【作者】 王晨；

【导师】 朱虎明；

【作者基本信息】 西安电子科技大学 ， 计算机应用技术， 2023， 硕士

【摘要】 遥感图像是指通过人造卫星对地面进行探测所得到的图像信息。遥感图像目标检测指在遥感图像中检测出感兴趣目标的类别和位置,并将其用边界框进行标注。与自然图像目标检测相比,遥感图像目标检测具有目标多尺度、目标旋转多方向、背景复杂等特点,直接将自然图像目标检测的模型迁移至遥感图像中不能取得较好的检测效果。为此,需要针对遥感图像的特点设计目标检测模型。随着深度学习的快速发展,人们需要更多的在特定背景下满足需求的神经网络结构。但是,人工设计神经网络需要具有专业的知识能力,同时要在设备上进行大量实验来验证模型的准确性,耗费了大量的时间和资源。神经网络结构搜索(Neural Architecture Search,NAS)的出现解决了上述问题。NAS由基本的三要素组成:搜索空间定义了模型中所有可能出现的网络结构;搜索策略定义了在搜索空间中搜索网络结构的方法;性能评估策略定义了如何评估搜索出的网络结构的性能。在现有研究的基础上,围绕如何设计更适合遥感图像的目标检测模型,做了如下工作:针对基于Transformer的端到端的目标检测框架DETR(DEtection TRansformer)对小目标检测效果差、模型训练时间长的问题,设计了优化后的融合卷积Transformer的遥感图像目标检测网络OC-DETR(Oriented Convolution-DEtection TRansformer)。该网络设计了融合卷积注意力的编码器,增强了模型对图像局部信息的提取能力;设计了交叉注意力模块解耦的解码器,独立获取目标的类别信息和位置信息;设计了基于旋转感知匹配的二分图匹配损失函数,提升了集预测过程的匹配效率。为了进一步提高OC-DETR的性能,提出使用NAS的方法对该模型进行网络结构搜索。将该搜索空间设计为分层搜索空间,降低了搜索空间复杂度,提升了网络结构搜索效率。基于SPOS(Single Path One Shot)搜索策略构建了基于搜索空间的单路径超网,设计了共享权重的选择模块以及动态分布的选择模块。在遥感图像数据集DOTA和HRSC2016上的m AP分别为78.49%和88.61%。针对不同的特征提取主干网络,设计合适的特征融合网络是一个困难的任务。为此,通过NAS搜索,基于特征金字塔结构(Feature Pyramid Networks,FPN)建立了一个特征融合超网,并基于可微分搜索设计了包含卷积核复用技术和动态通道优化技术的搜索策略,同时针对模型提取图像特征能力设计了增强通道注意的主干网络。实验结果表明,本方法搜索出的特征融合网络模型相比其他对比方法在遥感图像数据集上对多尺度目标均取得了较好的检测结果,在NWPU VHR-10数据集上的m AP为93.95%。针对遥感图像目标检测的轻量级网络,在提升检测速度的同时保持检测精度是一个困难的任务。为此,本章设计了基于深度可分离卷积以及带激励模块的移动反转瓶颈卷积模块的轻量级主干网的搜索空间。设计了引入跨阶段局部网络模块的路径聚合金字塔颈部网络,解决了梯度信息重复问题,同时将部分传统卷积替换为深度可分离卷积,降低模型的内存占用。采用联合网络的One-Shot NAS搜索策略,以检测精度和模型复杂度作为监督指标,同时对主干网和颈部网络进行结构搜索。轻量级搜索空间使得One-Shot算法能搜索出更紧凑的神经网络结构,在检测性能和计算成本之间达到平衡。在NWPU VHR-10数据集和RSOD数据集上的m AP分别为91.65%和90.16%,且在NWPU VHR-10数据集上的FPS达到了75.59。更多还原

【Abstract】 Remote sensing image refers to the image information obtained from the ground detection by artificial satellites.Remote sensing image object detection refers to detecting the class and location of the object of interest in remote sensing images and labeling them with bounding boxes.Compared with natural image object detection,remote sensing image object detection has the characteristics of multi-scale object,multi-directional object rotation and complex background,etc.Directly transferring the model of natural image object detection to remote sensing image cannot achieve better detection effect.For this reason,object detection models need to be designed for the characteristics of remote sensing images.With the rapid development of deep learning,there is a need for more neural network structures that meet the needs in specific contexts.However,manual design of neural networks requires specialized knowledge capabilities and a lot of experiments on the device to verify the accuracy of the model,which consumes a lot of time and resources.The emergence of Neural Architecture Search(NAS)has solved these problems.NAS consists of three basic elements: a search space that defines all possible network structures in the model;a search strategy that defines how to search for network structures in the search space;and a performance evaluation strategy that defines how to evaluate the performance.In this thesis,based on the existing research,the following work is done around how to design a more suitable object detection model for remote sensing images:Aiming at the problem that the end-to-end object detection framework DETR(DEtection TRansformer)based on Transformer is ineffective in detecting small objects and the model training time is long,the optimized object detection network OC-DETR(Oriented Convolution-DEtection TRansformer)for remote sensing images with fused convolution Transformer is designed.The network is designed with a fused convolution-attention encoder to enhance the model’s ability to extract local information from the image;a decoder with decoupled cross-attention module to independently obtain the object’s category information and location information;and a bipartite graph matching loss function based on rotationally-aware matching to improve the matching efficiency of the set prediction process.In order to further improve the performance of OC-DETR,it is proposed to use the method of NAS to search the network structure of this model.This search space is designed as a hierarchical search space,which reduces the complexity of the search space and improves the efficiency of network structure search.A search space-based single path supernet is constructed based on the SPOS(Single Path One Shot)search strategy,and a selection module for shared weights as well as a selection module for dynamic distribution are designed.The m AP on remote sensing image datasets DOTA and HRSC2016 are 78.49%and 88.61%,respectively.Designing a suitable feature fusion network for different feature extraction backbone networks is a difficult task.To this end,a feature fusion super-network is built based on Feature Pyramid Networks(FPN)through NAS search,and a search strategy containing convolutional kernel multiplexing technique and dynamic channel optimization technique is designed based on differentiable search,while a backbone network with enhanced channel attention is designed for the model’s ability to extract image features.The experimental results show that the feature fusion network model searched by this method achieves better detection results for multi-scale objects on remote sensing image datasets compared to other comparative methods,with a m AP of 93.95% on the NWPU VHR-10 dataset.For lightweight networks for remote sensing image object detection,it is a difficult task to improve the detection speed while maintaining the detection accuracy.To this end,this chapter designs a search space for lightweight backbone networks based on depth-separable convolution as well as a moving reversal bottleneck convolution module with excitation module.A path aggregation pyramid neck network that introduces a cross-stage local network module is designed to solve the gradient information duplication problem,while replacing part of the traditional convolution with depth-separable convolution to reduce the memory footprint of the model.The One-Shot NAS search strategy of the joint network is adopted to search the structure of both the backbone and the neck network simultaneously,using the detection accuracy and model complexity as supervisory metrics.The lightweight search space enables the One-Shot algorithm to search for more compact neural network structures,striking a balance between detection performance and computational cost.The m AP on the NWPU VHR-10 dataset and RSOD dataset are 91.65% and 90.16%,respectively,and the FPS on the NWPU VHR-10 dataset reaches 75.59.更多还原