Survey on object detection in tilting box for remote sensing images

Object detection is a basic issue in remote sensing images processing, especially with the breakthrough of deep learning and the development of remote sensing images acquisition, object detection in aerial imagery based on deep learning has attracted extensive attention. Though considerable progress has been made, many obstacles still exist because of the large-scale and highly complex backgrounds of optical remote sensing images. Besides, when detecting densely arranged and arbitrary oriented objects in aerial imagery, approaches based on horizontal proposals for common object detection often suffer from the issue of mismatch. Many domestic and foreign scholars have proposed a series of tilting box object detection algorithms based on deep learning, which promoted the improvement of object detection effect of remote sensing image.In order to enable researchers in related fields to have a comprehensive understanding of the theory, process and existing problems of remote sensing image tilting box object detection based on deep learning, this paper will systematically organize and summarize them. In this paper, we first analyze the limitations of the horizontal bounding box（HBB） object detection algorithms SENSING applied to remote sensing images, i.e. the introduction of background noise, inappropriate post-processing operation NMS and the inability to determine the orientation of objects accurately, it can be better solved by using the tilting bounding box object detection method. Next, we list the classical HBB object detection algorithms based on deep learning, and briefly introduce the principle of some of them. Then, emphatically introduces the development of tilting bounding box object detection algorithm, and the improvement process of two-stage tilting bounding box object detection algorithm is introduced from three aspects of feature extraction network, anchor boxes and proposal region design and loss function, about the one-stage detection algorithm, there are few researches at the present stage, hence two algorithms are simply introduced. At the fourth section, the detection performance of the existing tilt box object detection algorithms on two public and challenging aerial datasets (i.e. DOTA and HRSC2016) is demonstrated. From the comparison results of the three tables, it can be seen that: it is necessary to design a certain object feature enhancement module for the particularity of the objects in the remote sensing images, the problem of RSE in tilting bounding box detection algorithm still needs to be solved further. Although the one-stage detection algorithm is slightly worse than the two-stage algorithm in accuracy, it has obvious advantages in efficiency, so it still has certain research value. Finally, we summarize the existing problems of tilt box target detection algorithm from six aspects, and look forward to its future development trend. NATIONAL REMOTE SENSINGBULLETIN www.jors.cn. © 2022 National Remote Sensing Bulletin. All rights reserved.