一种基于自动特征学习的陨石坑区域检测算法

基于陨石坑的视觉导航技术成为一种新颖的高精度空间探测自主导航方式，如何从导航图像中精确地提取陨石坑区域是实现基于陨石坑视觉导航的首要条件。针对这一问题，根据陨石坑导航图像特点，提出了一种基于自动特征学习的陨石坑区域检测算法。首先，基于最大稳定极值区域检测算法提取陨石坑候选区域；其次，利用卷积神经网络（CNN）自动学习提取候选区域的特征；最后，通过支持向量机（SVM）实现候选区域的精确分类，得到真实的陨石坑区域。大量的仿真实验表明:与传统的基于人工特征的陨石坑区域检测算法相比，提出的基于自动特征学习的陨石坑区域检测算法具有更高的检测精度和更好的鲁棒性，在通用火星表面陨石坑数据集上，所提算法的F₁度量指标较于传统算法高出8%，可以广泛地应用于基于陨石坑的视觉导航算法中的陨石坑区域提取，为基于陨石坑视觉导航算法提供精确的导航路标输入。      

基于深度学习网络的星表非结构化岩石目标辨识方法研究

目标检测技术广泛应用于交通、医疗、安保和航天等领域.目前,目标检测技术面临目标微弱、背景复杂、目标被遮挡等挑战[1].针对星表非结构化模拟地形采集的图像中岩石和石块小目标检测识别率低、误识别率高的问题,研究了当下效果最好、模型轻量化的YOLOv5目标检测算法,在其基础上进行改进优化器与优化检测框重复检测效果的一种满足实时性要求的岩石目标检测算法.具体通过引入空标签负样本、结合随机梯度下降法SGD(stochastic gradient descent)优化模型和非极大值抑制参数调节方法,提升YOLOv5网络模型的特征描述能力和分类准确率.利用在地面试验场采集的复杂地形图像作为数据集,并采用mAP(mean average pre-cision)、画面每秒传输帧数(FPS)、准确率和召回(P-R)曲线等作为性能指标,对所提出的目标检测网络进行了试验验证.结果 表明本文提出的改进网络拥有更高的准确率和更低的虚警率,同时保持原有算法的实时性.     

改进YOLOv4的表面缺陷检测算法

为解决航空发动机部件表面缺陷检测精度低、检测速度慢的问题,提出一种改进的YOLOv4算法进行智能检测。在路径聚合网络（PANet）结构中融合浅层特征与深层特征,增大特征检测尺度,同时去除自下而上的路径增强结构,提高小目标检测精度和整体检测速度;根据各类缺陷数量不同的情况,优化聚焦损失中的平衡参数,增加权重因子调节各类缺陷的损失权重,将改进后的聚焦损失代替分类误差中的交叉熵损失函数,降低样本不平衡和难易样本对检测精度的影响。实验表明：相比于原始YOLOv4算法,改进后的YOLOv4算法在测试集上的平均精度均值（mAP）为90.10%,提高了2.17%;检测速度为24.82 fps,提高了1.58 fps,检测精度也高于单发多框检测（SSD）算法、EfficientDet算法、YOLOv3算法和YOLOv4-Tiny算法。     

遥感图像飞机目标高效搜检深度学习优化算法

为了实现大幅面遥感图像中飞机目标的高效检测与准确定位,通过深度神经网络（DNN）的级联组合,提出了一种新颖的搜寻与检测相集成的飞机目标高效检测算法。首先,运用高性能的端到端DNN网络,进行停机坪与跑道区域的像素级高效精准分割,从而大幅度缩小飞机目标的搜索范围,以降低虚警发生概率,完成飞机目标候选检测区域的快速搜寻。然后,针对分割所得停机坪与跑道区域,借助手工数据集对YOLO网络模型进行迁移式强化训练,一方面可以弥补训练集在样本类型与数据规模上的不足,另一方面借助YOLO网络的强时效性优势对飞机目标的位置进行回归求解,可以显著提高飞机目标的检测效率。停机坪与跑道区域分割DNN网络在分割精度与时效性上具有显著优势,而迁移式强化训练YOLO网络不仅具有很高的检测效率,在检测精度上也能保持良好的性能。通过一系列综合实验与对比分析,验证了提出的搜寻与检测相集成的DNN级联组合式飞机目标高效检测算法的性能优势。      

一种新的基于信号导频的射频指纹识别方法

现有基于深度学习的射频指纹识别技术大多采用原始数据样本作为网络输入，未考虑信号携带内容对分类结果产生的影响，网络结构相对单一。为此，针将信号导频部分作为网络输入展开了研究，提出了一种新的导频提取算法，对10个ADALM-PLUTO软件定义无线电设备(SDR)辐射出的信号提取其导频，并建立了3种不同距离条件下的导频数据集。提出将Inception网络结构用于射频指纹识别，在10 m无线传输距离下达到了98.58%的分类精度，相较于现有基于AlexNet网络改进的卷积神经网络(CNN)，分类精度有所提升。      

一种基于多尺度边缘提取的陨石坑检测算法

针对星体表面的陨石坑可用于探测器的自主导航、障碍识别等任务,提出一种基于多尺度边缘提取的陨石坑检测算法。该算法首先利用高斯金字塔得到不同尺度的陨石坑图像;其次,针对不同尺度的陨石坑图像,利用EDPF边缘提取算法对陨石坑进行边缘提取,并连接关键边缘像素点为直线段来近似表示图像边缘;然后将具有相同偏转方向的边缘直线段连接成圆弧,并将有相似半径和中心的圆弧拟合成候选圆和椭圆;最后对候选圆、椭圆进行验证。该算法的优点在于,能够准确地检测出陨石坑,有较高的检测率,且对存在较多陨石坑的图像有较好的检测结果。     

一种复杂场景下的人眼检测算法

针对复杂场景下的人眼检测问题，间接方法和直接方法具有一定的局限性。提出了一种不依赖人脸检测的直接型人眼检测算法，以解决复杂场景下多尺度尤其是小尺度人眼检测问题。算法通过减少下采样因子并加入扩张残差单元以提升小尺度人眼检测能力，且对多尺度特征相互拼接以保证多尺度人眼检测的精度。同时，算法借助于压缩特征输出通道降低了模型复杂度，使人眼检测效率得以提升。实验结果表明:所提模型可以在小尺度下有效地进行左右眼区分，并在红外数据上表现良好。经在DIF数据集上进行训练与测试，所提模型在较小尺度下人眼检测精度达到82.59%，检测效率达到30.5 fps。      

显著性引导的低光照人脸检测

针对卷积神经网络难以对低光照环境拍摄的图像进行人脸检测的问题。提出了一种将图像显著性检测算法和深度学习相结合的算法，并应用于低光照人脸检测。所提算法将图像的显著性信息与图像原始RGB通道融合，用于神经网络训练。在低光照人脸数据集DARK FACE上进行了充分的实验，结果表明:所提方法在DARK FACE数据集上获得了比当前主流人脸检测算法更好的检测精度，进而验证了所提算法的有效性。      

基于R-D SSD模型航空发动机安装工位检测算法

为解决航空发动机在安装过程中大多实行人工安装、定位不精确等问题，在研究其自动化安装方法中，针对航空发动机安装工位的检测需求，提出了一种残差网络与膨胀卷积相融合的SSD改进算法(R-D SSD)。将经典SSD模型的主干网络VGG16替换为残差网络ResNet-101，并增加其输出特征图上的预选框数量，解决了原始算法对底层特征抓取能力不足的问题，进而弥补了对小目标检测效果较差的缺陷；利用膨胀卷积扩大网络的感受野，获取足够的安装工位边缘特征细节信息，在不改变网络结构的同时，保证了模型良好的实时性和对目标的检测精度。实验表明:对于小目标数据集和整个数据集，R-D SSD算法的平均检测精度较原始算法分别提高了8.6%和4.0%，可以满足航空发动机安装时平均检测精度不低于85%的要求。      

基于层间分类一致性准则的舰船目标细粒度识别方法

舰船目标的层次化、细粒度识别在军事和民用领域均有重要意义。现有细粒度识别方法一般需要部件级精细标注或采用注意力机制提取关键特征，但并未有效利用舰船目标层次化分类体系中本身所蕴含的隶属关系信息提高细粒度识别精度。针对舰船目标的层次化分类问题，建立了舰船目标多层级一致性分类数学模型，提出了一种基于层间强一致性分类准则的细粒度识别方法，设计了层间一致性分类损失函数，并构建了多层级兼容舰船目标细粒度识别网络(MLCDet)。经试验验证，该方法有效、鲁棒，资源开销小，能够有效利用分类体系中各类别间的隶属关系提升目标识别精度。在无需部件级标注信息的前提下，将mAP提高了1.3%,与此同时，模型总参数量仅增加0.02%,推断速度不变。     

Bayesian network-based extraction of lunar impact craters from optical images and DEM data

Impact craters are among the most noticeable geomorphological features on the planetary surface and yield significant information about terrain evolution and the history of the solar system. Thus, the recognition of impact craters is an important branch of modern planetary studies. Aiming at addressing problems associated with the insufficient and inaccurate detection of lunar impact craters, a decision fusion method within the Bayesian network (BN) framework is developed in this paper to handle multi-source information from both optical images and associated digital elevation model (DEM) data. First, we implement the edge-based method for efficiently searching crater candidates which are the image patches that can potentially contain impact craters. Secondly, the multi-source representations of an impact crater derived from both optical images and DEM data are proposed and constructed to quantitatively describe the two-dimensional (2D) and three-dimensional (3D) morphology, consisting of Histogram of Oriented Gradient (HOG), Histogram of Multi-scale Slope (HMS) and Histogram of Multi-scale Aspect (HMA). Finally, a BN-based framework integrates the multi-source representations of impact craters, which can provide reductant and complementary information, for distinguishing craters from non-craters. To evaluate the effectiveness and robustness of the proposed method, experiments were conducted on three lunar scenes using both orthoimages from the Lunar Reconnaissance Orbiter (LRO) and DEM data acquired by the Lunar Orbiter Laser Altimeter (LOLA). Experimental results demonstrate that integrating optical images with DEM data significantly decreases the number of false positives compared with using optical images alone, with F₁-score of 84.8% on average. Moreover, compared with other existing fusion methods, our proposed method was quite advantageous especially for the detection of small-scale craters with diameters less than 1000 m.     

A machine learning approach to crater detection from topographic data

Craters are distinctive features on the surfaces of most terrestrial planets. Craters reveal the relative ages of surface units and provide information on surface geology. Extracting craters is one of the fundamental tasks in planetary research. Although many automated crater detection algorithms have been developed to exact craters from image or topographic data, most of them are applicable only in particular regions, and only a few can be widely used, especially in complex surface settings. In this study, we present a machine learning approach to crater detection from topographic data. This approach includes two steps: detecting square regions which contain one crater with the use of a boosting algorithm and delineating the rims of the crater in each square region by local terrain analysis and circular Hough transform. A new variant of Haar-like features (scaled Haar-like features) is proposed and combined with traditional Haar-like features and local binary pattern features to enhance the performance of the classifier. Experimental results with the use of Mars topographic data demonstrate that the developed approach can significantly decrease the false positive detection rate while maintaining a relatively high true positive detection rate even in challenging sites.     

A new approach based on crater detection and matching for visual navigation in planetary landing

This paper provides an approach of crater detection and matching to visual navigation in planetary landing missions. The approach aims to detect craters on the planetary surface and match them to a landmark database during the descent phase of a planetary landing mission. Firstly an image region pairing method is proposed to detect the crater by using an image region feature detector. Then a WTA-rule is adopted to match the detected crater to the crater in database. To further reduce the false matching rate, an efficient method for reducing false matches using parameters of crater in 3-D database is proposed. Real images of planetary terrain and a semi-physical planetary landing simulation platform are utilized to test the performance of the approach, simulation results show the proposed approach is able to match the required number of craters to the database for pin-point planetary landing with a low rate of false detection and false matching, which will lead to an improved planetary landing precision.     

基于深度学习和图匹配的接线图检测与校核

传统的厂站一次接线图的绘制和管理主要依靠电网运行人员，费时费力且缺乏科学可校核的参考标准。提出了一种基于深度神经网络和数字图像处理相结合的厂站一次接线图的自动检测、识别和校核算法。首先，使用目标检测Faster R-CNN模型检测厂站接线图中的电器元件，并达到92%的检测准确率，同时使用端到端的文字检测识别模型识别厂站接线图中的文字信息，并达到94.2%的文字检测准确率和92%的文字识别准确率；然后，使用数字图像处理技术进行厂站接线图连接线、拓扑关系识别；最后，使用改进的VF2算法进行厂站一次接线图和人工维护的厂站一次接线图拓扑关系匹配校核，将拓扑数据抽象为无向图，通过轮廓序号得到元件的相对位置信息，根据改进的VF2算法得到2张图的匹配率，并通过匹配率与设定好的阈值来帮助核验，相比于节点遍历的匹配方法，核验准确率提高了37.5%。基于某供电公司提供的部分变电站的厂站一次接线图标注了接线图电器元件，贡献了一个小型接线图数据集。      

Single crater-aided inertial navigation for autonomous asteroid landing

In this paper, a novel crater-aided inertial navigation approach for autonomous asteroid landing mission is developed. It overcomes the major deficiencies of existing approaches in the literature, which mainly focuses on the case where craters are abundant in the camera field of view. As a result, traditional crater based methods require at least three craters to achieve crater matching, which limits their application in final landing phase where craters are scarce in the camera’s field of view. In contrast, the proposed algorithm enables single crater based crater matching based on a novel 2D-3D crater re-projection model. The re-projection model adopts inertial measurements as a reference, and re-projects the 3D crater model onto descent images to achieve the matching to its counterpart. An asteroid landing simulation toolbox is developed to validate the performance of the proposed approach. Through comparison with the state-of-the-art local image feature and crater based navigation algorithms, the proposed approach is validated to achieve a competitive performance in terms of feature matching and pose estimation accuracy with a much lighter computational cost.     

Detection of sub-kilometer craters in high resolution planetary images using shape and texture features

Counting craters is a paramount tool of planetary analysis because it provides relative dating of planetary surfaces. Dating surfaces with high spatial resolution requires counting a very large number of small, sub-kilometer size craters. Exhaustive manual surveys of such craters over extensive regions are impractical, sparking interest in designing crater detection algorithms (CDAs). As a part of our effort to design a CDA, which is robust and practical for planetary research analysis, we propose a crater detection approach that utilizes both shape and texture features to identify efficiently sub-kilometer craters in high resolution panchromatic images. First, a mathematical morphology-based shape analysis is used to identify regions in an image that may contain craters; only those regions - crater candidates - are the subject of further processing. Second, image texture features in combination with the boosting ensemble supervised learning algorithm are used to accurately classify previously identified candidates into craters and non-craters. The design of the proposed CDA is described and its performance is evaluated using a high resolution image of Mars for which sub-kilometer craters have been manually identified. The overall detection rate of the proposed CDA is 81%, the branching factor is 0.14, and the overall quality factor is 72%. This performance is a significant improvement over the previous CDA based exclusively on the shape features. The combination of performance level and computational efficiency offered by this CDA makes it attractive for practical application.     

Crater detection via genetic search methods to reduce image features

Recent approaches to crater detection have been inspired by face detection’s use of gray-scale texture features. Using gray-scale texture features for supervised machine learning crater detection algorithms provides better classification of craters in planetary images than previous methods. When using Haar features it is typical to generate thousands of numerical values from each candidate crater image. This magnitude of image features to extract and consider can spell disaster when the application is an entire planetary surface. One solution is to reduce the number of features extracted and considered in order to increase accuracy as well as speed. Feature subset selection provides the operational classifiers with a concise and denoised set of features by reducing irrelevant and redundant features. Feature subset selection is known to be NP-hard. To provide an efficient suboptimal solution, four genetic algorithms are proposed to use greedy selection, weighted random selection, and simulated annealing to distinguish discriminate features from indiscriminate features. Inspired by analysis regarding the relationship between subset size and accuracy, a squeezing algorithm is presented to shrink the genetic algorithm’s chromosome cardinality during the genetic iterations. A significant increase in the classification performance of a Bayesian classifier in crater detection using image texture features is observed.     

基于Faster R-CNN的精密零部件的识别方法

在航空航天领域中，惯性陀螺等精密器件装配精度要求较高，目前大多采用人工装配的方法，装配效率低、装配过程受人主观影响大。针对上述存在的问题，采用基于Faster R-CNN模型的目标识别算法，通过VGG16特征提取网络提取特征信息，在模型训练过程中利用COCO数据集的深度网络模型进行迁移训练，防止模型过拟合并加速参数的训练过程。同时，该方法还与其他深度学习模型以及传统的目标识别算法进行了对比，在自建的数据模型测试集上进行试验。结果表明，基于VGG16的Faster R-CNN目标识别模型在复杂环境及物体发生遮挡的情况下对于惯性陀螺的识别具有明显的优势，准确率可达到87.80%，召回率80.30%，识别速度可达到15FPS，能够满足实时性要求。     

免疫支持向量机用于航空发动机磨损故障诊断

航空发动机在使用寿命周期内会不断磨损最终出现故障,通过对发动机油液监测铁谱分析数据的挖掘可实现磨损故障的诊断。本文研究免疫算法优化的支持向量机(SVM)在航空发动机磨损故障诊断中的运用。首先,总结了支持向量机和免疫算法的运行流程和关键算法。然后,用改进的免疫算法优化支持向量机惩罚因子、松弛变量及核函数参数。某型航空发动机的油液铁谱分析数据和加入噪声数据验证结果表明,该方法可有效实现航空发动机磨损故障诊断且具有较好的鲁棒性。最后,研究了核函数、多分类决策方法、初始种群大小、亲和力计算公式、支持向量机优化方法和归一化方法对磨损故障诊断准确率的影响,得到了最佳诊断方法。