WO2015067208A1 - 检查方法和设备 - Google Patents

检查方法和设备 Download PDF

Info

Publication number
WO2015067208A1
WO2015067208A1 PCT/CN2014/090563 CN2014090563W WO2015067208A1 WO 2015067208 A1 WO2015067208 A1 WO 2015067208A1 CN 2014090563 W CN2014090563 W CN 2014090563W WO 2015067208 A1 WO2015067208 A1 WO 2015067208A1
Authority
WO
WIPO (PCT)
Prior art keywords
cigarette
image
interest
region
model
Prior art date
Application number
PCT/CN2014/090563
Other languages
English (en)
French (fr)
Inventor
张丽
陈志强
李强
张健
顾建平
崔锦
Original Assignee
同方威视技术股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 同方威视技术股份有限公司 filed Critical 同方威视技术股份有限公司
Priority to JP2016528054A priority Critical patent/JP6246355B2/ja
Priority to EP14860336.8A priority patent/EP3067823A4/en
Priority to KR1020187008606A priority patent/KR20180035930A/ko
Priority to US15/034,021 priority patent/US10013615B2/en
Priority to KR1020167015104A priority patent/KR20160083099A/ko
Publication of WO2015067208A1 publication Critical patent/WO2015067208A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/05—Recognition of patterns representing particular kinds of hidden objects, e.g. weapons, explosives, drugs

Definitions

  • Embodiments of the present invention relate to automatic detection of suspects in radiation images, and more particularly to methods of inspecting cigarettes and corresponding security inspection devices in large container scanning systems.
  • Cigarette smuggling is considered to be the second largest smuggling activity in the world after drugs. For more than a decade, although countries have actively participated in combating cigarette smuggling, the number of smuggled cigarettes is still growing. In 1996, smuggled cigarettes accounted for 6.5% of total cigarette sales. In 2009, this figure increased to 11.6%, reaching 675 billion, causing $400 billion in losses to governments. In addition, the harm of cigarette smuggling is not only reflected in threatening human health, causing government tax losses, but also becoming a source of funds for criminal organizations and terrorist organizations, and even providing funding for drug smuggling. The EU even declared: "All international criminal organizations are suspected of involvement in cigarette smuggling." The danger of cigarette smuggling is increasing, reflecting that the relevant testing methods are not in place, giving criminals a chance.
  • Object detection is a hot issue in the field of computer vision and pattern recognition.
  • image retrieval there are many research results that can be used for reference.
  • the Oriented Gradients feature, the DPM-Deformable Part-based Model algorithm, and the Deep Learning method have led to a significant increase in object detection.
  • the invention draws on relevant research and conducts targeted research on the cigarette model in the radiation image, and obtains better results.
  • the embodiment of the present invention draws on the research of automatic object detection based on the container DR image, and proposes a new automatic cigarette smuggling detection method.
  • the purpose is to automatically detect whether there is a cigarette in the image by using an algorithm by means of radiographic scanning in the process of container inspection. If a cigarette is present, the position of the cigarette in the image is given, thereby assisting in manually determining whether there is a smuggling case.
  • the performance of the algorithm is also a task that must be considered for the purpose of assisting labor.
  • the detection algorithm must have a low false positive rate and a false negative rate, and must meet the requirements of real-time detection. Embodiments of the present invention have achieved good results by targeted research on cigarette patterns.
  • the algorithm false positive rate is less than 0.5%, the false negative rate is less than 10%, and the calculation is completed within 1 second, which satisfies the above application requirements.
  • the algorithm may also preferably have a self-learning function.
  • the algorithm can self-learn images that are confirmed as cigarettes but cannot be detected, and meet the detection requirements of changes in cigarette patterns when cigarette packages are changed.
  • the confirmation mechanism includes manually inputting a cigarette image and automatically acquiring a cigarette image through a customs declaration or the like.
  • a method of automatically detecting a cigarette in a container security perspective image comprising the steps of: obtaining a perspective image of the object to be inspected; processing the perspective image to obtain a region of interest; The region of interest is automatically detected using a cigarette model to determine if the region of interest of the fluoroscopic image belongs to a cigarette.
  • the image resolution does not need to be too high in order to achieve detection under real-time conditions.
  • a large number of cigarettes have a striped texture in the image.
  • This characteristic reflects the packaging of the cigarette rather than the cigarette itself.
  • the root cause of this feature is that in order to reduce the harm of smoking and prevent personal packaging from attracting young people, cigarette packages in various countries tend to be homogenized, which provides conditions for effective testing.
  • the present invention utilizes the above features and accurately and quickly identifies cigarettes in the container security perspective image through steps of image acquisition, region of interest acquisition, cigarette model establishment, and automatic detection.
  • the method before the step of automatically detecting the region of interest using the cigarette model, the method further includes the step of establishing a cigarette model, the step of establishing the cigarette model comprising: establishing a library of cigarette images; The middle images are processed to obtain respective regions of interest;
  • Feature extraction is performed on the region of interest of the image in the cigarette image library; the cigarette model is generated based on the extracted feature training classifier.
  • the step of establishing a library of cigarette images comprises: scanning a plurality of cigarette images in different shapes in the container, obtaining a positive sample library of the cigarette model; and collecting objects similar to cigarettes and random objects The image forms a negative sample library of cigarette models, each of the negative sample libraries containing no cigarettes.
  • the scanned images of the various regular stacks of cigarettes are divided into a plurality of modes of different widths from which a positive sample bank of the cigarette model is generated.
  • the step of performing feature extraction on the region of interest of the image in the library of cigarette images comprises: manually marking the position of the cigarette in units of boxes, forming a positive sample feature set in the plurality of modes; Multiple samples are randomly selected from the sample library to perform feature extraction to form a negative sample feature set.
  • the step of training the classifier based on the extracted features comprises: (1) training the classifier under given conditions of the positive and negative sample feature sets; (2) classifying the positive and negative samples with the classifier; 3) According to the confidence of the result, remove the easily-divided negative samples and re-add enough random negative samples; (4) repeat the above steps (1), (2), and (3) until the classification results of the classifier change enough. Small, or until enough iterations are reached.
  • the method further comprises, for the image of the undetected cigarette, automatically analyzing the image by manual labeling or customs declaration to obtain a new cigarette image and update the cigarette model
  • the cigarette image may exhibit a large difference with respect to the established model.
  • the method further includes a self-learning function. By continuously updating the cigarette model, the method is able to adapt to different detection environments, scanning devices and more packaging forms of cigarettes.
  • the image size is scaled to a uniform resolution, and the grayscale stretching is performed to fill the image value range with the entire possible range of values. Perform normalization.
  • the air portion of the fluoroscopic image is excluded to prevent air noise from producing a detection result.
  • the fluoroscopic image in the step of processing the fluoroscopic image to obtain a region of interest, using the method of air luminance threshold, the fluoroscopic image is binarized with an air value, and only the image portion below the threshold is performed. Cigarette testing.
  • the step of automatically detecting the region of interest using the cigarette model comprises: normalizing and acquiring the region of interest, generating an HOG feature of the region of interest, given a image to be detected; using a sliding window Traversing, seeking its maximum confidence in each window in multiple modes; and the location of the cigarette at a confidence greater than a certain threshold.
  • the cigarette is detected on a plurality of scales.
  • An embodiment of the present invention further provides an inspection apparatus including: a digital radiography apparatus that performs radiographic inspection on an object to be inspected to acquire a perspective image of an object to be inspected; and an image processing apparatus that processes the perspective image to obtain a feeling The region of interest, and the region of interest is automatically detected using a cigarette model to determine if the region of interest of the fluoroscopic image belongs to a cigarette.
  • an inspection apparatus including: a digital radiography apparatus that performs radiographic inspection on an object to be inspected to acquire a perspective image of an object to be inspected; and an image processing apparatus that processes the perspective image to obtain a feeling The region of interest, and the region of interest is automatically detected using a cigarette model to determine if the region of interest of the fluoroscopic image belongs to a cigarette.
  • the device is for automatically detecting cigarette smuggling.
  • the embodiment of the invention performs cigarette detection on the scanned image of the cargo, especially the container, which can avoid the problem that the traditional method of detecting the loophole and the manual judgment effect is poor, and is important for combating cigarettes.
  • the invention firstly proposes a technical solution for automatically judging cigarette smuggling by a detection algorithm in the prior art, and has been verified by practice, has excellent performance, and has strong practicability.
  • FIG. 1 is a schematic structural view of an inspection apparatus according to an embodiment of the present invention
  • FIG. 2 is a block diagram showing the structure of an inspection apparatus according to an embodiment of the present invention.
  • FIG. 3 shows a flow chart of a method of automatically detecting cigarettes in accordance with an embodiment of the present invention
  • FIG. 4 shows a flow chart of establishing a cigarette model in accordance with an embodiment of the present invention
  • Figure 5 shows six placements of cigarettes
  • Figure 6 shows a schematic view of scanning cigarettes in a container
  • Figure 7 shows three modes of cigarette images
  • Figure 8 shows a flow chart of automatic detection in accordance with an embodiment of the present invention
  • Figure 9 is a diagram showing the results of cigarette detection according to an embodiment of the present invention.
  • Figure 10 shows a flow chart of a self-learning process in accordance with an embodiment of the present invention.
  • an inspection method is proposed for the problem that the smuggling of cigarettes cannot be automatically checked in the prior art.
  • a perspective image of the object to be inspected is first acquired.
  • the fluoroscopic image is then processed to obtain a region of interest.
  • the region of interest is automatically detected using a cigarette model to determine if the region of interest of the fluoroscopic image belongs to a cigarette.
  • FIG. 1 is a schematic view of an inspection apparatus in accordance with one embodiment of the present invention.
  • the radiation source 110 generates X-rays, which are collimated by the collimator 120 to perform a security check on the moving container truck 140, and the detector 150 receives the rays that penetrate the truck to obtain a transmission image.
  • the transmission image is also processed by the image processing device 160 such as a computer to determine whether or not there is a cigarette. According to some embodiments, in the case of judging a cigarette, marking it on the image, or reminding the panelist that the container The truck carries cigarettes.
  • FIG. 2 shows a deployment diagram of an inspection apparatus according to an embodiment of the present invention.
  • the radiation source and detector 250 of the accelerator 210 are hardware devices
  • the data acquisition and control module 251 is coupled to the detector 250 and controls the detector 250
  • the control device 211 is coupled to the accelerator 210 and controls the accelerator 210. Bunch and stop.
  • the image inspection station 280 and the workstation running the inspection station 270 for the scanning device are connected to the data acquisition and control module 251 via the switch 260 and communicated, and the image results can be printed by the printer 290.
  • the scanner 271 is connected to the operation check station 270 for inputting other information such as customs declaration data. In other embodiments, other input devices such as a keyboard may be used to input information.
  • the accelerator 210 produces X-rays that pass through the X-rays of the object under inspection 240 and are received by the detector 250 for transmission imaging of the object under inspection.
  • the transmission image is processed by an image processing device such as a computer (for example, image inspection station 280) to determine whether or not a cigarette is present.
  • an image processing device such as a computer (for example, image inspection station 280) to determine whether or not a cigarette is present.
  • the image is marked or reminded to the panelist that the container truck carries the cigarette.
  • step S301 a perspective image of the object to be inspected is acquired.
  • step S302 the fluoroscopic image is processed to obtain a region of interest.
  • step S303 the region of interest is detected using the cigarette model to determine whether the region of interest of the fluoroscopic image belongs to a cigarette.
  • the step of automatically detecting includes image feature extraction, decision, and identification of suspect regions, and the like.
  • Different scanning devices have different images depending on the energy/dose of the radiation source, so the images obtained are not the same.
  • the image can first be normalized to reduce this difference.
  • the container cabinet is large, so that the detection can be realized under real-time conditions, and the image resolution does not need to be too high.
  • a 5mm/pixel image as an example, a large number of cigarettes have a striped texture in the image.
  • This characteristic reflects the packaging of the cigarette rather than the cigarette itself.
  • the root cause of this feature is that in order to reduce the harm of smoking and prevent personal packaging from attracting young people, cigarette packages in various countries tend to be homogenized, which provides conditions for effective testing. Therefore, without loss of generality, the image size can be scaled to a uniform resolution of, for example, about 5 mm/pixel, and grayscale stretching is performed to fill the image value range with the entire range of possible values, and the normalization operation is completed.
  • the image pixel resolution exemplified above is 5 mm/pixel, however, those skilled in the art will understand that the above resolution is not the only option, and the above resolution can be appropriately modified for the actual size of the cigarette package.
  • the air portion of the image is excluded prior to detection to avoid air noise producing a detection result.
  • the image is binarized with an air value using a method of air brightness threshold.
  • the brightness threshold of the air may be preset, and the area where the brightness exceeds the brightness threshold is considered to be the air area. Therefore, cigarette detection is performed only on the image portion below the threshold.
  • Object detection mainly includes two types of methods based on sliding window and sub-region based (Sub Region).
  • the former adopts traversal mode, and slides on various scales of the image in a fixed-size window to calculate whether each pixel is a target object on each scale; the latter adopts image segmentation and other methods to perform local features of non-fixed shapes in the image. Extraction and target detection.
  • the sliding window method is much larger than the local area method, the accuracy is much higher. Therefore, in an embodiment of the invention, it is preferred to use a sliding window method.
  • the step of establishing a cigarette model includes: step S401, establishing a cigarette image library acquisition; step S402, obtaining a region of interest, step S403, feature extraction and step S404, and training the classifier to generate a cigarette model.
  • the cigarette model models the image pattern of cigarettes, so the extraction of image patterns is the focus of the algorithm.
  • the packaging form, placement method, and quantity of cigarettes may cause differences in cigarette images.
  • the convergence of the packaging form makes this problem simple, a small number of modes can generalize most of the situation; the difference in quantity will cause the image gray level and the stripe intensity to change, which needs to be overcome in the feature extraction algorithm; The impact is very large, and the inventors attribute it to three modes.
  • the form of cigarette placement may be any of the six placements in Figure 5. Since the container scanning basically uses a fan beam (as shown in FIG.
  • the lateral edges are three-dimensionally superimposed in the ray direction, so that the cigarettes exhibit vertical stripes of different widths.
  • the vertical stripe texture of cigarettes can be summarized into the following three modes as shown in Fig. 7, namely: Fig. 5 (a) and (d) produce the wide stripe in Fig. 7 (a), Fig. 5 (b) and (e) produce a slightly narrower stripe in Fig. 7(b), while Figs. 5(c) and (f) produce a narrow stripe in Fig. 7(c).
  • the inventors proposed three cigarette patterns in different placements and used algorithms to build models.
  • the present invention subtly utilizes the characteristics of the above three modes, so that after the model is built, it can automatically train and/or learn without relying on the initial manual labeling, thereby greatly saving labor costs and improving detection efficiency and accuracy.
  • Those skilled in the art will appreciate that in other embodiments, Use more placement to create other patterns.
  • the model includes feature extraction algorithms and classifiers. According to the above analysis, the model is established as follows:
  • the classifier can be optimized using existing algorithms in the field of machine learning and pattern recognition. For example, referring to the Boosting idea, the easy-sort samples are continuously discarded, new random samples are added, and the C and D steps are repeated until the error rate is low enough or the algorithm converges to obtain the classifier.
  • the existing algorithm can be used to adjust the position of the positive sample to reduce the impact of manual labeling errors. For example, referring to the idea of the hidden variable model in DPM, it is assumed that the position of the manual label is not completely accurate. After the classifier is obtained in D, the position of the positive sample in A is offset, and the position of the positive sample in the sense of the classifier is obtained. The E step gets an optimized classifier.
  • the creation of a cigarette model can be performed independently of the automatic detection of the cigarette, i.e., the automated cigarette detection of the fluoroscopic image can be performed using the established cigarette model after the cigarette model is established.
  • Figure 6 shows the process of inspecting a cigarette, the cigarette moving relative to the source, producing a transmission image, and then performing an automatic detection.
  • the automatic detection process is actually a subset of the model establishment process: as shown in Fig. 8, in step S801, the image to be inspected is input. At step S802, feature extraction is performed on the image to be inspected. According to some embodiments, the image may also be processed during this process to obtain a region of interest.
  • the pixel model is used to traverse the pixel points in each region of interest for the image to be tested, and the neighborhood image around the point is extracted and subjected to feature extraction and classification judgment to obtain whether the point is a cigarette region, and Confidence in conclusion.
  • step S804 it is determined whether or not there is a cigarette. If so, in step S805, after obtaining the result of whether each pixel is a suspected area of the cigarette, the cigarette area is obtained by extracting through the connected area. Otherwise, in step S806, "undetected" is output or the image to be detected is discarded.
  • Fig. 9 is a view showing the detection of a cigarette in the form of a rectangular frame.
  • the algorithm has a self-learning function that enables it to adapt to different detection environments, scanning devices and more packaging forms of cigarettes.
  • the algorithm used for self-learning is basically the same as the model. The difference is that it only gets positive samples during the detection process (as shown in Figure 10).
  • FIG. 10 shows a flow chart of a self-learning process in accordance with an embodiment of the present invention.
  • step S1001 an image to be inspected is input.
  • step S1002 feature extraction is performed on the image to be inspected.
  • the image may also be processed during this process to obtain a region of interest.
  • step S1003 the inspection is performed using the cigarette model.
  • step S1004 it is determined whether or not there is a cigarette. If so, in step S1005, after obtaining the result of whether each pixel is a suspected area of the cigarette, the cigarette area is obtained by extracting through the connected area.
  • step S1006 if the manual discovery algorithm does not detect the cigarette and is marked in the detection process, the algorithm adds the marked area to the positive sample, and performs training or incremental learning in step S1007 to obtain the updated model. However, if the resulting model still cannot detect the sample that was just added, the update is discarded.
  • HOG HOG
  • DPM Deep Learning
  • Deep Learning can be directly applied to the present invention
  • the detection function can be realized by extracting and classifying the positive and negative sample features.
  • various gradient-based descriptors, texture descriptors, and the like such as HOG, local binary pattern (LBP-Local Binary Pattern), and maximum response set (Maximum Response Sets) may be used.
  • MR8 in addition to the BOW-Bag of Words, texture representation (Texton), sparse representation (Sparse Representation), etc. for structured feature processing; in the classifier, you can use a variety of linear, nonlinear, Integrated classifiers, neural networks, such as Fisher classifiers, support vector machines, Boosting, random forests, multi-layer perceptrons, etc.
  • the algorithm of the present invention is generally applicable to all types of large cargo/container scanning equipment.
  • the image needs to be normalized when the image is acquired.
  • the original two-dimensional image signal be X, according to the physical parameters of the scanning device, zoom the X resolution to 5mm/pixel, and perform grayscale stretching to obtain a normalized image.
  • a pixel larger than the threshold t a is considered to be air and no detection calculation is performed.
  • the number of images is about 100, which can meet the practical requirements.
  • a similar object and a random object are collected to form a negative sample library ⁇ Y ⁇ , and each image in the negative sample library ⁇ Y ⁇ does not contain a cigarette.
  • the embodiment employs the HOG feature.
  • Classifier C is used to classify positive and negative samples. According to the confidence of the results, the negative samples that are easy to divide are removed, and enough random negative samples are added again. Then, repeat steps D) and E) until the classifier changes are small enough, or enough iterations are reached.
  • step B) brings a large error, making the relative position of the cigarette stripes unclear.
  • a classifier C for surrounding the image I ij B) is detected, to find the best position I ij will update I ij I ij ', and then iteratively C) to F) a step Until the classifier changes small enough, or enough iterations are reached.
  • the HOG feature H of the ROI region is generated by normalization and ROI extraction. Use the sliding window to traverse H and find its maximum confidence in each of the three modes. The location of the cigarette is the confidence level above a certain threshold.
  • cigarette detection can be performed on multiple scales (ie, different scales). For example, the detection of the fluoroscopic image on the three scales of [0.9, 1.0, 1.1] can better solve the problem of a certain deflection when the cigarette is placed.
  • the detection result can be post-processed to remove noise. For example, median filtering is performed on the above confidence map, or the above-mentioned confidence map is binarized and connected area area filtering is performed.
  • the self-learning process involves a model update strategy.
  • the algorithm used for self-learning is basically the same as the model. The difference is that it only gets positive samples during the detection process (as shown in Figure 10).
  • the algorithm adds the marked area to the positive sample, re-trains or incrementally learns, and obtains the updated model. However, if the resulting model still cannot detect the sample that was just added, the update is discarded.
  • aspects of the embodiments disclosed herein may be implemented in an integrated circuit as a whole or in part, as one or more of one or more computers running on one or more computers.
  • a computer program eg, implemented as one or more programs running on one or more computer systems
  • implemented as one or more programs running on one or more processors eg, implemented as one or One or more programs running on a plurality of microprocessors, implemented as firmware, or substantially in any combination of the above, and those skilled in the art, in accordance with the present disclosure, will be provided with design circuitry and/or write software and / or firmware code capabilities.
  • signal bearing media include, but are not limited to, recordable media such as floppy disks, hard drives, compact disks (CDs), digital versatile disks (DVDs), digital tapes, computer memories, and the like; and transmission-type media such as digital and / or analog communication media (eg, fiber optic cable, waveguide, wired communication link, wireless communication link, etc.).

Abstract

本发明公开了一种检查方法和设备。所述方法包括以下步骤：获取被检查物体的透视图像；对所述透视图像进行处理得到感兴趣区域；以及利用香烟模型对感兴趣区域进行自动检测，以确定所述透视图像的感兴趣区域是否属于香烟。本发明对货物特别是集装箱扫描图像进行香烟检测，可以避免传统方式的检测漏洞与人工判图效果较差的问题，对于打击香烟走私有重要意义。

Description

检查方法和设备 技术领域
本发明的实施例涉及辐射图像中的嫌疑物自动检测，具体而言涉及大型集装箱扫描系统中，检查香烟的方法以及相应的安全检查设备。
背景技术
打击香烟走私有着重要的意义。香烟走私被认为是世界上仅次于毒品的第二大走私活动。十几年来，虽然各国都积极参与打击香烟走私，但走私香烟的数量还是呈增长趋势。1996年，走私的香烟占香烟总销量的6.5％。而2009年，这一数字增长到了11.6％，达到6750亿支，给各国政府造成4千亿美元的损失。另外，香烟走私的危害不仅体现在威胁人体健康、造成政府税收损失，而且还成为犯罪组织、恐怖组织的资金来源，甚至给毒品走私提供经费。欧盟甚至宣称：“所有的国际犯罪组织都涉嫌参与香烟走私”。香烟走私危害日益加剧，反映了相关检测手段不到位，给犯罪分子可乘之机。
目前国际上缺乏检测香烟走私的有效手段。虽然各种类型的条码、水印已经在香烟上大量使用，但这种手段难于在香烟过境时得到有效查验，因此收效甚微。据研究，香烟走私主要是通过集装箱大量运送。辐射成像通过对货物、行李等透视成像，达到无侵入性检查的目的。目前已经在机场、海关、车站、大型集会等场所广泛应用，是违禁品安检领域最为重要的手段。在集装箱检查的过程中，虽然已经得到了集装箱货物图像，但由于货物种类千差万别，判图员的经验参差不齐，走私品出现概率又比较低，使得人工判断效果差强人意。
近年来，随着模式识别、图像处理等相关学科的快速发展，违禁品的自动检测成为业界和学界关注焦点。但目前来看，大型集装箱数字射线成像(DR-Digital Radiography)图像中的自动检测相关文献仍然很少。限于成像手段、应用领域、数据来源等多方面因素影响，研究多见于小型行李安检设备，比如双能DR、计算机断层扫描(CT-Computed Tomography)中。其中特别是爆炸物与枪支的自动检测，由于直接关系到航空安全，受到了更多研究者的关注。对于走私香烟这个特定领域，目前还没有针对性的公开文献。
物体检测(Object Detection)是当前计算机视觉、模式识别领域的热点问题，随着图像检索的发展出现许多可借鉴的研究成果。特别是方向梯度直方图(HOG-Histogram  of Oriented Gradients)特征、基于可变部件模型(DPM-Deformable Part-based Model)算法、深度学习(Deep Learning)方法的出现，使得物体检测效果大幅提升。本发明借鉴相关研究，对辐射图像中的香烟模型展开针对性研究，并得到较好效果。
发明内容
针对上述问题，基于市场需求，本发明的实施例在使用集装箱DR图像的基础上，借鉴自动物体检测相关研究，提出一种新的香烟走私自动检测方法。其目的在于：在集装箱查私过程中，使用射线扫描成像手段，通过算法自动检测图像中是否有香烟。如果存在香烟，则给出香烟在图像中的位置，以此辅助人工判断是否存在走私案情。
为达到辅助人工的目的，算法的性能也是必须考虑的任务。检测算法必须有较低的误报率和漏报率，另外必须满足实时检测的要求。本发明的实施例通过对香烟模式的针对性研究，得到了较好的效果。算法误报率小于0.5％，漏报率小于10％，且在1秒内完成计算，满足上述应用需求。
另外，为使得算法能够满足特定场合的应用，算法还可以优选地具备自学习功能。算法对确认为香烟但又检测不出的图像可以自学习，满足香烟包装变化时香烟模式变化的检测需求。其中，确认机制包括人工输入香烟图像和通过报关单等自动获取香烟图像等。
根据本发明的实施例，提供了一种集装箱安检透视图像中自动检测香烟的方法，所述方法包括以下步骤：获取被检查物体的透视图像；对所述透视图像进行处理得到感兴趣区域；以及利用香烟模型对感兴趣区域进行自动检测，以确定所述透视图像的感兴趣区域是否属于香烟。
由于集装箱箱体较大，为了在实时条件下实现检测，图像分辨率无需太高。以5mm/像素的图像为例，大量香烟在图像中呈条纹状纹理，这个特性反映的其实是香烟的包装而不是香烟本身。产生这个特点的根源在于：为降低吸烟的危害，防止个性化包装对青少年产生吸引力，世界各国的香烟包装趋于同一化，这恰恰为有效的检测提供了条件。本发明正是利用了上述特点，并通过图像获取、感兴趣区域获取、香烟模型建立以及自动检测等步骤，以准确、迅速地识别集装箱安检透视图像中的香烟。
根据一些实施例，在所述利用香烟模型对感兴趣区域进行自动检测的步骤之前，还包括建立香烟模型的步骤，所述建立香烟模型的步骤包括：建立香烟图像库；对所述香烟图像库中图像进行处理得到各自的感兴趣区域；
对所述香烟图像库中图像的感兴趣区域进行特征提取；基于提取的特征训练分类器，生成所述香烟模型。
根据一些实施例，所述建立香烟图像库的步骤包括：扫描在集装箱中各个摆放形式、不同数量下的香烟图像，得到香烟模型的正样本库；以及采集与香烟类似的物体和随机物体的图像，形成香烟模型的负样本库，所述负样本库中每幅图像均不包含香烟。
根据一些实施例，在所述建立香烟模型的步骤中，香烟的各种规则堆叠的扫描图像被划分为宽度不同的多种模式，由所述多种模式生成所述香烟模型的正样本库。
根据一些实施例，对所述香烟图像库中图像的感兴趣区域进行特征提取的步骤包括：以箱为单位，人工标注香烟位置，形成所述多种模式下的正样本特征集；以及在负样本库中随机抽取多个样本，进行特征提取，形成负样本特征集。
根据一些实施例，基于提取的特征训练分类器的步骤包括：(1)在给定正、负样本特征集的条件下，训练分类器；(2)用分类器对正、负样本分类；(3)根据结果置信度，将易分的负样本去掉，重新加入足够的随机负样本；(4)重复上述步骤(1)、(2)、和(3)，直到分类器的分类结果变化足够小，或达到足够迭代次数为止。
根据一些实施例，所述方法还包括对于未检测到香烟的图像，通过人工标注或报关单自动分析，得到新的香烟图像并更新香烟模型
在实际检测中，在不同的检测环境(例如但不限于：扫描方式、集装箱状况等)、扫描设备和摆放模式下，香烟图像可能相对于已建立的模型而呈现出较大的差异。为使本发明的方法适用于更加具体的检测环境、扫描设备和摆放模式，根据本发明的实施例，所述方法还包括自学习功能。通过不断更新香烟模型，使得所述方法能够适应不同的检测环境、扫描设备和更多包装形式的香烟。
根据一些实施例，在所述获取被检查物体的透视图像的步骤中，将图像尺寸缩放到均一的分辨率下，并进行灰度拉伸，使图像值域充满整个可能的取值范围，来进行归一化操作。
根据一些实施例，在所述对所述透视图像进行处理得到感兴趣区域的步骤中，把所述透视图像中空气部分排除掉，以避免空气噪声产生检测结果。
根据一些实施例，在对所述透视图像进行处理得到感兴趣区域的步骤中，采用空气亮度阈值的方法，用空气值对所述透视图像进行二值化，只对阈值之下的图像部分进行香烟检测。
根据一些实施例，在所述利用香烟模型对感兴趣区域进行自动检测的步骤包括：给定要检测的图像，经过归一化和获取感兴趣区域，生成感兴趣区域的HOG特征；用滑动窗口遍历，求它在多种模式下，每个窗口中的最大置信度；以及置信度大于特定阈值处即为香烟位置。
根据一些实施例，在多个尺度上检测香烟。
本发明的实施例还提供了一种检查设备，包括：数字放射摄影装置，对被检查物体进行射线检查，获取被检查物体的透视图像；以及图像处理装置，对所述透视图像进行处理得到感兴趣区域，并利用香烟模型对感兴趣区域进行自动检测，以确定所述透视图像的感兴趣区域是否属于香烟。
根据一些实施例，所述设备用于自动检测香烟走私。
本发明的实施例对货物特别是集装箱扫描图像进行香烟检测，可以避免传统方式的检测漏洞与人工判图效果较差的问题，对于打击香烟走私有重要意义。本发明在本领域中首次提出通过检测算法自动判断香烟走私的技术方案，且已经过实际验证，具有优良的性能，具有很强的实用性。
附图说明
通过参考以下描述的实施例，本发明的这些和其他方面将是清楚的并得到阐述。在附图中：
图1示出了根据本发明实施例的检查设备的结构示意图；
图2示出了根据本发明实施例的检查设备的结构框图；
图3示出了根据本发明的实施例的自动检测香烟方法的流程图；
图4示出了根据本发明的实施例的建立香烟模型的流程图；
图5示出了香烟的六种摆放形式；
图6示出了对集装箱中的香烟进行扫描的示意图；
图7示出了香烟图像的三种模式；
图8示出了根据本发明的实施例的自动检测的流程图；
图9示出了根据本发明的实施例的香烟检测结果的示意图；以及
图10示出了根据本发明的实施例的自学习过程的流程图。
具体实施方式
下面将详细描述本发明的具体实施例，应当注意，这里描述的实施例只用于举例说明，并不用于限制本发明。在以下描述中，为了提供对本发明的透彻理解，阐述了大量特定细节。然而，对于本领域普通技术人员显而易见的是：不必采用这些特定细节来实行本发明。在其他实例中，为了避免混淆本发明，未具体描述公知的电路、材料或方法。
在整个说明书中，对“一个实施例”、“实施例”、“一个示例”或“示例”的提及意味着：结合该实施例或示例描述的特定特征、结构或特性被包含在本发明至少一个实施例中。因此，在整个说明书的各个地方出现的短语“在一个实施例中”、“在实施例中”、“一个示例”或“示例”不一定都指同一实施例或示例。此外，可以以任何适当的组合和/或子组合将特定的特征、结构或特性组合在一个或多个实施例或示例中。此外，本领域普通技术人员应当理解，在此提供的附图都是为了说明的目的，并且附图不一定是按比例绘制的。应当理解，当称元件“耦接到”或“连接到”另一元件时，它可以是直接耦接或耦接到另一元件或者可以存在中间元件。相反，当称元件“直接耦接到”或“直接连接到”另一元件时，不存在中间元件。相同的附图标记指示相同的元件。这里使用的术语“和/或”包括一个或多个相关列出的项目的任何和所有组合。
下面结合附图对本发明的对移动目标进行成像检查的设备进行说明。如图所示，以集装箱检查设备作为一个实施例进行说明。下述说明只是为了结合实例对本发明进行说明，并不是为了将本发明限制于下述内容。
根据本发明的一些实施例，针对现有技术中无法自动检查香烟走私的问题，提出了一种检查方法。根据该方法，首先获取被检查物体的透视图像。然后对所述透视图像进行处理得到感兴趣区域。接下来，利用香烟模型对感兴趣区域进行自动检测，以确定所述透视图像的感兴趣区域是否属于香烟。这样，当对诸如集装箱车辆这样的移动目标(即被检查武器)进行检查时，能够自动检查车辆上是否有香烟走私，并且能够提醒判图员，或者将其准确地定位在图像上。
图1是根据本发明一个实施方式的检查设备的示意。如图1所示，射线源110产生X射线，经过准直器120准直后，对移动的集装箱卡车140进行安全检查，由探测器150接收穿透卡车的射线，得到透射图像。在进行上述检查的同时，还利用诸如计算机之类的图像处理装置160对透射图像进行处理，判断是否存在香烟。根据一些实施例，在判断出香烟的情况下，在图像上标注出来，或者向判图员提醒，该集装箱 卡车上携带了香烟。
图2示出了本发明实施例的检查设备的部署图。如图2所示，例如加速器210的射线源和探测器250为硬件设备，数据采集与控制模块251与探测器250进行连接并控制探测器250，控制装置211与加速器210连接并控制加速器210出束和停束。
图像检查站280和运行检查站270为扫描设备配套的工作站，通过交换机260与数据采集和控制模块251连接并进行通信，图像结果可以通过打印机290打印出来。扫描仪271与运行检查站270连接，用于输入报关单数据等其他信息。在其他实施例中，也可以用诸如键盘之类其他的输入装置来输入信息。
借助于上述的设备，在一些实施例中，加速器210产生X射线，穿透被检查物体240的X射线，被探测器250接收以对被检查物体进行透射成像。利用诸如计算机之类的图像处理装置(例如图像检查站280)对透射图像进行处理，判断是否存在香烟。根据一些实施例，在判断出香烟的情况下，在图像上标注出来，或者向判图员提醒，该集装箱卡车上携带了香烟。这样，当对诸如车辆这样的移动目标进行检查时，在存在香烟的情况下，将其准确地定位在该目标的图像上。
图3示出了根据本发明的实施例的一种集装箱安检透视图像中自动检测香烟的方法。如图3所示，在步骤S301，获取被检查物体的透视图像。在步骤S302，对透视图像进行处理得到感兴趣区域。然后，在步骤S303，利用香烟模型对感兴趣区域进行检测，以确定所述透视图像的感兴趣区域是否属于香烟。根据一些实施例，所述自动检测的步骤包括待测图像特征提取、判决、以及对嫌疑区域标记等。
以下将详细地描述上述方法中各个步骤的实现方式。本领域技术人员能够理解，以下的实现方式仅是示例性的，而非限制性的。
图像获取
不同的扫描设备由于射线源的能量/剂量不同，探测器尺寸不同，所以得到的图像不尽相同。优选地，首先可以对图像归一化处理以减少这个差异。
集装箱箱体较大，为在实时条件下实现检测，图像分辨率无需太高。以5mm/像素的图像为例，大量香烟在图像中呈条纹状纹理，这个特性反映的其实是香烟的包装而不是香烟本身。产生这个特点的根源在于：为降低吸烟的危害，防止个性化包装对青少年产生吸引力，世界各国的香烟包装趋于同一化，这恰恰为有效的检测提供了条件。因此，不失一般性，可以将图像尺寸缩放到例如约5mm/像素的均一的分辨率下，并进行灰度拉伸，使图像值域充满整个可能的取值范围，完成归一化操作。
以上例举的图像像素分辨率为5mm/像素，然而本领域技术人员能够理解，上述分辨率并不是唯一的选择，针对香烟包装的实际尺寸，可以适当地修改上述分辨率。
感兴趣区域提取
在检测之前，优选地，把图像中空气部分排除掉，以避免空气噪声产生检测结果。可选地，采用空气亮度阈值的方法，用空气值对图像进行二值化。例如，可以预设空气的亮度阈值，认为亮度超过所述亮度阈值的区域为空气区域。因此，只对阈值之下的图像部分进行香烟检测。
模型建立
物体检测主要包括基于滑动窗口(Sliding Window)和基于局部区域(Sub Region)的两大类方法。前者采用遍历方式，以固定大小窗口在图像的各个尺度上滑动，计算每个像素在每个尺度上是否是目标物体；后者采用图像分割等方法，对图像中非固定形状的局部区域进行特征提取与目标检测。
根据发明人的对比研究，滑动窗口法虽然计算量远大于局部区域法，但精度却要高很多。因此在本发明的实施例中，优选地以使用滑动窗口法。
如图4所示，所述建立香烟模型的步骤包括：步骤S401，建立香烟图像库获取；步骤S402，得到感兴趣区域，步骤S403，特征提取和步骤S404，训练分类器生成香烟模型。
香烟模型是把香烟的图像模式模型化，因此图像模式的提取是算法的重点内容。不失一般性，香烟的包装形式、摆放方法、数量不同会造成香烟图像的差异。其中，包装形式的趋同性使得这个问题简单化，少量的模式就可以概况绝大部分情况；数量的不同会造成图像灰度、条纹强度变化，这一点需要在特征提取算法中克服；摆放形式带来的影响非常大，发明人将其归结为三种模式。不失一般性，香烟摆放形式可能是图5中六种摆放形式的任意一种。而由于集装箱扫描基本都采用扇形射线束(如图4所示)，因此横向的边缘在射线方向上立体叠加，使得香烟呈现宽度不同的竖条纹。根据发明人的研究，香烟的竖条纹状纹理可归纳为如图7所示的以下三种模式，即：图5(a)和(d)产生图7(a)中的宽条纹，图5(b)和(e)产生图7(b)中稍窄的条纹，而图5(c)和(f)产生图7(c)中的窄条纹。在上述分析的基础上，发明人提出不同摆放形式下的三种香烟模式，并使用算法建立模型。本发明巧妙地利用了上述三种模式的特点，使得在建立模型后，能够自动地训练和/或学习而不依赖于初始人工标注，因此大大节约了人工成本并提高了检测效率和准确度。本领域的技术人员应该理解，在其他实施例中，也可 以使用更多的摆放形式来产生其他的模式。
模型中包括特征提取算法与分类器。根据以上分析，模型的建立过程为：
A)建立香烟检测数据库。采集相当数量的香烟图像，使得其中六种摆放形式均匀分布；数量以最少一箱，最多占满集装箱宽度的条件下均匀分布。在扫描得到的二维图像上，以箱为单位，人工标注矩形的香烟区域。所有标注矩形框中的图像形成正类样本库。此外，随机扫描其它各类货物，形成负类样本库；
B)对A中所有香烟区域中的图像提取特征，形成三种模式下的三个特征集；
C)在负样本中随机抽取三个样本量足够大的负样本集，并提取特征；
D)对B、C中得到的正、负样本训练分类器；
E)可采用机器学习、模式识别领域的现有算法优化分类器。比如借鉴Boosting思想，不断丢弃易分样本，加入新的随机样本，重复C、D步骤直到错误率足够低或算法收敛，得到分类器。
F)可采用现有算法调整正样本位置，降低人工标注误差带来的影响。比如借鉴DPM中隐变量模型的思想，假设人工标注位置不完全准确，在D中得到分类器后，对A中正样本位置进行一定的偏移，得到分类器意义下正样本的位置，重复B到E步骤得到优化分类器。
在本发明的上下文中，建立香烟模型可以独立于自动检测香烟而进行，即，可以在建立了香烟模型之后，使用已建立的香烟模型对透视图像进行自动香烟检测。
自动检测过程
图6示出了对香烟进行检查的过程，香烟相对于射线源运动，产生透射图像，然后进行自动检测。自动检测过程实际上是模型建立过程的一个子集：如图8所示，在步骤S801，输入待检查的图像。在步骤S802，对待检查的图像进行特征提取。根据一些实施例，在此过程中还可以对图像进行处理来得到感兴趣区域。在步骤S803，利用香烟模型对于待测图像，遍历每个感兴趣区域内的像素点，提取该点周围邻域图像并经过特征提取、分类判断，即可得到该点是否为香烟区域，以及这个结论的置信度。在步骤S804，判断是否存在香烟。如果存在，则在步骤S805，在得到每个像素是否为香烟嫌疑区域的结果后，通过连通区提取，即可得到香烟区域。否则，在步骤S806，输出“未检测到”或者丢弃待检测图像。图9示出了以矩形框形式标注出香烟的检测示意图。
自学习过程
优选地，算法具备自学习功能，使得它能够适应不同的检测环境、扫描设备和更多包装形式的香烟。自学习使用的算法和模型建立时基本一致，不同之处在于它在检测过程中仅获取正样本(如图10所示)。
图10示出了根据本发明的实施例的自学习过程的流程图。如图10所示，在步骤S1001，输入待检查图像。在步骤S1002，对待检查的图像进行特征提取。根据一些实施例，在此过程中还可以对图像进行处理来得到感兴趣区域。在步骤S1003，利用香烟模型进行检查。在步骤S1004，判断是否存在香烟。如果存在，则在步骤S1005，在得到每个像素是否为香烟嫌疑区域的结果后，通过连通区提取，即可得到香烟区域。在步骤S1006，在检测过程中，如果人工发现算法未检测到香烟并进行了标注，则算法把标注区域加入正样本，在步骤S1007重新进行训练或进行增量学习，得到更新的模型。但如果得到的模型仍无法检测到刚加入的样本，则放弃这次更新。
另外，通过自动报关单分析(如电子报关单中的关键字段或通过字符识别)，获知某次扫描图像中包含香烟，则算法从检测过程中得到的置信度图中，选取分数最大的区域，默认为香烟并进行重新训练。但如果得到的模型仍无法检测到这次扫描中的香烟，则放弃这次更新。
实例
在以上的描述中给出了所述方法的一般形式，在具体实施算法层面上，可以选用多种已有的算法实现。例如，上文提到的HOG、DPM、Deep Learning均可以直接应用于本发明，通过对正负样本特征提取、分类即可实现检测功能。具体来说，在图像特征提取方面，可以使用各种基于梯度的描述子、纹理描述子等，如HOG、局部二值模式(LBP-Local Binary Pattern)、最大响应集(Maximum Response Sets，或称MR8)，另外可以使用词袋法(BOW-Bag of Words)、纹理表示(Texton)、稀疏表示(Sparse Representation)等进行结构化特征处理；在分类器方面，可以使用各类线性、非线性、集成分类器、神经网络，如Fisher分类器、支持向量机、自助法(Boosting)、随机森林、多层感知机等。
不失一般性，这里给出一个使用具体算法的实施例。然而能够理解，根据本发明的教导，本领域技术人员可以在不偏离本发明思想的情况下对于实施例中的具体算法进行改变或替换。
一、图像获取
本发明的算法普遍适用于各类大型货物/集装箱扫描设备。为保证算法有效性， 需要在获取图像时对图像进行归一化。设原始二维图像信号为X，按照扫描设备物理参数，将X分辨率缩放到5mm/像素，并进行灰度拉伸，即可得到归一化图像
Figure PCTCN2014090563-appb-000001
二、感兴趣区域(ROI，Region Of Interest)提取
检测
Figure PCTCN2014090563-appb-000002
中的空气部分，并将其排除在检测过程之外。将空气部分排除不仅能够提高运算速度，而且还能避免在空气中出现误报。
统计
Figure PCTCN2014090563-appb-000003
的直方图，在直方图中计算最亮峰值a，并拟合以其为中心的空气正态分布(a,σa)，则设定阈值为ta＝a-3*σa。
Figure PCTCN2014090563-appb-000004
中大于阈值ta的像素被认为是空气，不进行检测计算。
三、模型建立
A)建立香烟检测数据库
扫描在集装箱中各个摆放形式、不同数量下的香烟图像，获取香烟图像库
Figure PCTCN2014090563-appb-000005
图像数量在100幅左右即可满足实用要求。采集类似物体和随机物体，形成负样本库{Y}，该负样本库{Y}中的每幅图像均不包含香烟。
B)正样本特征提取
以箱为单位，人工标注香烟位置，形成三种模式下的正样本库P＝P1∪P2∪P3＝{I1m，m∈[1，M]}∪{I2n，n∈[1，N]}∪{I3k，k∈[1，K]}。其中，Pi＝{Iij}表示模式i下得到的正样本库，并且其中每幅图像Iij都是
Figure PCTCN2014090563-appb-000006
中图像的局部，在正样本库中的每幅图像仅包含香烟，M、N、K分别为三种模式各自的样本量，且m、n、k分别是范围为[1,M],[1,N],[1,K]的变量。优选地，对于上述三种香烟模式进行独立训练，因此不要求样本量M、N、K相等。但为达到相当的性能，三种模式样本量应保持基本一致。
提取Iij的特征。不失一般性，实施例采用HOG特征。然而本领域技术人员能够理解，根据本发明的教导，同样可以采用其他已有的方法来提取特征。特征提取后，二维图像Iij转化为高维矢量Fij。样本库P转化为特征库F＝{F1m}∪{F2n}∪{F3k}。
由于三种模式可以是独立训练的，所以下面仅以训练一个模式为例。
C)随机负样本特征提取
在{Y}中随机抽取足够的样本，进行特征提取，形成负样本特征集{N}。注意HOG算法中，特征维数由图像大小决定，因此三种模式维数不同。负样本抽取在不同模式下区域大小不同。
D)在给定正、负样本集的条件下，训练分类器C。训练方法可采用模式识别中 各类已有算法。不失一般性，在本实施例中采用开源工具LIBSVM训练分类器C。
E)挑选难分类样本优化分类器
用分类器C对正、负样本分类。根据结果置信度，将易分的负样本去掉，重新加入足够的随机负样本。然后，重复D)和E)步骤，直到分类器变化足够小，或达到足够迭代次数为止。
F)调整正样本，优化分类器
在步骤B)中的人工标注带来较大的误差，使得香烟条纹相对位置不明确。为减小这一因素的影响，用分类器C对B)中Iij图像周围进行检测，找到Iij的最佳位置，将Iij更新为Iij′，然后迭代进行C)至F)步骤，直到分类器变化足够小，或达到足够迭代次数为止。
四、检测
给定要检测的图像，经过归一化和ROI提取，生成ROI区域的HOG特征H。用滑动窗口遍历H，求它在三种模式下，每个窗口中的最大置信度。置信度大于特定阈值处即为香烟位置。
为提高检出率，可以在多个尺度(即，不同的缩放比例)上进行香烟检测。比如在[0.9,1.0,1.1]三个尺度上对所述透视图像进行检测，可以较好地解决香烟摆放时存在一定偏转的问题。
为降低误报率，可以对检测结果进行后处理去除噪声。比如对上述置信度图进行中值滤波，或对上述置信度图进行二值化后连通区面积滤波等。
此外，为实现实时检测，可以通过均匀/随机采样的方法，仅对图像上一部分点进行检测。实际上，大多特征提取算法如HOG也都采用不重叠的窗口，以窗口代替邻域像素，达到以点代面，局部逼近遍历的效果。
五、自学习过程
自学习过程涉及模型更新策略。自学习使用的算法和模型建立时基本一致，不同之处在于它在检测过程中仅获取正样本(如图10所示)。
在检测过程中，如果人工发现算法未检测到香烟并进行了标注，则算法把标注区域加入正样本，重新进行训练或进行增量学习，得到更新的模型。但如果得到的模型仍无法检测到刚加入的样本，则放弃这次更新。
另外，通过自动报关单分析(如电子报关单中的关键字段或通过字符识别)，获知某次扫描图像中包含香烟，则算法从检测过程中得到的置信度图中，选取分数最大的 区域，默认为香烟并进行重新训练。但如果得到的模型仍无法检测到这次扫描中的香烟，则放弃这次更新。
以上的详细描述通过使用示意图、流程图和/或示例，已经阐述了自动检查香烟的方法和设备的众多实施例。在这种示意图、流程图和/或示例包含一个或多个功能和/或操作的情况下，本领域技术人员应理解，这种示意图、流程图或示例中的每一功能和/或操作可以通过各种结构、硬件、软件、固件或实质上它们的任意组合来单独和/或共同实现。在一个实施例中，本发明的实施例所述主题的若干部分可以通过专用集成电路(ASIC)、现场可编程门阵列(FPGA)、数字信号处理器(DSP)、或其他集成格式来实现。然而，本领域技术人员应认识到，这里所公开的实施例的一些方面在整体上或部分地可以等同地实现在集成电路中，实现为在一台或多台计算机上运行的一个或多个计算机程序(例如，实现为在一台或多台计算机系统上运行的一个或多个程序)，实现为在一个或多个处理器上运行的一个或多个程序(例如，实现为在一个或多个微处理器上运行的一个或多个程序)，实现为固件，或者实质上实现为上述方式的任意组合，并且本领域技术人员根据本公开，将具备设计电路和/或写入软件和/或固件代码的能力。此外，本领域技术人员将认识到，本公开所述主题的机制能够作为多种形式的程序产品进行分发，并且无论实际用来执行分发的信号承载介质的具体类型如何，本公开所述主题的示例性实施例均适用。信号承载介质的示例包括但不限于：可记录型介质，如软盘、硬盘驱动器、紧致盘(CD)、数字通用盘(DVD)、数字磁带、计算机存储器等；以及传输型介质，如数字和/或模拟通信介质(例如，光纤光缆、波导、有线通信链路、无线通信链路等)。
虽然已参照几个典型实施例描述了本发明，但应当理解，所用的术语是说明和示例性、而非限制性的术语。由于本发明能够以多种形式具体实施而不脱离发明的精神或实质，所以应当理解，上述实施例不限于任何前述的细节，而应在随附权利要求所限定的精神和范围内广泛地解释，因此落入权利要求或其等效范围内的全部变化和改型都应为随附权利要求所涵盖。

Claims (14)

  1. 一种检查方法，包括以下步骤：
    获取被检查物体的透视图像；
    对所述透视图像进行处理得到感兴趣区域；以及
    利用香烟模型对感兴趣区域进行自动检测，以确定所述透视图像的感兴趣区域是否属于香烟。
  2. 根据权利要求1所述的方法，其中在所述利用香烟模型对感兴趣区域进行自动检测的步骤之前，还包括建立香烟模型的步骤，所述建立香烟模型的步骤包括：
    建立香烟图像库；
    对所述香烟图像库中图像进行处理得到各自的感兴趣区域；
    对所述香烟图像库中图像的感兴趣区域进行特征提取；
    基于提取的特征训练分类器，生成所述香烟模型。
  3. 根据权利要求2所述的方法，其中所述建立香烟图像库的步骤包括：
    扫描在集装箱中各个摆放形式、不同数量下的香烟图像，得到香烟模型的正样本库；以及
    采集与香烟类似的物体和随机物体的图像，形成香烟模型的负样本库，所述负样本库中每幅图像均不包含香烟。
  4. 根据权利要求2或3所述的方法，其中在所述建立香烟模型的步骤中，香烟的各种规则堆叠的扫描图像被划分为宽度不同的多种模式，由所述多种模式生成所述香烟模型的正样本库。
  5. 根据权利要求4所述的方法，其中对所述香烟图像库中图像的感兴趣区域进行特征提取的步骤包括：
    以箱为单位，人工标注香烟位置，形成所述多种模式下的正样本特征集；以及
    在负样本库中随机抽取多个样本，进行特征提取，形成负样本特征集。
  6. 根据权利要求2所述的方法，其中基于提取的特征训练分类器的步骤包括：
    (1)在给定正、负样本特征集的条件下，训练分类器；
    (2)用分类器对正、负样本分类；
    (3)根据结果置信度，将易分的负样本去掉，重新加入足够的随机负样本；
    (4)重复上述步骤(1)、(2)、和(3)，直到分类器的分类结果变化足够小，或达到足够迭代次数为止。
  7. 根据权利要求1或2所述的方法，其中所述方法还包括对于未检测到香烟的图像，通过人工标注或报关单自动分析，得到新的香烟图像并更新香烟模型。
  8. 根据权利要求1或2所述的方法，其中在所述获取被检查物体的透视图像的步骤中，将图像尺寸缩放到均一的分辨率下，并进行灰度拉伸，使图像值域充满整个可能的取值范围，来进行归一化操作。
  9. 根据权利要求1或2所述的方法，其中在所述对所述透视图像进行处理得到感兴趣区域的步骤中，把所述透视图像中空气部分排除掉，以避免空气噪声产生检测结果。
  10. 根据权利要求9所述的方法，其中在对所述透视图像进行处理得到感兴趣区域的步骤中，采用空气亮度阈值的方法，用空气值对所述透视图像进行二值化，只对阈值之下的图像部分进行香烟检测。
  11. 根据权利要求1或2所述的方法，其中在所述利用香烟模型对感兴趣区域进行自动检测的步骤包括：
    给定要检测的图像，经过归一化和获取感兴趣区域，生成感兴趣区域的HOG特征；
    用滑动窗口遍历，求它在多种模式下，每个窗口中的最大置信度；以及
    置信度大于特定阈值处即为香烟位置。
  12. 根据权利要求1或2所述的方法，其中在多个尺度上检测香烟。
  13. 一种检查设备，包括：
    数字放射摄影装置，对被检查物体进行射线检查，获取被检查物体的透视图像；以及
    图像处理装置，对所述透视图像进行处理得到感兴趣区域，并利用香烟模型对感兴趣区域进行自动检测，以确定所述透视图像的感兴趣区域是否属于香烟。
  14. 根据权利要求13所述的设备，其用于自动检测香烟走私。
PCT/CN2014/090563 2013-11-07 2014-11-07 检查方法和设备 WO2015067208A1 (zh)

Priority Applications (5)

Application Number Priority Date Filing Date Title
JP2016528054A JP6246355B2 (ja) 2013-11-07 2014-11-07 検出方法及びその機器
EP14860336.8A EP3067823A4 (en) 2013-11-07 2014-11-07 Detection method and device
KR1020187008606A KR20180035930A (ko) 2013-11-07 2014-11-07 검출 방법 및 그 설비
US15/034,021 US10013615B2 (en) 2013-11-07 2014-11-07 Inspection methods and devices
KR1020167015104A KR20160083099A (ko) 2013-11-07 2014-11-07 검출 방법 및 그 설비

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201310546289.2A CN104636707B (zh) 2013-11-07 2013-11-07 自动检测香烟的方法
CN201310546289.2 2013-11-07

Publications (1)

Publication Number Publication Date
WO2015067208A1 true WO2015067208A1 (zh) 2015-05-14

Family

ID=53040922

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2014/090563 WO2015067208A1 (zh) 2013-11-07 2014-11-07 检查方法和设备

Country Status (7)

Country Link
US (1) US10013615B2 (zh)
EP (1) EP3067823A4 (zh)
JP (1) JP6246355B2 (zh)
KR (2) KR20180035930A (zh)
CN (1) CN104636707B (zh)
CL (1) CL2016000960A1 (zh)
WO (1) WO2015067208A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115601631A (zh) * 2022-12-15 2023-01-13 深圳爱莫科技有限公司（Cn） 一种卷烟陈列图像识别方法、模型、设备及储存介质
CN116052062A (zh) * 2023-03-07 2023-05-02 深圳爱莫科技有限公司 一种鲁棒的烟草陈列图像处理方法和装置

Families Citing this family (23)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10282623B1 (en) * 2015-09-25 2019-05-07 Apple Inc. Depth perception sensor data processing
CN108108744B (zh) * 2016-11-25 2021-03-02 同方威视技术股份有限公司 用于辐射图像辅助分析的方法及其系统
US10268924B2 (en) 2016-12-05 2019-04-23 Sap Se Systems and methods for integrated cargo inspection
CN108303435B (zh) * 2017-01-12 2020-09-11 同方威视技术股份有限公司 检查设备和对集装箱进行检查的方法
CN108303747B (zh) * 2017-01-12 2023-03-07 清华大学 检查设备和检测枪支的方法
CN107392931A (zh) * 2017-08-08 2017-11-24 南京敏光视觉智能科技有限公司 条烟品牌分类装置及方法
US11058143B2 (en) * 2017-10-19 2021-07-13 R.J. Reynolds Tobacco Company Smoking-related article inspection systems and associated methods
PL3476228T3 (pl) 2017-10-25 2020-11-16 International Tobacco Machinery Poland Sp. Z O.O. Sposób i urządzenie do napełniania pojemników transportowych artykułami prętopodobnymi przemysłu tytoniowego
KR101930062B1 (ko) * 2017-12-27 2019-03-14 클리어라인 주식회사 인공지능기술을 이용한 단계별 자동 교정 시스템
JP6863326B2 (ja) * 2018-03-29 2021-04-21 日本電気株式会社 選別支援装置、選別支援システム、選別支援方法及びプログラム
CN108596820B (zh) * 2018-04-11 2022-04-05 重庆第二师范学院 一种基于信息安全的图像处理系统
CN109410190B (zh) * 2018-10-15 2022-04-29 广东电网有限责任公司 基于高分辨率遥感卫星影像的杆塔倒断检测模型训练方法
CN109446961B (zh) * 2018-10-19 2020-10-30 北京达佳互联信息技术有限公司 姿势检测方法、装置、设备及存储介质
CN109829466A (zh) * 2019-01-23 2019-05-31 中国建筑第八工程局有限公司 一种基于机器视觉的吸烟行为人工智能检测方法
CN110837856B (zh) * 2019-10-31 2023-05-30 深圳市商汤科技有限公司 神经网络训练及目标检测方法、装置、设备和存储介质
CN112633274A (zh) * 2020-12-21 2021-04-09 中国航天空气动力技术研究院 一种声呐图像目标检测方法、装置、电子设备
CN112819001B (zh) * 2021-03-05 2024-02-23 浙江中烟工业有限责任公司 基于深度学习的复杂场景卷烟烟包识别方法和装置
US20230012871A1 (en) * 2021-07-07 2023-01-19 Nanyang Technological University Methods and Systems for Watermarking Neural Networks
IT202100023954A1 (it) * 2021-09-17 2023-03-17 Gd Spa Apparato assemblatore e metodo di assemblaggio per produrre articoli da fumo multicomponente.
IT202100023942A1 (it) * 2021-09-17 2023-03-17 Gd Spa Metodo e sistema per ispezionare articoli da fumo
IL290416A (en) * 2022-02-07 2023-09-01 Seetrue Screening Ltd Identifying prohibited material hidden in an item using image processing
CN114792369B (zh) * 2022-06-29 2022-09-23 上海启迪睿视智能科技有限公司 基于光投影的烟条盒填充状态检测方法及系统
CN115205432B (zh) * 2022-09-03 2022-11-29 深圳爱莫科技有限公司 一种香烟终端陈列样本图像自动生成的仿真方法与模型

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2000011456A1 (en) * 1998-08-20 2000-03-02 Csir Inspection of containers
CN1301960A (zh) * 1999-12-29 2001-07-04 李广寅 集装箱中卷烟探测成像装置
US7706502B2 (en) * 2007-05-31 2010-04-27 Morpho Detection, Inc. Cargo container inspection system and apparatus
CN102483803A (zh) * 2009-05-26 2012-05-30 拉皮斯坎系统股份有限公司 识别特定目标物品的x 射线断层摄影检查系统

Family Cites Families (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6195444B1 (en) 1999-01-12 2001-02-27 Analogic Corporation Apparatus and method for detecting concealed objects in computed tomography data
GB0128659D0 (en) * 2001-11-30 2002-01-23 Qinetiq Ltd Imaging system and method
US8223919B2 (en) * 2003-04-25 2012-07-17 Rapiscan Systems, Inc. X-ray tomographic inspection systems for the identification of specific target items
US7596247B2 (en) 2003-11-14 2009-09-29 Fujifilm Corporation Method and apparatus for object recognition using probability models
KR100695136B1 (ko) 2005-01-04 2007-03-14 삼성전자주식회사 영상의 얼굴검출장치 및 방법
CA2608119A1 (en) * 2005-05-11 2006-11-16 Optosecurity Inc. Method and system for screening luggage items, cargo containers or persons
US7483511B2 (en) * 2006-06-06 2009-01-27 Ge Homeland Protection, Inc. Inspection system and method
JP4818997B2 (ja) 2007-06-29 2011-11-16 オリンパス株式会社 顔検出装置及び顔検出プログラム
CN101403711B (zh) 2007-10-05 2013-06-19 清华大学 液态物品检查方法和设备
JP5444718B2 (ja) * 2009-01-08 2014-03-19 オムロン株式会社 検査方法、検査装置および検査用プログラム
JP5407774B2 (ja) * 2009-11-10 2014-02-05 株式会社島津製作所 放射線撮影装置
JP5707570B2 (ja) 2010-03-16 2015-04-30 パナソニックＩｐマネジメント株式会社 物体識別装置、物体識別方法、及び、物体識別装置の学習方法
CN101853393B (zh) * 2010-04-22 2013-02-13 深圳市鼎为科技有限公司 机器视觉系统检测算法的自动产生和自动学习方法
CN102608673A (zh) * 2012-02-29 2012-07-25 北京无线电计量测试研究所 人体安检系统可疑物品图像显示方法
CN102608672A (zh) * 2012-02-29 2012-07-25 北京无线电计量测试研究所 人体安检系统可疑物品图像显示装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2000011456A1 (en) * 1998-08-20 2000-03-02 Csir Inspection of containers
CN1301960A (zh) * 1999-12-29 2001-07-04 李广寅 集装箱中卷烟探测成像装置
US7706502B2 (en) * 2007-05-31 2010-04-27 Morpho Detection, Inc. Cargo container inspection system and apparatus
CN102483803A (zh) * 2009-05-26 2012-05-30 拉皮斯坎系统股份有限公司 识别特定目标物品的x 射线断层摄影检查系统

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3067823A4 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115601631A (zh) * 2022-12-15 2023-01-13 深圳爱莫科技有限公司（Cn） 一种卷烟陈列图像识别方法、模型、设备及储存介质
CN116052062A (zh) * 2023-03-07 2023-05-02 深圳爱莫科技有限公司 一种鲁棒的烟草陈列图像处理方法和装置

Also Published As

Publication number Publication date
CL2016000960A1 (es) 2016-10-28
JP6246355B2 (ja) 2017-12-13
JP2016535888A (ja) 2016-11-17
US20160335503A1 (en) 2016-11-17
US10013615B2 (en) 2018-07-03
CN104636707A (zh) 2015-05-20
KR20160083099A (ko) 2016-07-11
EP3067823A4 (en) 2017-06-14
CN104636707B (zh) 2018-03-23
EP3067823A1 (en) 2016-09-14
KR20180035930A (ko) 2018-04-06

Similar Documents

Publication Publication Date Title
WO2015067208A1 (zh) 检查方法和设备
EP3349048B1 (en) Inspection devices and methods for detecting a firearm in a luggage
EP3349050B1 (en) Inspection devices and methods for detecting a firearm
Rogers et al. Automated x-ray image analysis for cargo security: Critical review and future promise
US10042079B2 (en) Image-based object detection and feature extraction from a reconstructed charged particle image of a volume of interest
US10509979B2 (en) Inspection methods and systems
Rogers et al. A deep learning framework for the automated inspection of complex dual-energy x-ray cargo imagery
Jaccard et al. Tackling the X-ray cargo inspection challenge using machine learning
CN109522913B (zh) 检查方法和检查设备以及计算机可读介质
US20140010437A1 (en) Compound object separation
Rogers et al. Threat Image Projection (TIP) into X-ray images of cargo containers for training humans and machines
Jaccard et al. Using deep learning on X-ray images to detect threats
WO2017101514A1 (zh) 检查货物的方法、系统和装置
US20090226032A1 (en) Systems and methods for reducing false alarms in detection systems
JP6567703B2 (ja) 検査機器およびコンテナを検査する方法
Rogers et al. Detection of cargo container loads from X-ray images
US9846935B2 (en) Segmentation of sheet objects from image generated using radiation imaging modality
US20210256296A1 (en) Object identification system and computer-implemented method
US20190259160A1 (en) Item classification using localized ct value distribution analysis
Sharma et al. Automatic detection of novel corona virus (SARS-CoV-2) infection in computed tomography scan based on local adaptive thresholding and kernel-support vectors
Kehl et al. Distinguishing malicious fluids in luggage via multi-spectral CT reconstructions
Green et al. Investigation of a Classification-based Technique to Detect Illicit Objects for Aviation Security.

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 14860336

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 15034021

Country of ref document: US

ENP Entry into the national phase

Ref document number: 2016528054

Country of ref document: JP

Kind code of ref document: A

REEP Request for entry into the european phase

Ref document number: 2014860336

Country of ref document: EP

WWE Wipo information: entry into national phase

Ref document number: 2014860336

Country of ref document: EP

NENP Non-entry into the national phase

Ref country code: DE

REG Reference to national code

Ref country code: BR

Ref legal event code: B01A

Ref document number: 112016010097

Country of ref document: BR

ENP Entry into the national phase

Ref document number: 20167015104

Country of ref document: KR

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 112016010097

Country of ref document: BR

Kind code of ref document: A2

Effective date: 20160504