Current Issue Cover
航空遥感图像深度学习目标检测技术研究进展

石争浩,仵晨伟,李成建,尤珍臻,王泉,马城城(西安理工大学计算机科学与工程学院;西安翔腾微电子科技有限公司集成电路与微系统设计航空科技重点实验室)

摘 要
航空遥感图像目标检测旨在定位和识别遥感图像中感兴趣的目标,是航空遥感图像智能解译的关键技术,在情报侦察、灾害救援、资源勘探等领域具有重要应用价值。然而由于航空遥感图像具有尺寸大、目标小且密集、目标呈任意角度分布、目标易被遮挡、目标类别不均衡、背景复杂等诸多特点,航空遥感图像目标检测目前仍然是极具挑战的任务。近年来,基于深度卷积神经网络的航空遥感图像目标检测方法因具有精度高、处理速度快等优点,受到了越来越多的关注。为推进基于深度学习的航空遥感图像目标检测技术的发展,本文对当前主流的遥感图像目标检测方法,特别是近三年提出的检测方法,进行了系统梳理和总结。首先梳理了基于深度学习目标检测方法的研究发展演化过程,然后对基于卷积神经网络和基于Transformer目标检测方法中的代表性算法进行分析总结,再后针对不同遥感图象应用场景的改进方法思路进行归纳,分析了典型算法的思路和特点,介绍了现有的公开航空遥感图像目标检测数据集,给出了典型算法的实验比较结果,最后给出现阶段航空遥感图像目标检测研究中所存在的问题,并对未来研究及发展趋势进行了展望。
关键词
Object Detection Techniques Based on Deep Learning for Aerial Remote Sensing Images: A Survey

Shi Zhenghao,Wu Chenwei,Li Chengjian,You Zhenzhen,Wang Quan,Ma Chengcheng()

Abstract
Benefiting from the booming development of aerospace technology, high-resolution remote-sensing images have entered daily research work. While the earlier low-resolution images limit researchers" interpretation of image information, today"s high-resolution remote sensing images contain richer geographic and entity detail features and are rich in spatial structure and semantic information, which can greatly promote the development of research in this field. Aerial remote sensing image object detection aims to give the category and location of the target of interest in aerial remote sensing images and provide evidence for further information interpretation reasoning, which is a key technology for aerial remote sensing image interpretation and has important applications in intelligence reconnaissance, target surveillance, disaster rescue and so on. The early remote sensing image object detection task mainly relied on manual interpretation to complete, and the interpretation results were greatly affected by subjective factors such as the experience and energy of the interpreters, and the timeliness was low. With the progress and development of machine learning technology, various remote sensing image object detection based on machine learning technology have been proposed. Traditional machine learning-based object detection techniques generally use manually designed models to extract feature information such as feature spectrum, gray value, texture, and shape of remote sensing images after generating sliding windows. They then feed the extracted feature information into classifiers such as SVM and AdaBoost to achieve object detection in remote sensing images. These methods design the corresponding feature extraction models for specific targets, with strong interpretability, but weak feature expression capability, poor generalization, time-consuming computation, and low accuracy, which are difficult to meet the needs of accurate and efficient object detection tasks in complex and variable application scenarios. In recent years, with the wide application of deep learning techniques such as deep convolutional neural networks and generative adversarial neural networks in the fields of natural image object detection, classification, and recognition, especially the excellent performance in the task of large-scale natural scene image object detection, the research on the application of deep learning in remote sensing image processing has also received great attention and become a research hotspot in this field in recent years, giving birth to Many excellent works have been born. At present, object detection in aerial remote sensing images mainly faces challenges such as large-size and high-resolution images, interference from complex backgrounds, target direction diversity, dense targets, dramatic scale changes, and small targets, etc. For these challenges, there are now corresponding model improvement methods. For large-scale, high-resolution aerial remote sensing images, because the target scale in the image is widely distributed, to ensure the integrity of small target detail information, so the most commonly used detection and recognition method is to segment the image during data pre-processing, that is, the large image is segmented into regular image size and then sent to the object detection algorithm for detection and recognition in turn, and finally in the subsequent processing Finally, all the detection results are stitched together and reset in the subsequent processing, to complete the detection of the whole image. At the same time, the aerial remote sensing image with ultra-high resolution has a complex background, and the target to be detected is easily interfered with by various similar objects, and the similar targets to be detected also present different characteristics, so it is easy to produce false detection when carrying out detection. Therefore, the usual methods to solve complex background interference can be divided into the following two types: extracting the contextual information in the image and improving the attention mechanism. For the complex situation of multi-directional and multi-target, because the aerial remote sensing images are all top-down images, the targets to be detected in the images are multi-directional, and the aspect ratio range of the targets to be detected is more diverse than that of the targets in the natural images, so the interference between the targets is more serious, which affects the accuracy of the final target localization and classification. For the problems of directional diversity and dense arrangement distribution of targets to be detected, there are three effective improvement ideas at present: image rotation enhancement, design of rotation invariant module, and design of accurate position regression method. To meet the challenge of drastic changes of target scales in aerial remote sensing images, the designed model needs to have good scale invariance, i.e., the model still has high recognition ability under the drastic changes of multiple scales of multiple targets, so the common improvement scheme is multi-scale feature fusion. For the detection of small targets in aerial remote sensing images, the current algorithms are mainly improved from feature enhancement, multi-level feature map detection, and the design of precise positioning strategies. In summary, for the various challenges and difficulties in object detection in aerial remote sensing images, there have been good improvement solutions for the pain points in a single direction. For example, the large size and high resolution of aerial remote sensing images inevitably lead to a complex background in the images and a sharp increase in the category and number of small targets to be detected, and most of the small targets are susceptible to strong interference from the complex background, resulting in the localization and classification recognition accuracy being At the same time, improvements for one challenge are also applicable to other difficulties, e.g., improvements for multi-scale target feature enhancement are beneficial for almost all challenges. Therefore, it is important to analyze and improve the problems in the field from a global perspective. To facilitate scholars to comprehensively understand and grasp the latest progress of aerial remote sensing image object detection research based on deep learning, based on the full study of the latest reviews and related research works, this paper systematically compares and summarizes the deep learning object detection algorithms for aerial remote sensing images, especially the research methods at home and abroad in the past three years, to provide better object detection research for aerial remote sensing images. Firstly, it introduces the deep learning-based image object detection model, then systematically composes the deep learning-based aerial remote sensing image detection methods, introduces the publicly available aerial remote sensing image object detection datasets, and compares the performance of typical methods through experiments; Finally, the problems in the current research of aerial remote sensing image object detection are given, and future research and development trends have been prospected.
Keywords

订阅号|日报