基于姿态表示的航空影像旋转目标检测网络

由于航空影像复杂多变的视角, 目标呈现出拥挤、聚集及旋转等特点, 传统目标检测中的水平边框难以契合地表示目标的几何轮廓及位置信息。本文提出了单阶段基于姿态表示的旋转目标检测网络。该网络将不同旋转角目标表示成不同姿态, 通过检测目标的中心位置及回归4个顶点相对坐标来实现旋转目标的检测。同时使用了自适应特征金字塔网络, 利用可学习权重自动从多尺度特征中选择更具判别性的特征。针对航空影像高分辨率的特点, 提出选择性采样策略以提高网络训练效率和缓解网络正负样本不平衡问题。本方法在DOTA遥感数据集旋转目标检测任务上的平均精度(mean Average Precision, mAP)达到74.9%, 超过了现有单阶段甚至部分双阶段的方法。定性与定量的对比实验表明, 基于姿态表示的旋转目标检测网络具有设计简单、检测性能更高的优势。

Pose-based Oriented Object Detection Network for Aerial Images

Horizontal bounding box representation in traditional object detection is not appropriate for ubiquitous oriented objects in aerial images because of the variant perspective, the crowded, cluttered and oriented objects. Therefore, a one-stage pose-based oriented object detection network is proposed, which represents oriented object as different pose and detect the oriented objects by locating the center and regressing four offsets between center and four vertices. Meanwhile, an adaptive feature pyramid network with learnable weights is utilized to automatically select more discriminative features. Moreover, according to the high resolution of aerial images, selective sampling strategy is proposed to improve the efficiency of network training and alleviate the imbalance problem of positive and negative samples. The proposed method achieves 74.85 mAP on oriented detection task of DOTA dataset, which outperforms the existing one-stage or even two-stage methods. The qualitative and quantitative comparative experiments show that the proposed pose-based oriented object detection network is simple and has competitive detection performance.