Arbitrary-Oriented Object Detection in Aerial Images with Dynamic Deformable Convolution and Self-Normalizing Channel Attention

Abstract

Objects in aerial images often have arbitrary orientations and variable shapes and sizes. As a result, accurate and robust object detection in aerial images is a challenging problem. In this paper, an arbitrary-oriented object detection method for aerial images, based on Dynamic Deformable Convolution (DDC) and Self-normalizing Channel Attention Mechanism (SCAM), is proposed; this method uses ReResNet-50 as the backbone network to extract rotation-equivariant features. First, DDC is proposed as a replacement for the conventional convolution operation in the Convolutional Neural Network (CNN) in order to cope with various shapes, sizes and arbitrary orientations of the objects. Second, SCAM embedded into the high layer of ReResNet-50, which allows the network to enhance the important feature channels and suppress the irrelevant ones. Finally, Rotation Regions of Interest (RRoI) are generated based on a Region Proposal Network (RPN) and a RoI Transformer (RT), and the RoI-wise classification and bounding box regression are realized by Rotation-invariant RoI Align (RiRoI Align). The proposed method is comprehensively evaluated on three publicly available benchmark datasets. The mean Average Precision (mAP) can reach 80.91%, 92.73% and 94.1% on DOTA-v1.0, DOTA-v1.5 and HRSC2016 datasets, respectively. The experimental results show that, when compared with the state-of-the-arts methods, the proposed method can achieve superior detection accuracy.