主动迁移学习的海上任意方向船只目标检测

Ship Object Detection in Any Direction at Sea Based on Active and Transfer Learning

在基于深度学习的遥感图像目标检测任务中,船只目标通常呈现出任意方向排列的特性,而常见的水平框目标检测算法一般不能满足此类场景的应用需求。因此本文在单阶段Anchor-Free目标检测器CenterNet的基础上加入旋转角度预测分支,使其能输出旋转边界框,以用于海上船只目标的检测。同时针对海上船只遥感数据集仅有水平边界框标注,无法直接适用于旋转框目标检测,且人工手动标注旋转框标签成本较高的问题,提出一种主动迁移学习的旋转框标签生成方法。首先,提出一种水平框-旋转框约束筛选算法,通过水平真值边界框来对旋转预测框进行监督约束,筛选出检测精度较高的图像加入训练集,然后通过迭代这一过程筛选出更多的图像,最后通过标签类别匹配,完成对数据集的旋转框自动化标注工作。本文最终对海上船只遥感图像数据集BDCI中约65.59%的图片进行旋转框标注,并手动标注部分未标注的图片作为测试集,将本文方法标注的图片作为训练集进行验证,评估指标AP50达到90.41%,高于其他旋转框检测器,从而表明本文方法的有效性。

In remote sensing images object detection tasks based on deep learning,the ship usually exhibit features arranged in any direction.The common algorithms of object detection adopt horizontal detection that generally cannot meet the application requirements of such scenarios.Therefore,this paper adds a rotation angle prediction branch to the single-stage Anchor-Free object detector CenterNet,it can output a rotating bounding box for the detection of marine ship objects.At the same time,in view of the problem that maritime ship remote sensing data sets only have horizontal bounding box labels,which cannot be directly applied to rotating boxes object detection,and manual labeling of rotating boxes labels is expensive,an active and transfer learning method of rotating boxes label generation is proposed.Firstly,a horizontal box-rotating box constraint screening algorithm is proposed.The rotating prediction box is supervised and constrained by the horizontal ground truth bounding box.The image with higher detection accuracy is selected and added to the training set.Then this process is iterated to filter out more images.Finally,the automatic labeling of the rotating box of the data set is completed by matching the label categories.In this paper,about 65.59%of the pictures in the remote sensing image data set BDCI of marine ships are finally marked with a rotating boxes,and some unmarked pictures are manually marked as the test set.The pictures marked by the method in this paper are used as the training set for verification.The evaluation index AP50 reaches 90.41%,which is higher than other rotating boxes detectors,indicating the effectiveness of this method.