基于单阶段检测网络的遥感图像目标检测研究

【作者】 刘涛；

【导师】 汪西莉；

【作者基本信息】 陕西师范大学 ， 计算机应用技术， 2020， 硕士

【摘要】 目标检测是对图像中感兴趣的物体进行识别和定位,是图像理解和应用的基础任务。遥感图像目标检测是目标检测应用的重要研究方向,被广泛应用于城市规划、灾害监测、军事制导和无人机导航等领域,具有很高的军用和民用价值。随着深度学习目标检测理论的发展,遥感图像目标检测性能取得了巨大的提升,然而遥感图像的高分辨率、目标尺度差异大且分布不均匀、空间场景复杂难以与目标区分等特点,增大了遥感图像目标检测的研究难度,导致遥感图像目标检测的发展滞后于普通图像目标检测的发展进程。本文在现有研究和深度学习目标检测理论的基础上,对单阶段检测网络进行改进,使其更好地应用于遥感图像目标检测领域,提出快速有效的基于单阶段检测网络的遥感图像目标检测模型。其创新点主要有三个方面。(1)在卷积神经网络中,浅层特征图包含丰富的细节信息但缺乏语义信息,深层特征图包含丰富的语义信息但缺乏细节信息,为了利用不同深度特征图的优势以及检测多尺度目标,提出采用卷积核金字塔和空洞卷积的单阶段目标检测模型(Single-Stage Detection based on Filter Pyramid and Atrous convolution,AFP-SSD)。首先,所提模型使用像素相加方式融合多层不同分辨率的特征图,并在通道维度拼接不同阶段的特征图,形成具有丰富语义信息和细节信息的特征信息融合层作为模型的预测特征层。其次,所提模型在锚框机制中引入卷积核金字塔结构以检测多尺度的目标,同时采用空洞卷积减少大尺寸卷积核增加的参数量,合理地降低锚框数量。实验结果表明,所提模型不仅具有较快的检测速度和较高的精度,而且较好地解决了小目标以及重叠目标难以检出的问题。(2)提出单阶段遥感目标检测模型:特征注意力金字塔网络FAPNet(Feature Attention Pyramid Network)。首先,FAPNet模型使用特征注意力金字塔模块融合不同特征层的特征信息,形成多个不同分辨率的预测特征层,提升遥感图像中不同尺度目标的检测精度。其次,使用弱监督注意力模块调整预测特征层的特征分布,强化目标区域特征,提升遥感图像中遮挡目标的检测精度。在特征注意力金字塔模块中,使用通道拼接方式融合不同层的特征图,并使用SE(Squeeze-and-Excitation)注意力机制对融合特征图在通道维度重标定,自适应地调整不同层特征图在融合特征图中的丰富度。在弱监督注意力模块中,将分割结果预测值作为位置注意力权重,调整预测特征层的特征分布,提升目标检测网络的性能。实验结果表明,所提模型在两个遥感图像数据集上超越基线模型和其他先进模型,取得较好的检测精度。(3)不同尺度目标和遮挡目标的检测问题是遥感图像目标检测研究中的难点问题。特征金字塔结构是提高不同尺度目标检测精度的有效方法,但是在特征金字塔结构中,逐层传递特征信息的方式容易产生特征信息丢失问题,因此提出跳跃连接特征金字塔模块进一步增强特征金字塔结构中各个特征层的语义和细节信息。另外,用位置注意力强化目标区域特征是提高遮挡目标检测精度的有效方法,但是以往的位置注意力会同时强化产生不精确预测结果的特征区域,对最终预测结果产生干扰,因此提出基于锚框的位置注意力模块,聚焦于强化更可能产生精确预测结果的特征区域。将跳跃连接特征金字塔模块和基于锚框的位置注意力模块嵌入到RetinaNet模型中,形成端到端的单阶段遥感目标检测模型:基于锚框的注意力网络AANet(Anchor-based Attention Network)。实验结果表明,AANet模型是一个快速有效的单阶段遥感目标检测器。更多还原

【Abstract】 Object detection is to identify and locate objects of interest in the image,which is a basic task for image understanding and application.Remote sensing image object detection is an important research direction of object detection application.It can be widely used in urban planning,disaster monitoring,military guidance,drone navigation and other fields,and has high military and civilian value.With the development of object detection theory based on deep learning,the performance of remote sensing image object detection has been greatly improved.However,the characteristics of remote sensing image,such as high resolution,large scale variation and uneven distribution,complex scene and difficult to distinguish from the object,increase the research difficulty of remote sensing image object detection,resulting in the development of remote sensing image object detection still lags behind the development of ordinary image object detection.Based on existing research and deep learning object detection theory,Improves the single-stage detection network to make it better used in the field of remote sensing image object detection,and forms a fast and effective remote sensing image object detector based on the single-stage detection network.There are three main innovations.(1)In the convolution neural network,the shallow feature map contains rich details but lacks semantic information,while the deep feature map contains rich semantic information but lacks detailed information.In order to make use of the advantages of different depth feature maps and detect multi-scale objects,a single-stage object detection model using convolution filter pyramid and atrous convolution is proposed.Firstly,multi-layer feature maps of different resolutions are fused by pixel-by-pixel addition,and then feature maps of different stages are concatenated to form the fusion feature layer with rich semantic information and detail information as the prediction feature layer of the model.Secondly,the convolution filter pyramid structure is introduced into the anchor mechanism to detect multi-scale objects.At the same time,atrous convolution is used to reduce the parameters added by large-scale convolution filter and reasonably reduce the number of anchors.The experimental results show that the proposed model not only has faster detection speed and higher accuracy,but also solves the problems of small objects and occluded objects that are difficult to detect.(2)A one-stage remote sensing object detection model:Feature Attention Pyramid Network(FAPNet).Firstly,the FAPNet model uses the feature attention pyramid module to fuse the feature information of different feature layers to form multiple prediction feature layers with different resolutions,which improves the detection accuracy of objects at different scales in remote sensing images.Secondly,the weak supervised attention module is used to adjust the feature distribution of the prediction feature layer,strengthen the object region features,and improve the detection accuracy of occluded objects in remote sensing images.In the feature attention pyramid module,the feature maps of different layers are fused by means of channel concatenation,and the SE(Squeeze-and-Excitation)attention module is added to refine the fusion feature map in the channel dimension to adaptively adjust the richness of different feature layers in the fusion feature map.In the weak supervised attention module,the prediction values of segmentation results are taken as the spatial attention weight,adjust the feature distribution of prediction feature layers and improve the performance of object detection network.The experimental results show that the proposed model surpasses the baseline model and other advanced models in two remote sensing image data sets and achieves better detection accuracy.(3)Scale variation across object instances and occlusion issue are still challenging research topics in the remote sensing object detection task.Feature pyramids are an effective method for detecting objects at different scales.The way of transferring feature information layer by layer will lose feature information in the feature pyramids.Therefore,a feature pyramid network with shortcut connections is proposed,which can enhance the semantic and detail information of each feature layer of feature pyramid network.Moreover,using spatial attention to strengthen the possible object area feature is an effective method to solve the occlusion problem.But the available spatial attention will strengthen feature regions that produce the imprecise prediction results simultaneously,thus will interfere with the final prediction results.For this purpose,an anchor-based spatial attention module is proposed.It mainly strengthens feature regions that are more likely to produce accurate prediction results.The feature pyramid network with shortcut connections and the anchor-based spatial attention module are embedded into the RetinaNet to form an end-to-end single-stage remote sensing object detector,namely AANet(Anchor-based Attention Network).The experimental results show that AANet is a fast and effective single-stage remote sensing object detector.更多还原