区域增强和多特征融合的X光图像违禁品识别

目的 对旅客行李进行安全检查是维护公共安全的措施之一,安检智能化是未来的发展方向。基于X光图像的安检因不同的安检机成像方式不同,同一类违禁品在不同设备上的X光图像在颜色分布上有很大差异,导致安检图像智能识别算法在训练与测试数据分布不同时,识别性能明显降低,同时X光行李图像中物品的混乱复杂增加了违禁品识别的难度。针对上述问题,本文提出一种区域增强和多特征融合模型。方法 首先,通过注意力机制的思想提取一种区域增强特征,消除颜色分布不同的影响,保留图像整体结构并增强违禁品区域信息。然后,采用多特征融合策略丰富特征信息,使模型适用于图像中物品混乱复杂情况。最后,提出一种三元损失函数优化特征融合。结果 在公开数据集SIXray数据集上进行整体识别性能和泛化性能的实验分析,即测试本文方法在相同和不同颜色分布样本上的性能。在整体识别性能方面,本文方法在平均精度均值(mean average precision,mAP)上相较于基础模型ResNet18和ResNet34分别提升了4.09%和2.26%,并优于一些其他识别方法。对于单类违禁品,本文方法在枪支和钳子类违禁品上的平均识别精度为94.25%和...     

Creating a reference database of cargo inspection X-ray images using high energy radiographs of cargo mock-ups

Customs continue to use a wide range of technology in protecting against terrorism and the movement of illicit trade and prohibited imports. The throughput of scanned vehicles and cargo increases and just keeps on growing. Therefore, the need of automated algorithms to help screening officers in inspection, examination or surveillance of vehicles and containers is crucial. In this context, the successful collaboration between manufacturers and customs offices is of key importance. Facing this topic, within the seventh framework program of the European Commission, the project ACXIS “Automated Comparison of X-ray Images for cargo Scanning” arose. The main objective of this project is to develop a manufacturer independent reference database for X-ray images of illicit and licit cargo. Historic images of real detections, images of illegal cargo mock-ups as well as images of legitimate cargo will be integrated into the reference database. For this, procedures and algorithms to uniform X-ray images of different cargo scanners was developed, as well as an automated identification method of potentially illicit cargo. Finally, these developments were incorporated in creating a training simulator and a toolbox for inspection officers enhanced X-ray screening competence.     

基于域适应的X光图像的目标检测

随着卷积神经网络的发展,X光安全检查图像的自动目标检测算法已经取得了重大进步。但是,当将这些目标检测算法应用到不同于训练集数据的新数据,即训练域数据和测试域数据的图像数据服从不一致的分布时,这些检测算法的性能通常会降低。根据X光成像的变化,提出一种基于上下文的透射率自适应域对齐方法,用于解决检测算法的域不适应问题。首先,通过利用X光图像中存在的颜色信息,设计了一种注意力机制来分别处理X光图像的每个颜色通道特征,解决不同X光机器的颜色差异问题。接着,提出一种多分辨率特征对齐方法,以解决不同厂商不同X光图像之间的数据分布差异。最后,使用上下文向量作为对抗训练的正则化,利用邻域信息提高测试精度。基于X光图像数据集和Cityscape数据集的实验表明,所提方法解决了目标检测算法在不同于训练域数据的测试域中精度下降的问题。     

基于三级图像融合与CLAHE的机场安检X光图像增强算法

针对CLAHE增强机场安检X光图像出现的颜色失真及背景噪声放大问题, 提出一种基于三级图像融合与CLAHE的X光图像增强算法. 首先将X光图像分别转换成RGB、HSV图像并分别进行CLAHE增强, 将增强后的图像通过欧几里得范数实现第1级融合. 然后将融合后的图像进行USM锐化, 锐化过程中根据掩膜进行图像第2级融合. 最后将锐化后的图像与原图像按系数进行合并, 完成第3级融合. 仿真实验结果表明, 本文提出的算法能够有效提高安检X光图像的对比度, PSNR平均值提高了7 dB, 同时可以抑制增强图像颜色失真及背景噪声. 该算法有助于提高X光图像中违禁物品的识别准确率, 对平安机场的构建具有积极意义.     

基于Polygon-RefineNet的违禁品X线图像自动标注方法

近年来,随着深度学习的快速发展,其在智慧安检领域的应用已经成为了当下的研究热点.众所周知,深度学习方法是以海量训练数据为基础的,然而手工标注真值(ground truth)是一项十分繁琐的工作.为此,本文提出一种基于Polygon-RefineNet的违禁品X线图像自动标注方法,该方法在用户设定的包含感兴趣区域的初始边框(bounding box)内自动预测出目标的多边形轮廓,旨在生成可用真值的情况下最大限度地减少标注时间.由于违禁品X线图像存在大量的重叠现象,导致图像背景十分杂乱、违禁品轮廓模糊不清,因此本文首先引入多路径优化机制,通过有效利用深度网络下采样过程中提取的底层空间信息和高层语义信息来优化多边形轮廓的边缘细节,从而提高标注精度;其次,本文设计一种混合损失函数用于优化多边形轮廓的整体形状和位置,并同时消除真值本身存在的主观性误差使模型具有强大的泛化能力.最后,为了验证所提出方法的有效性,本文建立了一个违禁品X线数据集,该数据集包含2623张经过手工标注的X线图像,共10类7257个违禁品带有像素级真值.实验表明,本文提出的方法在标注违禁品时达到了93.1%的准确率,且速度约是手工标注的3.7倍.本文进一步证明了该方法在Cityscapes数据集、MS COCO数据集等其它域外数据集上的有效性.     

Change of search time and non-search time in X-ray baggage screening due to training

As found in studies of aircraft structural inspection, the time used for judging if a part of an aircraft shows tiny cracks is composed of search time, used for actively scanning, and non-search time, used for matching and decision while fixating a region of interest (Drury et al. 1997 Drury, C. G., Spencer, F. W. and Schurman, D. . Measuring human detection performance in aircraft visual inspection. Proceedings of the 41st annual Human Factors and Ergonomics Society meeting. Albuquerque, NM. pp.304–308. NM. Santa Monica, CA: HFES.  [Google Scholar]). These findings can be applied to detection of threats by X-ray screening of passenger bags at airports. To investigate whether search time and non-search time change when an experienced screener is given additional training in recognising threat objects in passenger bags, data from a European airport were analysed. A comparison of detection performance and reaction time between two large groups of screeners, one trained for 6 months, shows a large impact of training on overall performance and on both search and non-search components of the task. There was also a small but consistent decline in performance measures with screener age. This study shows a way to localise the effect of training on threat detection performance for aviation security screening. Analysis of the time needed for screening each passenger bag showed that training had a significant effect, particularly on the non-search part of the searching process (i.e. identification, recognition, decision, response execution, etc.).     

融合多尺度特征与全局上下文信息的X光违禁物品检测

目的 X光图像违禁物品检测一直是安检领域的一个基础问题,安检违禁物品形式各异,尺度变化大,以及透视性导致大量物体堆放时出现重叠遮挡现象,传统图像处理模型很容易出现漏检误检,召回率低。针对以上问题,提出一种融合多尺度特征与全局上下文信息的特征增强融合网络（feature enhancement fusion network,FEFNet）用于X光违禁物品检测。方法 首先针对特征主干网络darknet53,加入空间坐标的注意力机制,将位置信息嵌入到通道注意力中,分别沿两个空间方向聚合特征,增强特征提取器对违禁目标的特征提取能力,抑制背景噪声干扰。然后,将特征提取主干网络输出的特征编码为1维向量,利用自监督二阶融合获取特征空间像素相关性矩阵,进而获取完整的全局上下文信息,为视觉遮挡区域提供全局信息指导。针对违禁物品尺度不一的问题,提出多尺度特征金字塔融合模块,增加一层小感受野预测特征用于提高对小尺度违禁目标的检测能力。最后,通过融合全局上下文特征信息和局部多尺度细节特征解决违禁物品之间的视觉遮挡问题。结果 在SIXRay-Lite （security inspection X-ray）数据集上进行训练和验证,并与SSD （single shot detection）、Faster R-CNN、RetinaNet、YOLOv5（you only look once）和ACMNet （asymmetrical convolution multi-view neural network）模型进行了对比实验。结果表明,本文模型在SIXray-Lite数据集上的mAP （mean average precision）达到85.64%,特征增强融合模块和多尺度特征金字塔融合模块较原有模型分别提升了6.73%和5.93%,总体检测精度较原有检测网络提升了11.24%。结论 提出的特征增强融合检测模型能够更好地提取显著差异特征,降低背景噪声干扰,提高对多尺度以及小型违禁物品的检测能力。同时利用全局上下文特征信息和多尺度局部特征相结合,有效地缓解了违禁物品之间的视觉遮挡现象,在保证实时性的同时有效地提高了模型的整体检测精度。     

A fuzzy system to support the configuration of baggage screening devices at an airport

An airport is a complex engineering system; it is composed of many elements interconnected with numerous internal relations with a strongly pronounced role of the human factor. One of the specific tasks carried out by the airport managing entity (AME) is to configure the airport security system (ApSS) so that to attain the expected level of confidence in the airport safety and security. This task consists in selection of infrastructure, technical equipment, allocation of personnel and financial means that are necessary to perform all functions of the ApSS. One of the aspects of the configuration of the ApSS is the allocation of available X-ray baggage screening devices searching for items prohibited for transportation. To make this allocation, we need to know how effective these devices are (in terms of detecting prohibited items). This assessment is dependent on several factors which are treated as linguistic variables and are input to fuzzy inference system: the ability to detect explosives, the number of detection lines, the effectiveness of the TIP (Threat Image Projection) system and the age of the machine. Some of these elements are difficult to objective assessment, as they are heavily dependent on the human factor or the information is uncertain or incomplete. So fuzzy ApSS analysis is proposed. The output from the fuzzy inference system is linguistic variable Device evaluation. The meaning of this variable is the ability to protect the aircraft against prohibited items. The proposed new method of assessing the airport baggage screening system involves the construction of a hierarchical fuzzy inference system. The usefulness of the method is exemplified for Katowice–Pyrzowice International Airport, for which an assessment of devices has been performed. The results show that not only allocation of specific devices for specific control points is important for the security of passengers. Also important are the locally accepted principles of their work, which so far are not specified by international regulations. This applies for instance to the selection of the number and frequency of TIP images. Experiments show that the proposed approach can be effective as part of an expert system for supporting the airport operator in configuring ApSS.     

A mathematical framework for passenger screening optimization via a multi-objective evolutionary approach

Concerns regarding the smuggling of dangerous items into commercial flights escalated after the failed Christmas day bomber attack. As a result, the Transportation Security Agency (TSA) has strengthened its efforts to detect passengers carrying hazardous items by installing novel screening technologies and by increasing the number of random pat-downs performed at security checkpoints nationwide. However, the implementation of such measures has raised privacy and health concerns among different groups thus making the design and evaluation of new inspection strategies strongly necessary. This research presents a mathematical framework to design passenger inspection strategies that include the utilization of novel and traditional technologies (i.e. body scanners, explosive detection systems, explosive trace detectors, walk-through metal detectors, and wands) offered by multiple manufacturers, to identify three types of items: metallic, bulk explosives (i.e. plastic, liquids, gels), and traces of explosives. A multiple objective optimization model is proposed to optimize inspection security, inspection cost, and processing time; an evolutionary approach is used to solve the model. The result is a Pareto set of quasi-optimal solutions representing multiple inspection strategies. Each strategy is different in terms of: (1) configuration, (2) the screening technologies included, (3) threshold calibration, and consequently, (4) inspection security, inspection cost, and processing time.     

Why stop after 20 minutes? Breaks and target prevalence in a 60-minute X-ray baggage screening task

Current EU regulation restricts continuously reviewing X-ray images of passenger baggage to 20-min duration as a precautionary measure to prevent performance decrements in airport security officers (screeners). However, this 20-min limit is not based on clear empirical evidence on how well screeners can sustain their performance over time. Our study tested screeners in a 60-min simulated X-ray cabin baggage screening task. One group took 10-min breaks after 20 min of screening; the other group worked without breaks. We found no decrease in performance over 60 min in either group. Breaks did not affect performance, but they did reduce the amount of subjective distress. By varying target prevalence, we found that d_a with a slope of about 0.6 is a more valid measure of detection performance than d'. Target prevalence caused a criterion shift. Our results provide a basis for conducting field studies of prolonged screening durations, and open the discussion on whether more flexible break policies and work schedules should be considered.     

优化YOLOv4算法的安检X光图像检测网络

针对目前安检场景中违禁品种类复杂、人工检查效率低易出错等问题,文章提出一种名为Res152-YOLO的网络架构,该架构基于YOLOv4(You Only Look Once)优化目标检测网络.为提高对X光图像中危险品的检测精度,Res152-YOLO采用ResNet-152网络代替原YOLOv4中的CSPDarknet-53网络,将改进后的ResNet残差网络与YOLOv4网络连接.实验中利用YOLOv4、Res152-YOLO等一系列网络在同一数据集上进行对比实验,分别比较上述网络的损失曲线、对各类危险品的检测性能以及各网络的总体性能.结果表明,Res152-YOLO网络在以上实验中性能优于原YOLOv4网络,并且满足安检设备的帧率要求.改进后的网络有效提高了安检的准确率,能够消除一定的安全隐患.     

基于智能识别技术的铁路安检辅助分析装置研究

针对铁路安检X光图像判图高度依赖人工,时有发生漏检的问题,提出一种基于智能识别技术的铁路安检辅助分析装置。通过采用视频图像接口的硬件设计,解决与安检仪的适配问题。通过采用跟踪进程、分析进程、推送进程的多进程思路设计,解决60Hz刷新率下的高精度分析显示问题。跟踪进程采用关键点差分算法实现安检X光图像跟踪,分析进程采用改进的残差网络(ResNet)实现特征提取,采用CenterNet算法实现禁限物品检测,推送进程通过检测键盘中断实现报警图像推送。经实验测试,该装置禁限物品识别准确率达92%,显示帧率达60帧/秒,可适配主流品牌安检仪,满足铁路车站安检辅助分析需求。     

违禁品X光图像检测技术应用研究进展综述

X光图像在安检中应用十分广泛,目前大部分安检工作还要依靠人工完成,但X光安检巨大的工作量和工作强度使自动安检成为必然趋势。如何根据X光图像自动检测其中物体成为研究热点。随着基于深度学习技术的目标检测取得巨大进展,在X光图像违禁品检测中也大量应用深度学习模型进行研究并获得大量成果。为全面、详细总结现有研究,首先介绍X光成像特点、X光图像检测的传统方法以及基于深度学习的方法,然后对比传统方法与深度学习方法的检测效果并分析现有自动安检研究进展,最后指出未来值得关注的研究方向,以期给X光图像违禁品检测的研究提供参考。     

Computerized working memory training: Can it lead to gains in cognitive skills in students?

Given that working memory is an important cognitive skill that is linked to academic success, there is increasing attention given to exploring ways to support working memory problems in struggling students. One promising approach is computerized training, and the aim of the present study focused on whether computerized working memory training could result in both near and far transfer training effects; and whether such effects would be maintained over time. Students were allocated into one of three groups: Nonactive Control, Active Control, where they trained once a week (WMT-Low frequency); Training group, where they trained four times a week (WMT-High frequency). All three groups were tested on measures of working memory, verbal and nonverbal ability, and academic attainment before training; and re-tested on the same measures after training, as well as 8 months later. The data indicate gains in both verbal and visuo-spatial working memory tasks for the high-frequency Training group. Improvements were also evidenced in tests of verbal and nonverbal ability tests, as well as spelling, in the high-frequency Training group. There were some maintenance effects when students were tested 8 months later. Possible reasons for why the computerized working memory training led to some far transfer effects in the high-frequency Training group are included in the discussion.     

A computer-based outpatient medical record for a teaching hospital.

We developed a computer-based outpatient medical record system to facilitate direct physician interaction with the clinical computing system at the Beth Israel Hospital in Boston. During the 2 years since the medical record system was installed, 20 staff physicians, 5 fellows, 64 residents, and 11 nurse practitioners have entered 15,121 active problems and 1996 inactive problems for 3524 patients, as well as 12,651 active medications and 1894 discontinued medications for 3430 patients. Another 20,321 items were entered on health-promotion and disease-prevention screening sheets, and with the help of automatic updating by the computer, an additional 21,897 entries on screening sheets were made for 8686 patients. Clinicians wrote 10.9 +/- 12.8 (mean +/- SD) words per problem when they were working at the computer, as compared with 4.3 +/- 2.5 words per problem when they were writing in the paper medical record. We conclude that physicians will readily enter data directly into a computing system when they are given appropriate tools, and that they consider the computer-based problem list to be a valuable improvement over its paper counterpart. Use of a computer-based medical record system has obvious benefits for data management and patient care.     

基于小波分析和伪彩色处理的轮毂图像增强

针对轮毂X射线图像随机噪声大、对比度低、分辨率较低、不利于对缺陷进行检测的问题,该文充分利用小波降噪增强技术和伪彩色增强技术的优点,提出基于小波分析和伪彩色处理的轮毂X射线图像增强方法。首先将轮毂X射线图像用改进的小波增强方法进行降噪增强处理,提高图像的对比度及细节可见度,然后进行基于HSI颜色空间的伪彩色处理进一步提高图像分辨率,最后得到增强后的轮毂图像。实验结果表明,该文提出的方法提高了轮毂X射线图像的分辨率,在降噪的同时产生更好的视觉效果,方便进行轮毂缺陷的检测。     

Vector Ordering and Morphological Operations for Colour Image Processing: Fundamentals and Applications

: The extension of concepts of greyscale morphology to colour image processing requires the use of a proper ordering of vectors (colours) and the definitions of infimum and supremum operators in an appropriate colour space. In this paper, a new approach to colour image morphology is proposed. It is based on a new ordering of vectors in the HSV colour space that is partial ordering. The proposed approach is hue preserving, and it is not a component-wise technique. Its basic characteristic is that it is compatible to the standard greyscale morphology: its fundamental and secondary operations possess the same basic properties as their greyscale counterparts, and furthermore, it is identical to greyscale morphology when it is applied to greyscale images. Examples that illustrate the application of the defined operations to colour images are provided. Moreover, the usefulness of the new method in various colour image processing applications, such as colour image edge detection, object recognition, vector top-hat filtering and skeleton extraction, is demonstrated. Received: 14 July 2000, Received in revised form: 24 April 2001, Accepted: 19 June 2001     

A factor analysis based selection process for predicting successful university color guard club members

The purpose of this paper is to determine how to place students in the most appropriate club, to save training time and expenses, and maximize club performance. To this end, this research utilizes the example of the color guard club at Chienkuo Technology University (CTU). First, the authors researched the characteristics needed to be a color guard member. Second, a series of tests for measuring these characteristics was designed. Third, the authors administered the tests to all the club members and recorded the results. Fourth, after the club members had received one year of training, the authors ran regression analysis by using data from the tests of those who successfully passed the training. From this, the authors obtained a regression model. The authors then ran logistic regression analysis and discriminant analysis on the test data of all initial color guard club members, including those who eventually passed the training and those who eventually withdrew, to establish screening norms. Last, using factor analysis, the authors found the latent factors. These factors, along with the screening norms, can serve as a foundation for future selection of color guard members. This process of selecting club members scientifically may be adopted by other clubs in order to match students and clubs most effectively.     

基于稀疏表示与字典学习的彩色图像去噪算法

针对彩色图像在去噪时易产生模糊现象和伪色彩的问题,提出多信息结合字典算法。首先提出了基于RGB颜色空间各通道模值的加权梯度定义,并在此基础上建立了由彩色图像的亮度、加权梯度、颜色信息结合的一种过完备结构字典。其次利用噪声图像的稀疏性,通过不断更新迭代的字典训练过程,找到最优稀疏系数和最优学习字典,从而将噪声信息和图像有用信息分离开,精确重构图像并单求其颜色,进而得到去噪后的彩色图像。实验结果显示,与已有算法相比,本文提出的算法在不同的噪声强度下都取得了更好的视觉效果和更高的客观评价指标值,表明该算法具有良好的去噪性能。     

基于多部位多示例学习的人体检测

基于部位的检测方法能处理多姿态及部分遮挡的人体检测,多示例学习能有效处理图像的多义性,被广泛应用于图像检索与场景理解中。文中提出一种基于多示例学习的多部位人体检测方法。首先,根据人体生理结构将图像分割成若干区域,每个区域包含多个示例,利用AdaBoost多示例学习算法来训练部位检测器。然后利用各部位检测器对训练样本进行测试得到其响应值,从而将训练样本转化为部位响应值组成的特征向量。再用SVM方法对这些向量进行学习,得到最终的部位组合分类器。在INRIA数据集上的实验结果表明该方法能改进单示例学习的检测性能,同时评价3种不同的部位划分及其对检测性能的影响。