Multiclass Object Detection in UAV Images Based on Rotation Region Network

Abstract

The object detection in UAV application is a challenging task due to the diversity of target scales, variation of views, and complex backgrounds. To solve several challenges, including dense objects, objects with arbitrary orientation, and diversity of aspect ratios, this article proposes an end-to-end object detection method based on the convolutional neural network. In this article, the feature extraction performance is enhanced by utilizing a deep residual neural network. Multiscale feature maps are obtained through fusion with different convolutional layers, thus combining the high-level semantic information and low-level detail information. A rotation region proposal network is adopted to generate rotated regions, which makes the bounding box sensitive to dense objects in aerial images. Meanwhile, the RoIAlign is used and a convolution layer is appended in the classification stage, and focal loss is used in the classification stage. The proposed method focuses on arbitrary-oriented and dense objects in UAV images. After a comprehensive evaluation with several state-of-the-art object detection algorithms, the proposed method is proved to be effective to detect multiclass artificial objects in aerial images. Extensive experiments are conducted on the DOTA, VEDAI, and the VisDrone UAV image datasets, which demonstrate that the proposed method can obtain discriminative features through the improved multiscale feature extraction and the rotating region network. The results on the above datasets show that our method obtains gains in mean average precision compared with several state-of-the-art methods.