基于中心点和语义信息的多方向遥感舰船检测

在高分辨率遥感影像解译中, 舰船目标的检测一直是研究热点。针对遥感影像中近岸舰船排列密集、方向各异以及背景复杂等问题, 本文提出一种基于旋转中心点网络和语义信息(rotated CenterNet using semantic information, RSI-CenterNet)的多方向遥感舰船目标检测方法。首先, 基于关键点检测网络, 在检测阶段添加目标角度回归分支, 以预测目标方向; 其次, 添加语义分割分支, 并将其输出的特征与检测部分的输入特征进行融合以强化目标区域的特征信息; 最后, 引入注意力模块, 以强化目标显著区域与通道的特征, 提升检测精度。实验结果表明, 与其他多种先进方法相比, 本文方法具有更高的检测精度与检测速度, 在高分辨率船舶数据集(High Resolution Ship Collections 2016, HRSC2016)上的平均精度达到88.31%, 检测速度达到17.8 FPS。

Multi-direction remote sensing ship detection based on center point and semantic information

Ship detection is a hotspot in the field of high-resolution remote sensing image interpretation. Aiming at the challenges of dense arrangement, different directions, and complex background of ships in remote sensing images, a rotated CenterNet using semantec information (RSI-CenterNet) multidirectional ship target detection method is proposed. First, we add a branch to predict the angle of objects in CenterNet. Second, a semantic segmentation branch is added, and the feature of semantic segmentation branch is used to guide the detection head to locate the center point more accurately. Finally, an attention module is introduced to enhance the feature of the significant areas and channels. Experimental results show that compared with other advanced methods, the proposed method has higher detection accuracy and detection speed, with an average accuracy of 88.31% and detection speed of 17.8 FPS in HRSC2016 dataset.