Consistency-based Active Learning for UAV Object Detection
Haohao Hu, Tianyu Han, Wanjun Zhong, Yuerong Wang, Peng Zan · 2023
In recent years, with the development of artificial intelligence, many new technologies have emerged in computer vision, such as object detection. At the same time, unmanned aerial vehicle (UAV) as a strategic emerging industry has demon-strated significant value and has been widely used. Therefore, using deep neural networks for object detection of UAV images has become one of the current research hotspots. Currently, mainstream object detection networks are divided into two-stage and one-stage networks. Their task is to identify all the target objects to be recognized in the measured image and determine their position and category. However, due to the specific characteristics of UAV images, such as small target size, high density, overlapping areas, and large angle changes, object detection tasks based on UAV platforms still face considerable challenges. This paper proposes to use an active learning method named consistency-based active learning for UAV object detection (CALD- UAV) based on a deep convolutional neural network. The method uses image consistency and mutual information as indicators to search for unlabeled images with high information content for labeling. It obtains a more effective model with fewer training samples. To verify the effectiveness of our method, we conducted relevant experiments on the VisDrone2019 dataset and compared it with other active learning methods. The results show that our method, CALD- UAV, achieves good results in UAV data.