【作者】 张帅；

【导师】 纪文宇；

【作者基本信息】 吉林大学 ， 光学， 2024， 硕士

【摘要】 遥感目标检测是对利用遥感技术获取的图像数据,通过计算机视觉算法进行处理,识别和提取图像中的目标物体信息的一种技术。在遥感图像中,目标物体可以是建筑、道路、桥梁、农田、河流,以及轮船、飞机和汽车等人造物体。随着深度学习和计算机视觉技术的进步,以及遥感卫星和无人机技术的发展,遥感目标检测技术逐渐在农业技术、城市规划、交通管理和军事侦查等领域发挥越来越重要的作用。但是面向遥感图像的旋转目标检测任务仍然是一项具有挑战性的任务。许多基于深度卷积神经网络的遥感目标检测算法在近年来被不断提出,在这些算法中,标签分配策略对检测任务效果的提升起着至关重要的作用。然而,现有的基于交并比的标签分配策略无法准确捕获旋转物体的形状和方向信息,不能有效地区分和真实框的交并比相同,但重叠情况不同的候选样本的定位潜力。此外,在遥感目标检测网络中,分类任务和定位任务之间的不一致可能会降低模型的检测性能,这种不一致体现在两个方面:一方面是分类与定位任务之间的置信度分数不一致,另一方面是分类和定位分支的特征敏感区域不一致。本文旨在解决上述问题并做出以下贡献:本文首先对深度学习和目标检测技术进行了介绍,然后介绍了现有的针对遥感图像的目标检测算法。同时详细说明了遥感目标检测算法的数据处理研究和作用机制。为了解决上述问题,本文提出了一种用于遥感目标检测的无锚检测器——高斯感知旋转检测器GARDet(Gaussion Aware Rotated Detector)。它包含两项改进:高斯动态标签分配策略GDLA(Gaussian Dynamic Label Assignment Strategy)和特征对齐模块FAM(Feature Alignment Module)。高斯动态标签分配策略使用基于候选样本与真实框之间的高斯分布,和预测框与真实框之间的高斯距离设计的高斯匹配置信度GMC(Gaussian Matching Confidence),作为一种新的衡量候选样本定位质量的判据,并依据高斯匹配置信度GMC动态地分配用于训练的正负样本。为了提取不同的任务敏感特征,特征对齐模块FAM由旋转特征对齐卷积OFA(Oriented Feature Alignment)和空间特征对齐卷积SFA(Spatial Feature Alignment)两部分组成。旋转特征对齐卷积提取对分类任务有益的物体内部特征和旋转不变特征。空间特征对齐卷积提取对定位任务有益的空间坐标特征。此算法在涵盖多个场景、多模态图像和多个对象类别的多个公共数据集上进行了大量实验,实验结果证明了所提出的模型可以在几个具有挑战性的遥感目标检测数据集(DOTA数据集、HRSC2016数据集和UCAS-AOD数据集)上达到了较好结果。在DOTA数据集实验中,GARDet实现了79.97%的检测性能,在无锚遥感目标检测方法中达到了最好的结果。而在UCAS-AOD和HRSC2016数据集的实验中,GARDet分别达到了97.95%和96.70%的检测精度,实现了最先进的检测精度。综上所述,论文对光学遥感图像目标检测与识别中存在的难点进行了分析,并对相关理论进行了研究。本文基于深度学习的卷积神经网络,提出了一种新型的针对遥感目标的目标检测算法,并通过大量实验证明了算法的有效性。论文的研究成果为光学遥感目标检测技术的发展提供了理论基础,并为该领域的研究和应用做出了贡献。更多还原

【Abstract】 Remote sensing object detection is a technology that processes image data obtained using remote sensing technology through computer vision algorithms to identify and extract object information in the image.In remote sensing images,objects can be buildings,roads,bridges,farmland,rivers,and artificial objects such as ships,airplanes,and cars.With the advancement of deep learning and computer vision technology,as well as the development of remote sensing satellite and drone technology,remote sensing object detection technology has gradually played an increasingly important role in agricultural technology,urban planning,traffic management,and military reconnaissance.However,the rotated object detection task for remote sensing images is still challenging.Many remote sensing object detection algorithms based on deep convolutional neural networks have been continuously proposed in recent years.Among these algorithms,label assignment strategies are crucial in improving the detection task effect.However,existing label assignment strategies based on the Intersection-over-Union Ratio cannot accurately capture rotated objects’ shape and orientation information.They cannot effectively distinguish the localization potential of candidate samples with the same Intersection-over-Union with the ground truth but in different overlapping situations.In addition,in the remote sensing object detection network,the inconsistency between the classification task and the localization task may reduce the detection performance of the model.This inconsistency is reflected in two aspects: On the one hand,the confidence scores between the classification and positioning tasks are inconsistent.On the other hand,the feature-sensitive areas of the classification and localization branches are inconsistent.This article aims to address the above issues and make the following contributions:This article first introduces deep learning and object detection technology.Then,we introduce existing object detection algorithms for remote sensing images.At the same time,the data processing research and mechanism of the remote sensing object detection algorithm are explained in detail.In order to solve the above problems,this paper proposes an anchor-free detector for remote sensing object detection: Gaussian Aware Rotated Detector(GARDet).It contains two improvements: Gaussian Dynamic Label Assignment Strategy(GDLA)and Feature Alignment Module(FAM).The Gaussian dynamic label assignment strategy uses Gaussian Matching Confidence(GMC)based on the Gaussian distribution between the candidate sample and the ground truth and the Gaussian distance between the predicted box and the ground truth as a new criterion for measuring the localization quality of the candidate sample.According to the Gaussian matching confidence,positive and negative samples for training are dynamically allocated.In order to extract different task-sensitive features,the feature alignment module FAM consists of two parts: Oriented Feature Alignment convolution OFA and Spatial Feature Alignment convolution SFA.Oriented Feature Alignment convolution extracts object internal features and rotation-invariant features beneficial for classification tasks.Spatial Feature Alignment convolution extracts spatial coordinate features beneficial for localization tasks.This paper conducts extensive experiments on this algorithm on multiple public datasets covering multiple scenes,multi-modal images,and multiple object categories.Experiment results demonstrate that the proposed model can achieve good performance on several challenging remote sensing object detection datasets(the DOTA dataset,HRSC2016 dataset,and UCAS-AOD dataset).In the DOTA dataset experiment,GARDet achieves a detection performance of 79.97%,achieving the best results among anchor-free remote sensing object detection methods.In experiments on the HRSC2016 and UCAS-AOD datasets,GARDet achieves detection accuracy of 97.95% and 96.70%respectively,achieving the most advanced detection accuracy.In summary,this paper analyzes the difficulties existing in object detection and recognition in optical remote sensing images and studies related theories.This paper proposes a new object detection algorithm for remote sensing based on the convolutional neural network of deep learning.It proves the effectiveness of the algorithm through a large number of experiments.The research results of this paper provide a theoretical basis for the development of optical remote sensing object detection technology and contribute to research and application in this field.更多还原