ABSTRACT

Anchor-based methods, which require a large number of pre-set anchors, have been widely used for oriented object detection in remote sensing images. However, the definitions of anchors’ sizes, aspect ratios and quantities are heuristic and the use of anchors is time-consuming. In this paper, we propose an oriented one-stage anchor-free detector for aerial image object detection. Arbitrary oriented object detection is based on oriented bounding box regression. By adaptively fusing the features from the neighbour layers of the feature pyramid network (FPN), a finer adaptive feature fusion network is proposed to align the features with ground truths. The proposed network can avoid the ambiguous heuristic-guided feature selection caused by scale variations of aerial image objects. We also design a foreground enhancement module to obtain more discriminative features from the fused FPN. Experiments on remote sensing image public datasets show that our method can outperform current one-stage anchor-free methods and achieve comparable performance with state-of-the-art two-stage anchor-based methods.

摘要

需要大量预设锚点的基于锚点的方法已广泛用于遥感图像中的定向对象检测。然而，锚的尺寸，长宽比和数量的定义是试探性的，并且锚的使用是费时的。在本文中，我们提出了一种定向的一级无锚检测器，用于航空图像目标检测。任意定向对象检测基于定向包围盒回归。通过自适应融合来自特征金字塔网络（FPN）相邻层的特征，提出了一种更精细的自适应特征融合网络，以使特征与地面真相对齐。所提出的网络可以避免由航空影像对象的尺度变化引起的含糊的启发式引导特征选择。我们还设计了前景增强模块，以从融合的FPN中获得更多区分功能。遥感图像公开数据集上的实验表明，我们的方法可以胜过当前的一阶段无锚方法，并且可以与基于两阶段锚的最新方法相媲美。