Abstract:

Recently, deep learning has developed rapidly in the field of satellite image target detection, how to accurately and effectively locate the target object is the main challenge in the research of satellite image target detection. An improved YOLOv3 algorithm based on rotating rectangular space is proposed, which can accurately locate satellite image targets. This paper adds angle conversion as a preprocessing step, which can eliminate the distortion of instance angle during network training. Two kinds of rotating bounding box coordinates are proposed for regressing network and angle anchors are added to priori boxes, which can improve the effectiveness of network detection. A novel rotated non-maximum suppression is proposed to handle rotated bounding boxes, which can effectively remove redundant rotating prediction box. The experimental results show that the improved YOLOv3 algorithm has a significant improvement over the original effect, which can effectively and accurately locate the target object of the satellite image. Improved YOLOv3 has a better visualization effect and effectively avoids the occlusion problem of prediction frames in dense scenes and the mean average precision of detection is improved by 0.8 percentage points under the premise of ensuring real-time performance.

Key words: satellite imagery, target detection, deep learning, rotating bounding box

摘要：

近年来，深度学习在卫星影像目标检测领域得到了快速的发展，如何精准高效定位目标物体是卫星影像目标检测研究中的主要难点。提出了一种基于旋转矩形空间的YOLOv3改进算法来精准定位卫星影像目标，对原有网络进行改进，增加角度变换的数据预处理过程，防止实例角度变化对网络训练造成影响。使用双旋转坐标进行回归训练，增加了角度锚点，提高了网络对卫星目标的检测有效性。提出了基于旋转矩形空间的非极大值抑制改进算法，可以有效去除多余的旋转预测框。实验结果表明，改进YOLOv3算法相较于原始YOLOv3算法拥有更好的可视化效果，可以有效准确地定位卫星影像的目标物体，有效避免了密集场景下预测框的遮挡问题，在保证实时性的前提下，将均值平均精度提高了0.8个百分点。

关键词: 卫星影像, 目标检测, 深度学习, 旋转矩形框