【作者】 田玲；

【导师】 李德识；

【作者基本信息】 武汉大学 ， 通信与信息系统， 2021， 博士

【摘要】 图像目标检测任务是为了给出输入图像中待测物体的名称、位置及尺寸信息,是航天任务在轨遥感图像解译中的重点应用研究方向。现有遥感技术的发展,对新时代卫星在轨数据处理能力提出了更高的要求。因此,提升遥感边缘智能技术的灵活性和实用性,实现基于深度学习方法的目标检测识别显得尤为迫切。当前普遍认为以机器视觉为代表的人工智能获得突破性发展的三个条件是:算法、算力和数据,但这三者在被引入遥感数据的图像解译研究时均存在困难。一方面,“算法”包括卷积代表的线性空间变换,但对于遥感图像特殊的成像视角及复杂场景,传统卷积操作提取到的地物特征往往不够丰富;另一方面,“算力”包括实用级边缘设备上有限的计算资源,现有高精度算法往往难以协调边缘设备上计算复杂度与功耗之间的矛盾;最后,在“数据”问题上,遥感数据集中实际有效目标样本往往容易淹没在大量无用重复背景中,使得遥感目标检测性能往往更容易受到数据规模和质量的影响。因此,针对上述三方面限制,突破在算法、算力和数据方面的局限性,需要引入对深度学习框架和有效性的深入研究与分析,并在此基础上探索适应航天任务中卫星数据特点与边缘智能计算场景的模块及结构创新。为此本文提出一种基于算法优化、算力压缩和图像特征自动增强的遥感图像目标检测系统方案,实现高效的边缘智能与在轨遥感图像目标检测。本文主要工作与贡献如下:(1)针对遥感图像智能解译对算法性能的需求,同时考虑到设备的限制,通常考虑在卷积神经网络之间进行知识迁移以实现特征提取能力与模型规模之间的平衡。但传统卷积操作因其线性变换的特性导致较弱的数据表征能力,往往只能通过非线性激活函数或重复堆叠网络层来满足特征转化需求。由于曲面流形的计算位于非欧几里得空间,在空间变换中具有天然的非线性特性,同时具有数据表征能力强和网络结构简单的双重特点。因此,本文提出一种基于非线性格拉斯曼流形网络的知识蒸馏方法,打破流形网络与卷积神经网络之间的信息隔离,将原始输入图片转换为格拉斯曼流形中的点,进而形成流形网络的输入子空间,同时利用从复杂卷积神经网络中学到的信息来指导轻量化流形网络的训练,实现网络模型更优的空间转换能力与更强的边缘设备部署适应性。在三个遥感场景分类数据集上进行了一系列的实验表明该方法分别提高了1.09%、1.73%和0.95%的分类精度,同时实现了更快的收敛速度。该方法适用于不同种类的卷积神经网络与流形网络,可承担目标检测中主干特征提取的任务。(2)针对边缘计算环境下系统算力受限的约束情况,通常可以通过降低特征提取密度来有效扩大感受野,使重点特征突出并降低计算的复杂性。但作为目前使用最广泛的稀疏密度特征提取方法的跨步卷积,该操作会失去相邻特征点的信息,导致了一些有用特征的遗漏和检测精度的降低。因此本文引入提升方案探索构建一个硬件友好的参数化、可学习、统一的神经网络架构模块基础单元。提升方案作为小波变换的快速实现方式,具有速度快,不需要辅助内存的特点,其与卷积神经网络中特征提取后采样的操作具有等效性。因此本文基于提升方案和卷积之间的近似关系提出了提升方案层作为一种新的稀疏密度特征提取方法,在提高检测精度的同时,有效降低计算复杂度。在遥感目标检测数据集的实验结果表明,该方法同时提高了检测性能和网络效率,可以实现计算复杂度和检测精度的最优。(3)由于在基础数据问题上,卷积神经网络在检测任务中对样本差异的抵抗力较差,数据规模、背景和质量的巨大差异都会影响目标检测任务的性能。针对遥感图像中目标的特点,本文提出了一种顾及目标原始特征分布的遥感图像增强方法。首先,借鉴生成式对抗网络的思想,并设计一种顾及图像色彩、内容及纹理的网络优化方法,在扩增训练集规模的同时提高了图像中目标的成像质量;此外,还设计密集特征提取模块与感受野扩展模块,提高特征在网络中的重用效率,获取更广泛的深度语义信息,增强对不同大小目标特征的提取能力。该方法在两个光学遥感图像数据集上进行了实验,可以分别实现较高的平均精度提升。(4)在上述研究基础上,构建面向边缘智能的遥感图像目标检测原型系统。分析所研究方法在国产AI芯片上的性能水准与适用场景,推动本论文研究的实际应用。评估实现边缘智能网络架构和参数更新的可行性和高效性。更多还原

【Abstract】 The task of image target detection is to give the name,position,and size information of the object to be measured in the input image.At present,the image resolution collected by remote sensing satellites is getting higher and higher,while the bandwidth of down transmission data is limited.Therefore,it is urgent to improve the on-orbit data processing capacity of satellite systems and enhance the autonomy and flexibility of remote sensing edge intelligence technology.Therefore,it is particularly urgent to conduct on-orbit data processing and realize target detection and recognition based on the deep learning method.At present,it is generally believed that artificial intelligence represented by machine vision has three bases for breakthrough development: algorithm,computing power,and data.However,these three bases are difficult to be introduced into image interpretation research of remote sensing data.On the one hand,"algorithm" includes linear space transformation represented by convolution,but for the special imaging perspective and complex scene of remote sensing images,the features extracted by traditional convolution operation are often not rich enough.On the other hand,"computing power" includes the limited computing resources on practical edge devices,and the existing high-precision algorithms are often difficult to coordinate the contradiction between computing complexity and power consumption on edge devices.Finally,in terms of "data",the actual effective target samples in remote sensing data sets are often easily submerged in a large number of useless repetition backgrounds,which makes the detection performance of remote sensing targets more susceptible to the impact of data scale and quality.Therefore,to overcome the limitations in algorithms,computing power,and data,it is necessary to introduce in-depth research and analysis on the framework and effectiveness of deep learning,and on this basis to explore module and structural innovation adapted to the characteristics of satellite data and edge intelligent computing scenarios in space missions.Therefore,this paper proposes a remote sensing image target detection system based on algorithm optimization,computational force compression,and image feature automatic enhancement to achieve efficient edge intelligence and in-orbit remote sensing image target detection.The main work and contributions of this paper are as follows:(1)In view of the requirement of intelligent interpretation of remote sensing images on algorithm performance and the limitation of equipment,knowledge transfer between convolutional neural networks is usually considered to achieve the balance between feature extraction capability and model scale.However,due to its linear transformation characteristics,the traditional convolution operation has weak data representation ability and can only meet the requirement of feature transformation through nonlinear activation function or repeated stacked network layer.Because the calculation of curved manifold is located in non-Euclidean space,it has the characteristics of natural nonlinear in space transformation,strong data representation ability,and simple network structure.Therefore,this article puts forward a kind of based on non-linear Glassman flow type network knowledge distillation method,breaking the information isolation between the manifold network with the convolutional neural network,the original input image into a point in Glassman manifolds,form manifold network input subspace,use at the same time learn from complicated convolution neural network information to guide the training of the lightweight manifold network,The network model can achieve better spatial transformation ability and stronger adaptability of edge device deployment.A series of experiments on three remote sensing scene classification datasets show that the proposed method improves the classification accuracy of 1.09%,1.73%,and 0.95% respectively,and achieves faster convergence speed.This method is suitable for different kinds of convolutional neural networks and manifold networks and can undertake the task of extracting main features in target detection.(2)In view of the constraints of system computing power in the edge computing environment,it is usually possible to effectively expand the receptive field by reducing the feature extraction density to highlight key features and reduce the complexity of calculation.However,as the most widely used sparse density feature extraction method at present,the step convolution operation will lose the information of adjacent feature points,resulting in the omission of some useful features and the reduction of detection accuracy.Therefore,this paper introduces the promotion scheme to explore the construction of a hardware-friendly parameterized,learnable,and unified neural network architecture module basic unit.As a fast implementation method of the wavelet transform,the lifting scheme has the characteristics of fast speed and no need for auxiliary memory,and it is equivalent to the operation of feature extraction and post-sampling in the convolutional neural network.Therefore,based on the approximate relationship between lifting scheme and convolution,this paper proposes the lifting scheme layer as a new sparse density feature extraction method.While improving detection accuracy,the computational complexity is almost the same as that of series convolution.Experimental results on remote sensing target detection data set show that the proposed method improves both detection performance and network efficiency,and achieves the optimization of computational complexity and detection accuracy.(3)As for the basic data problem,the resistance of convolutional neural network to sample difference in detection task is poor,and the huge difference of data size,background,and quality affects the performance of target detection task;According to the characteristics of ship target in remote sensing image,a remote sensing image enhancement method considering the original feature distribution of target is proposed in this paper.Firstly,a network optimization method considering the color,content,and texture of the image is designed based on the idea of a generative adversarial network,which improves the image quality of the ship target in the image while enlarging the training set size.In addition,the intensive feature extraction module and the receptive field expansion module are designed to improve the reuse efficiency of features in the network,obtain a wider range of deep semantic information,and enhance the feature extraction capability of targets of different sizes.The method is tested on two optical remote sensing images datasets and the average accuracy can be improved a lot.(4)Based on the above research,the prototype system of remote sensing image target detection oriented edge intelligence is constructed.The performance level and application scenarios of the proposed method in domestic AI chips are analyzed to promote the practical application of this paper.Evaluate the feasibility and efficiency of implementing edge intelligent network architecture and parameter updating.更多还原