Treffer: 航空遥感图像旋转目标检测技术研究综述.

Object detection mainly involves classification and regression. It is a basic task in computer vision and has been widely studied. Early research mainly focused on horizontal object detection in natural image scenes. In recent years, the development of Convolutional Neural Networks (CNNs) has facilitated the establishment of many general object detection methods. These methods have achieved good results on natural images, which greatly promoted the advancement of object detection tasks in aerial remote sensing images. Compared with natural images, aerial remote sensing images have complex backgrounds and generate objects that are densely distributed in arbitrary orientation. Therefore, traditional horizontal object detection is no longer suitable for object detection in aerial remote sensing images. According to the requirements of object detection tasks in aerial remote sensing images, the task of oriented object detection that relies on Oriented Bounding boxes (OBBs) to detect oriented objects accurately has gradually emerged. Most oriented object detection relies on the horizontal proposal/anchor of mainstream horizontal object detection frameworks to predict the OBBs. Although substantial progress has been made in performance, some fundamental flaws still exist, including the introduction of object-irrelevant information in the region features, as well as more complex rotation information calculation. At the same time, traditional CNNs cannot explicitly model the orientation variation of objects, which seriously affects the detection performance. Most existing oriented object detectors are based on horizontal object detectors by introducing an additional channel in the regression branch to predict the angle parameters. Angle regression-based methods have shown promising results, but they still encounter certain fundamental limitations. Compared with horizontal object detectors, angle regression detectors introduce new problems, mainly including 1) inconsistency between metrics and losses, 2) boundary discontinuity, and 3) square-like problems. At the same time, small objects in arbitrary orientation present great challenges to existing detectors, especially small oriented objects with extreme geometric shapes and limited features, which can lead to serious feature mismatch. These challenges have attracted widespread attention and prompted in-depth consideration from researchers in the relevant fields. To further promote the development of oriented object detection, this study mainly summarizes and analyzes the research status of oriented object detection in aerial remote sensing images. Currently, various pipelines exist for object detection, with the most common approach adding an angle output channel to the regression branch to predict the OBBs. As a result, many oriented object detectors are built upon the horizontal object detection framework. This study begins by reviewing key representative horizontal object detectors. With the development of object detection, many well-designed oriented object detection methods have been proposed and exhibit good performance. The main challenges encountered in current oriented object detection research are summarized in this study into five aspects: 1) feature alignment in object detection, 2) inconsistency between metric and loss, 3) boundary discontinuity and square-like problems, 4) low recognizability of small objects, and 5) label annotation problem. This study provides a comprehensive analysis of these challenges, introduces methods to address them, and discusses representative solutions in detail. The limitations and shortcomings of existing methods are examined, with potential directions for future exploration being considered. In the field of oriented object detection in aerial remote sensing images, several widely used and representative public benchmark datasets with OBB annotations are comprehensively summarized and introduced. The experimental results and visualizations of representative state-of-the-art detectors are compared and analyzed using the commonly employed datasets, namely, DOTA, HRSC2016, DIOR-R, and STAR. Current one-stage detectors are simple and effective, with anchor-free object detectors demonstrating comparable detection accuracy to traditional two-stage detectors. On this basis, this study integrates the current research status of object detection in aerial remote sensing images and anticipates future research trends. Future research could enhance the detection accuracy of oriented object detection while optimizing model complexity by developing anchor-free oriented object detectors, extracting rotation-invariant features, and minimizing annotation costs through weak supervision. Through this review, we aim to provide a valuable reference for researchers interested in exploring oriented object detection in aerial remote sensing images. [ABSTRACT FROM AUTHOR]

目标检测是计算机视觉领域中一项核心且具有挑战性的任务。近年来，目标检测在自然图像上取得了巨大成功，针对航空遥感图像目标检测的研究也取得了显著进展。与自然图像中水平目标不同，航空遥感图像中的目标往往以任意方向密集分布于复杂多变的背景之中。为精确高效定位方向目标并识别其类别，以水平检测为基础的旋转目标检测任务被提出。基于深度学习尤其是卷积神经网络CNNs（Convolutional Neural Network）的旋转目标检测虽然受到越来越多的关注，但当前对其存在的挑战缺乏系统性研究。本文重点阐述了航空遥感图像目标检测的研究现状，系统性地剖析了旋转目标检测存在的挑战性难题，目标在于推动相关检测技术的发展。首先，梳理归纳了水平目标检测的通用框架，它们也是旋转目标检测框架的设计基础；其次，重点剖析了旋转目标检测任务面临的主要挑战，总结了应对每项挑战而产生的主要研究成果及其优势和局限性；第三，简要介绍了常用的遥感图像目标检测数据集，并在DOTA、HRSC2016、DIOR-R、STAR等公开遥感图像基准数据集上对当前先进的旋转目标检测器进行了评估对比，在验证当前研究成果显著成效的同时，也初步揭示了它们在处理极端几何形状与复杂场景下存在的局限性；最后，对航空遥感图像旋转目标检测任务的发展趋势与进一步研究方向进行了展望。 [ABSTRACT FROM AUTHOR]

Copyright of Journal of Remote Sensing is the property of Editorial Office of Journal of Remote Sensing & Science Publishing Co. and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)