有很强的检测能力,但对于任意方向的船舶检测,都不能给出满意的检测结果。在此提出了一种新型的船舶检测算法,可以检测任意方向的船舶。在该方法中,提出了一种利用旋转候选区域来生成具有船舶方位角信息的多方向候选信息,并且还对边框的方位角进行了回归,使倾斜船型区域的生成更加准确,最终提取识别特征。实验结果表明,提出的方法在处理任意方向的船舶检测任务时优于现有的方法。

Using images for ship detection can provide important information for maritime reconnaissance and surveillance.The previous detection methods have strong detection capabilities for horizontal or near-level targets,but they cannot provide satisfactory detection results for ships in any direction.Therefore,a new type of ship detection algorithm is proposed here,which can detect ships in any direction.In this method,a rotation region proposal is used to generate multi-directional proposal information with ship azimuth information.And the azimuth of the frame is regressed to make the generation of the inclined ship-shaped area more accurate,and finally extract the identification features.The experimental results show that the proposed method is better than the existing methods when dealing with ship detection tasks in any direction.