WO2016107481A1 - 检查方法和系统 - Google Patents

检查方法和系统 Download PDF

Info

Publication number
WO2016107481A1
WO2016107481A1 PCT/CN2015/098476 CN2015098476W WO2016107481A1 WO 2016107481 A1 WO2016107481 A1 WO 2016107481A1 CN 2015098476 W CN2015098476 W CN 2015098476W WO 2016107481 A1 WO2016107481 A1 WO 2016107481A1
Authority
WO
WIPO (PCT)
Prior art keywords
interest
region
goods
wine
alcoholic
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/CN2015/098476
Other languages
English (en)
French (fr)
Inventor
陈志强
张丽
赵自然
刘耀红
张多坤
顾建平
李强
张建
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nuctech Co Ltd
Original Assignee
Nuctech Co Ltd
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 Nuctech Co Ltd filed Critical Nuctech Co Ltd
Priority to SG11201607836WA priority Critical patent/SG11201607836WA/en
Priority to US15/129,389 priority patent/US10229336B2/en
Priority to MYPI2016703474A priority patent/MY189002A/en
Publication of WO2016107481A1 publication Critical patent/WO2016107481A1/zh
Anticipated expiration legal-status Critical
Priority to US16/255,738 priority patent/US10509979B2/en
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V5/00—Prospecting or detecting by the use of ionising radiation, e.g. of natural or induced radioactivity
    • G01V5/20—Detecting prohibited goods, e.g. weapons, explosives, hazardous substances, contraband or smuggled objects
    • G01V5/22—Active interrogation, i.e. by irradiating objects or goods using external radiation sources, e.g. using gamma rays or cosmic rays
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N23/00—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00
    • G01N23/02—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00 by transmitting the radiation through the material
    • G01N23/04—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00 by transmitting the radiation through the material and forming images of the material
    • G01N23/046—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00 by transmitting the radiation through the material and forming images of the material using tomography, e.g. computed tomography [CT]
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V5/00—Prospecting or detecting by the use of ionising radiation, e.g. of natural or induced radioactivity
    • G01V5/20—Detecting prohibited goods, e.g. weapons, explosives, hazardous substances, contraband or smuggled objects
    • G01V5/22—Active interrogation, i.e. by irradiating objects or goods using external radiation sources, e.g. using gamma rays or cosmic rays
    • G01V5/226—Active interrogation, i.e. by irradiating objects or goods using external radiation sources, e.g. using gamma rays or cosmic rays using tomography

Definitions

  • Embodiments of the present application relate to automatic detection of suspects in radiation images, and more particularly to methods and corresponding devices for automatic detection of alcoholic goods in a large container scanning system.
  • an inspection method and an inspection system are proposed to be able to check whether the goods are alcoholic goods.
  • an inspection method comprising: acquiring a transmission image of an inspected goods; processing the transmission image to obtain a region of interest; and using the alcohol cargo detection model to perform the region of interest A check is made to determine if the region of interest of the fluoroscopic image contains alcoholic goods.
  • the step of inspecting the region of interest using the alcohol cargo detection model comprises: extracting local target shape information and texture information for the region of interest; utilizing alcohol created based on local target shape features and texture features
  • the cargo detection model classifies local target shape information and texture information of the region of interest to determine whether the region of interest of the transmission image contains alcoholic goods.
  • the region of interest is inspected on a plurality of scales using a wine cargo detection model.
  • the step of processing the transmission image comprises: detecting an air region and an impervious region in the transmission image; excluding an air region and an impervious area from the region of interest.
  • the method further comprises the steps of: scanning images of some known cargo categories, which must include alcoholic goods, and some scanned images of goods similar to the alcoholic goods, but not the wine, manually Marking and training the wine cargo inspection model.
  • the process of manual labeling includes marking the position and placement of the alcoholic goods in the image.
  • the alcohol cargo detection model is established by weighting and summing local target shape features and texture features of the sample, the weights in the weighting process being related to the regional condition entropy.
  • the method further comprises the steps of: manually marking the goods that are not detectable by the alcohol cargo detection model, and detecting the goods that are wine but not actually alcohol, retraining the goods Alcohol cargo inspection model.
  • the alcohol cargo detection model is established for different placement postures of the alcoholic goods.
  • an inspection system comprising: a scanning imaging system that scans an inspected shipment to obtain a transmission image of the inspected shipment; and a data processing device that processes the transmission image to obtain The region of interest is examined using a wine cargo detection model to determine if the region of interest of the fluoroscopic image contains alcoholic goods.
  • the data processing apparatus extracts local target shape information and texture information for the region of interest, and uses the alcohol cargo detection model created based on the local target shape feature and the texture feature to locally target shape information of the region of interest.
  • the texture information is classified to determine whether the region of interest of the transmission image contains alcoholic goods.
  • the alcohol inspection can be automatically performed, and the inspection efficiency and accuracy can be improved.
  • FIGS. 1A and 1B are schematic structural views of an inspection system according to an embodiment of the present disclosure.
  • FIG. 2 is a flow chart describing an inspection method in accordance with an embodiment of the present disclosure
  • Figure 3 shows the posture of the wine placed (the dotted line is the direction of the ray);
  • FIG. 4 shows a schematic diagram of a process of conducting online learning of a liquor model in a method in accordance with an embodiment of the present disclosure
  • FIG. 5 is a flow chart describing performing a positive sample based online learning feedback process in a method in accordance with an embodiment of the present disclosure
  • FIG. 6 is a flow chart describing performing a negative sample based online learning feedback process in a method in accordance with an embodiment of the present disclosure
  • Fig. 7 is a view showing the result of wine detection obtained in the inspection system of the embodiment of the present disclosure.
  • a transmission image of an inspected goods is acquired by a radiation imaging apparatus, and then the transmission image is processed to obtain a region of interest.
  • the region of interest is examined using a wine cargo detection model to determine if the region of interest of the fluoroscopic image contains alcoholic goods.
  • alcoholic cargo testing relies primarily on detecting the presence of a container of wine, such as a particular type of wine bottle.
  • the inspection system of the embodiment of the present disclosure further has an online learning function, for example, online learning can be performed on a typical image that is confirmed as wine but not detected and detected as alcohol is not alcohol, and the update detection algorithm is automatically updated to eliminate the alcohol container. Or the effect of changes in posture on the test results.
  • online learning can be performed on a typical image that is confirmed as wine but not detected and detected as alcohol is not alcohol, and the update detection algorithm is automatically updated to eliminate the alcohol container. Or the effect of changes in posture on the test results.
  • FIGS. 1A and 1B are schematic structural views of an inspection system according to an embodiment of the present disclosure.
  • Figure 1A A top schematic view of the inspection system is shown, and
  • FIG. 1B shows a front elevational view of the inspection system.
  • 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 140 and converts them into numbers.
  • the signal is then obtained in a transmission image at a data processing device 160, such as a computer.
  • a data processing device 160 such as a computer.
  • the transmission image of the container truck 140 is obtained by scanning
  • the transmission image is processed by the data processing device 160 to obtain a region of interest, such as a body cargo region, and then interested in the wine cargo detection model.
  • the area is checked to determine if the region of interest of the fluoroscopic image contains alcoholic goods.
  • a transmission image of the inspected goods is obtained by the transmission scanning system.
  • a perspective image can be produced.
  • the energy/dose, detector size, mechanical structure, vehicle speed, and beaming frequency of the source of the different scanning devices are different, so that the respective rates of image acquisition are slightly different.
  • the fluoroscopic image is first normalized as needed, ie, brightness normalization and resolution are uniformly scaled to 5*5 (mm/pixel).
  • the data processing device 160 processes the transmission image to obtain a region of interest.
  • pre-processing can be performed as needed. Pre-processing is to remove the effects of air and impervious areas in the image.
  • the air region and the opaque region in the fluoroscopic image are directly detected by using a threshold value, and the regions are excluded to obtain the region of interest.
  • the effects of air areas and impervious areas can be avoided, which not only speeds up the detection but also reduces false alarms.
  • the data processing device 160 checks the region of interest using a wine cargo detection model to determine whether the region of interest of the fluoroscopic image contains alcoholic goods.
  • the region of interest is inspected using a wine cargo detection model trained in advance based on the sample image.
  • the sample image is subjected to the same pre-processing as above, and then the feature is extracted, and a wine cargo detection model is established based on the extracted features.
  • the model can then be trained using a pre-built sample library (eg, including a positive sample library and a negative sample library). Online learning and training as well as retraining can also be performed during the testing process.
  • the local target shape information and the texture information are extracted for the region of interest, and then the local target shape information and the texture information of the region of interest are classified by using the alcohol cargo detection model created based on the local target shape feature and the texture feature. Determining whether the region of interest of the transmitted image contains Alcoholic goods.
  • the process of feature extraction, model building, and model updating is described in detail below.
  • Image feature extraction is the first step in object classification and detection.
  • Low-level feature extraction generally includes: based on point of interest detection and based on dense extraction. Based on point of interest detection, such as Harris corner point detector, FAST (Features from Accelerated Segment Test) operator, LoG (Laplacian of Gaussian), DoG (Difference of Gaussian), and the like.
  • Local features such as SIFT (Scale-invariant feature transform), HOG (Histogram of Oriented Gradient), LBP (Local Binary Pattern), etc., are generally employed based on the dense manner. Because the content in the fluoroscopic image is very different, using a feature representation can not meet the actual needs.
  • the HOG feature describes a local image gradient or a directional density distribution of an edge, which can be considered as a description of a local target representation and shape.
  • the LBP can be considered as describing a texture information. Shape information is more useful than texture information in target detection applications, but the two can complement each other.
  • the underlying features utilize a HOG-LBP adaptive fusion feature (Boosted HOG-LBP) to describe the fluoroscopic image for the features at (i, j):
  • Boosted HOG-LBP HOG-LBP adaptive fusion feature
  • ⁇ i,j is an amount related to the regional conditional entropy.
  • HOG i,j and LBP i,j represent the HOG feature and the LBP feature of the corresponding location , respectively.
  • the regional condition entropy is small, so that when the goods are dense, the texture features account for a larger proportion.
  • the most original information of the image may be described in an embodiment of the present disclosure with a mid-level feature, that is, an image slice.
  • This mid-level feature can be obtained by clustering from a large number of images.
  • the training of the classification model requires a large number of labeled positive samples (negative samples are not labeled).
  • the labeling of positive samples is mostly by manual labeling.
  • the weak labeling method can reduce the degree of human participation, making the labeling process simple and efficient.
  • two methods are used for labeling: 1) using a rectangular box to indicate the approximate position of the wine in the image, and using a variable to describe the posture of the wine placement, the position and posture can be used as hidden variables when the model is trained; 2) "Position + Detection" two steps to achieve weak labeling.
  • the “Positioning + Detection” method belongs to the weak labeling algorithm, that is, to give a sample of the fluoroscopic image, it is only necessary to inform whether there is wine in the image, and how to place the bottle.
  • the bottle In actual cargo transportation, the bottle is generally erect Stacked in containers, there are also cases of horizontal and vertical stacking, but the probability of occurrence is low.
  • the posture of the bottle is divided into five types: A erect, B inverted, C horizontal perpendicular to the ray direction, D horizontal parallel to the ray direction 1, E horizontal parallel to the ray direction 2, as shown in Figure 3. .
  • the bottles are stacked regularly and irregularly placed in the bulk of the bottle is very rare.
  • the image can be densely divided into image patches of equal size and then clustered by a clustering algorithm. Since the K-means application distance measure aggregates images that do not look similar in appearance, in order to solve this problem, the patent uses a method of detection clustering: first use K-means clustering, then for each A cluster uses a linear SVM to train a classifier, which is detected in the image respectively. The class that obtains the most detection results is a positive sample, and the detected area is the area where the wine is located. After the "positioning + monitoring" process, you can get the area of the wine in the perspective image, plus the wine placement gesture can be treated the same as the first labeling method. In a specific implementation of the present disclosure, either of the two methods can be selected to handle the positive sample labeling problem.
  • the approximate area and posture category of the wine were obtained through weak labeling. With this information, the training of the test model can be performed. For wines with similar bottle packaging, five component discriminants can be trained to detect five types of wine. Combining the five components to form a total classification detection model is the Mult-Component detection model.
  • ⁇ is the Mult-Component model parameter
  • ⁇ i is the i-th component
  • z and i are both hidden variables.
  • I is a feature vector. In the fluoroscopic image, the scale of the object changes little, and the number of layers of the feature pyramid can be compressed to only two layers during the training of the detection model. The specific training process can be divided into two major steps:
  • the first step is to randomly sample a certain size image in the area with wine as a positive sample, randomly sample a certain number of images as a negative sample in the area without wine, and use SVM to train to obtain a classifier.
  • Step 2 Using the classifier obtained in the first step, the image containing the wine is detected in the image of the wine as a new aligned positive sample, and the classifier obtained in the first step is detected on the image without the wine.
  • the image slice is used as a hard negative sample, and the negative sample described above is retrained to obtain a detection model.
  • the situation described above is the offline training (learning) process, the training data is prepared in advance, and the training model is gradually obtained to obtain the detection model.
  • a detection model based on a large amount of training data offline learning can meet the needs of wine detection.
  • the number of training According to the fact that no such data is collected in the library, the bottles that have not been learned in the offline training will have a poor effect when tested. In this case, manual participation is required. For typical bottle images that are not detected, manual weak marking is performed and re-training is called incremental learning or online learning.
  • Offline learning and online learning can be explained by a combination of "internal loop” and “outer loop” learning modes, as shown in Figure 4.
  • the inner loop means that for each sample to be tested, the system processes the unknown sample using the off-line trained model, and there is no external human participation in this process.
  • the external circulation means that the map inspectors participate in the detection system in an appropriate manner, supervise online learning of the alcohol cargo inspection model, update the detection model, and external personnel directly participate in the detection process. This method can cope with the detection situation in special occasions. When learning online, you can use the Online-SVM method to directly update the alcohol cargo inspection model, or you can use it as a component of the original model. The latter can be trained with SVM using a positive sample plus multiple negative samples.
  • step S51 a fluoroscopic image input system having a wine is used in step 52 to perform a detection process using the detection model.
  • step S53 if the result is not detected and the bottle is typical, one component can be trained to join the original model with one positive sample and a plurality of negative samples, and the detection model is updated in step S54 to increase the recall rate. If the result of the check is correct, the goods are marked and an alarm is issued in step S55.
  • the detection area can be added to the negative sample library, so that the detection boundary shrinks and the false alarm is reduced.
  • the model is updated. As shown in Fig. 6, a fluoroscopic image input system without wine is input at step S61 and a detection process is performed using the detection model at step S62. If there is a check result in step S63, the area image slice is added to the negative sample bank in step S64, and the model is updated in step S65.
  • a sliding window that is, using a traversal method, sliding the fixed-size windows on various scales of the image, and calculating whether each pixel is a target object on each scale.
  • This method is processed on multiple scales and the time cost is relatively high.
  • the scale change of the object in the image is small, so in some embodiments, it can be slid on two scales, so that acceleration is achieved under the premise of ensuring detection accuracy.
  • Fig. 7 is a view showing the result of wine detection obtained in the inspection system of the embodiment of the present disclosure. As shown in Fig. 7, the scheme according to the present embodiment can not only detect the result but also indicate the wine whose detection result is that posture.
  • the solution of the above embodiment can perform wine detection on goods, especially container scanned images, and can judge The figure staff is intelligently assisted.
  • 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.).

Landscapes

  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Geophysics (AREA)
  • General Life Sciences & Earth Sciences (AREA)
  • High Energy & Nuclear Physics (AREA)
  • Theoretical Computer Science (AREA)
  • Pathology (AREA)
  • Analytical Chemistry (AREA)
  • Biochemistry (AREA)
  • General Health & Medical Sciences (AREA)
  • Radiology & Medical Imaging (AREA)
  • Immunology (AREA)
  • Chemical & Material Sciences (AREA)
  • Pulmonology (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Engineering & Computer Science (AREA)
  • Analysing Materials By The Use Of Radiation (AREA)
  • Apparatus For Radiation Diagnosis (AREA)
  • Image Analysis (AREA)

Abstract

一种检查方法和系统，用于检查货物中是否包含酒类。该方法包括步骤：获取被检查物体的透射图像（S21）；对所述透射图像进行处理，得到感兴趣区域（S22）；利用酒类货物检测模型对所述感兴趣区域进行检查，以确定所述透视图像的感兴趣区域是否包含酒类货物（S23）。上述方案可以为判图人员提供智能辅助。

Description

检查方法和系统 技术领域
本申请的实施例涉及辐射图像中嫌疑物自动检测，具体而言涉及大型集装箱扫描系统中，对酒类货物的自动检测的方法及相应的设备。
背景技术
安全检查在反恐、打击贩毒走私等领域有十分重要的意义。世界各国对打击走私越来越重视，对海关集装箱、行李物品等的检查要求也越来越高。近年来，酒类走私不仅冲击酒类市场，扰乱正常市场经济秩序，使得相关国家税收遭受巨大损失，还损害消费者正常权益甚至给其他犯罪活动提供资金。由此可见，打击酒走私，具有重要意义。
大宗酒走私多通过集装箱运送，普通的检测手段很难对所有的酒都做出正确的检测，也不可能对所有的集装箱都开箱检查。而辐射成像技术通过对货物、行李等透视成像，可以达到不开箱进行检测的目的，已经在机场、海关、车站等场所广泛应用，是安检领域违禁品检测的重要手段。但是在集装箱透视图像检查过程中，由于货物种类千差万别，判图员经验参差不齐，走私品出现概率低且长期判图造成的疲劳，使得人工判图的效果大打折扣。这迫切需要针对违禁品的智能检测功能，对人工判图提供辅助。
发明内容
鉴于现有技术中的一个或多个问题，提出了一种检查方法和检查系统，能够检查货物是否为酒类货物。
在本公开的一个方面，提出了一种检查方法，包括：获取被检查货物的透射图像；对所述透射图像进行处理，得到感兴趣区域；利用酒类货物检测模型对所述感兴趣区域进行检查，以确定所述透视图像的感兴趣区域是否包含酒类货物。
优选地，利用酒类货物检测模型对所述感兴趣区域进行检查的步骤包括：对所述感兴趣区域提取局部目标形状信息和纹理信息；利用基于局部目标形状特征和纹理特征而创建的酒类货物检测模型对所述感兴趣区域的局部目标形状信息和纹理信息进行分类，以确定所述透射图像的感兴趣区域是否包含酒类货物。
优选地，利用酒类货物检测模型在多个尺度上对感兴趣区域进行检查。
优选地，对所述透射图像进行处理的步骤包括：检测所述透射图像中的空气区域和穿不透区域；将空气区域和穿不透区域排除在所述感兴趣区域之外。
优选地，所述的方法还包括步骤：通过一些已知货物类别的扫描图像，其中必须包含酒类货物，以及一些跟酒类货物图像类似，但不是酒的货物的扫描图像，通过人工方式进行标注，训练所述酒类货物检测模型。
优选地，人工标注的过程包括标注图像中的酒类货物的位置和摆放姿态。
优选地，所述酒类货物检测模型是通过对样本的局部目标形状特征和纹理特征加权求和而建立的，所述加权过程中的权重与区域条件熵相关。
优选地，所述的方法还包括步骤：对于未能够通过所述酒类货物检测模型检测到的货物，以及检测出为酒但实际不是酒类的货物，通过人工方式进行标注，重新训练所述酒类货物检测模型。
优选地，所述酒类货物检测模型是针对酒类货物的不同摆放姿态而建立的。
在本公开的另一方面，提出了一种检查系统，包括：扫描成像系统，对被检查货物进行扫描，以获取被检查货物的透射图像；数据处理装置，对所述透射图像进行处理，得到感兴趣区域，并利用酒类货物检测模型对所述感兴趣区域进行检查，以确定所述透视图像的感兴趣区域是否包含酒类货物。
优选地，数据处理装置对所述感兴趣区域提取局部目标形状信息和纹理信息，并利用基于局部目标形状特征和纹理特征而创建的酒类货物检测模型对所述感兴趣区域的局部目标形状信息和纹理信息进行分类，以确定所述透射图像的感兴趣区域是否包含酒类货物。
利用上述方案，能够对自动进行酒类检查，提高了检查效率和准确性。
附图说明
为了更好的理解本公开，将根据以下附图对本公开的实施例进行描述：
图1A和图1B是根据本公开一个实施方式的检查系统的结构示意图；
图2是描述根据本公开实施例的检查方法的流程图；
图3示出了酒摆放姿态(虚线为射线方向)；
图4示出了在根据本公开实施例的方法中对酒类模型进行在线学习的过程的示意图；
图5是描述在根据本公开实施例的方法中进行基于正样本的在线学习反馈过程的流程图；
图6是描述在根据本公开实施例的方法中进行基于负样本的在线学习反馈过程的流程图；
图7示出了在本公开实施例的检查系统中得到的酒检测结果的示意图。
附图没有对实施例的所有电路或结构进行显示。贯穿所有附图相同的附图标记表示相同或相似的部件或特征。
具体实施方式
下面将详细描述本公开的具体实施例，应当注意，这里描述的实施例只用于举例说明，并不用于限制本公开。在以下描述中，为了提供对本公开的透彻理解，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公开。在其他实例中，为了避免混淆本公开，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公开至少一个实施例中。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。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。
针对现有技术的问题，本公开的实施例提出，通过辐射成像设备获取被检查货物的透射图像，然后对所述透射图像进行处理，得到感兴趣区域。利用酒类货物检测模型对所述感兴趣区域进行检查，以确定所述透视图像的感兴趣区域是否包含酒类货物。这样，能够在不开集装箱的情况下，利用X射线对集装箱进行扫描，自动检测透射图像中是否有酒类货物。如果存在酒类货物，则透射图像中标明其相应位置，以此作为安全检查的一种辅助手段。在本公开的实施例中，酒类货物检测主要依靠检测是否存在盛酒的容器，如特定类型的酒瓶来实现。
另外，本公开实施例的检查系统还具备在线学习功能，例如可以对确认为酒但又检测不出和检测为酒实际不是酒的典型图像进行在线学习，自动更新升级检测算法以消除酒类容器或姿态变化对检测结果的影响。
图1A和图1B是根据本公开一个实施方式的检查系统的结构示意图。图1A 示出了检查系统的俯视示意图，图1B示出了检查系统的正视示意图。如图1A和图1B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140的射线，转换为数字信号，然后在诸如计算机之类的数据处理装置160得到透射图像。根据本公开的实施例，在通过扫描得到集装箱卡车140的透射图像后，在数据处理装置160对透射图像进行处理，得到感兴趣区域，例如主体货物区域，然后利用酒类货物检测模型对感兴趣区域进行检查，以确定透视图像的感兴趣区域是否包含酒类货物。
图2是描述根据本公开实施例的检查方法的流程图。如图2所示，在步骤S21，通过透射扫描系统获得被检查货物的透射图像。例如，当被检测集装箱卡车以一定的速度经过检测区域时，就可以产生一幅透视图像。不同扫描设备的射线源的能量/剂量、探测器尺寸、机械结构、车速和出束频率的不同，使得获取图像的分别率略有不同。在一些实施例中，根据需要对透视图像先做归一化处理，即亮度归一化和分辨率统一缩放到5*5(mm/pixel)。
在步骤S22，数据处理装置160对所述透射图像进行处理，得到感兴趣区域。在一些实施例中，针对透视图像，可以根据需要进行预处理。预处理就是要去除图像中的空气区域和穿不透区域的影响。本公开实施例中直接采用阈值的方法检测出透视图像中的空气区域和穿不透区域，排除这些区域，得到感兴趣区域。在进行酒类货物检测的时候，可以避免空气区域和穿不透区域的影响，不仅可以加快检测速度，还可以减少误报。在其他实施例中，也可以排除透射图像中的集装箱内主体货物区之外的部分，也就是将主体货物区作为感兴趣区域。
在步骤S23，数据处理装置160利用酒类货物检测模型对所述感兴趣区域进行检查，以确定所述透视图像的感兴趣区域是否包含酒类货物。例如，利用事先根据样本图像训练的酒类货物检测模型对感兴趣区域进行检查。在建立模型的过程中，对样本图像进行与上述相同的预处理，然后抽取特征，在所抽取的特征的基础上建立酒类货物检测模型。然后可以使用事先建立的样本库(例如包括正样本库和负样本库)对模型进行训练。也可以在检测过程中进行在线学习和训练以及再训练。
具体而言，对感兴趣区域提取局部目标形状信息和纹理信息，然后利用基于局部目标形状特征和纹理特征而创建的酒类货物检测模型对感兴趣区域的局部目标形状信息和纹理信息进行分类，以确定所述透射图像的感兴趣区域是否包含 酒类货物。下面详细描述特征提取、模型建立和模型更新的过程。
图像特征提取是物体分类与检测的第一步，底层特征(low-level feature)提取一般包括：基于兴趣点检测和基于密集提取的方式。基于兴趣点检测，如Harris角点检测子、FAST(Features from Accelerated Segment Test)算子、LoG(Laplacian of Gaussian)、DoG(Difference of Gaussian)等。基于密集提取(dense)的方式，一般采用局部特征，如SIFT(Scale-invariant feature transform)、HOG(Histogram of Oriented Gradient)、LBP(Local Binary Pattern)等。由于透视图像中内容千差万别，利用一种特征表示，无法满足实际需求。例如，HOG特征描述了局部图像梯度或者边缘的方向密度分布，可以认为是一种局部目标表象和形状的描述。而LBP可以认为描述的是一种纹理信息。在目标检测应用中，形状信息要比纹理信息更为有用，但二者可以互补。在透视图像中，货物密集堆在一起的时候，在图像上形状信息不在明显，而纹理特征表现明显；大型货物单个摆放时，纹理信息不明显，形状信息明显。为了应多货物类型的多样性，在本公开的实施例中底层特征利用一种HOG-LBP自适应融合特征(Boosted HOG-LBP)来描述透视图像，对于(i，j)处的特征：
Fi，j＝αi，jHOGi，j+(1-αi，j)LBPi，j
其中，αi，j是一个和区域条件熵相关的量。HOGi，j和LBPi，j分别表示相应位置的HOG特征和LBP特征。货物密集的区域其区域条件熵值小，这样货物密集时，纹理特征占的比例多些。
另外，为了提高检测性能，在本公开的实施例中可以用中层特征(mid-level feature)即图像片来描述图像的最原始信息。这种中层特征可以从海量的图像片中聚类获得。
分类模型(二分类即是检测，下文也说检测模型)的训练需要大量的经过标注的正样本(负样本不用标注)。目前，正样本的标注多采用人工标注的方式，虽然有一些辅助标注的算法和软件，但无法完全解脱人。弱标注的方法可以降低人参与的程度，使得标注过程简单高效。本公开实施例中采用两种方法进行标注：1)用一个矩形框标示出酒在图像中的大致位置，用一个变量描述酒摆放姿态，位置和姿态在模型训练的时候可以作为隐变量；2)“定位+检测”两步实现弱标注。
“定位+检测”方法属于弱标注算法，即给出一幅透视图像样本，只需要告知这幅图像中是否有酒，酒瓶如何摆放即可。在实际货物运输中酒瓶一般是正立 堆放在集装箱内，也有横着、倒着堆放的情况，但出现的概率较低。根据实际情况，酒瓶摆放姿态一共分为五种：A正立、B倒立、C水平垂直于射线方向、D水平平行于射线方向1、E水平平行于射线方向2，如图3所示。一幅已知酒瓶摆放姿态的透视图像中，酒瓶是规则摆放堆积在一起，不规则摆放在酒瓶大量运输中是很罕见的。这时可以将图像密集划分为等大小的图像块(image patch)，然后用聚类算法聚类。由于K-means应用距离测度会把视觉看起来并不相似的图像片聚成一类，为了解决这个问题，该专利使用检测式聚类的方法来处理：先利用K-means聚类，然后对每一个聚类用线性SVM训练一个分类器，分别在图像中检测，获得检测结果最多的类就是正样本，检测到的区域即是酒所在的区域。经过“定位+监测”处理，可以得到酒在透视图像中的区域，再加上酒摆放姿态就可以和第一种标注方式一样处理了。在本公开的具体实施中，可以选择两种方法的任意一种来处理正样本标注问题。
前文中经过弱标注获得了酒的大致区域和姿态类别，有了这些信息就可以进行检测模型的训练了。对于具有类似酒瓶包装的酒，可以训练出五个组件(Component)判别式用于检测五种摆放姿态的酒。把五个组件组合在一起构成总的分类检测模型即是Mult-Component检测模型。
在训练模型时，每个样本x的得分为：
Figure PCTCN2015098476-appb-000001
β是Mult-Component模型参数，βi是第i个组件，z和i都是隐变量，
Figure PCTCN2015098476-appb-000002
是特征向量。在透视图像中，物体的尺度变化都不大，在检测模型训练过程中特征金字塔的层数可以压缩到只有两层。具体训练过程可以分为两大步：
第一步：在有酒的区域随机采样一定大小的图像片作为正样本，把没有酒的区域随机采样一定数量的图像片作为负样本，利用SVM进行训练得到一个分类器。
第二步：利用第一步获得的分类器，在酒的图像中检测出包含酒的图像片作为新的经过对齐的正样本，利用第一步获得的分类器在没有酒的图像上检测得到的图像片作为难的负样本(hard negative sample)，再加上述的负样本重新训练得到检测模型。
前面描述的情况是离线训练(学习)过程，提前准备好训练数据，逐步进行训练得到检测模型。基于大量训练数据离线学习得到的检测模型可以满足酒检测的需求。但在实际检测应用中，会有不同种类的酒和不同种类的包装瓶，训练数 据库中没有采集到这样的数据，在离线训练中没有学习到的包装瓶，检测时效果会比较差。这种情况下，需要人工参与，对于没有检出的典型酒瓶图像，人工进行弱标注，重新进行训练，称为增量学习或在线学习。
离线学习和在线学习，其实可以用“内循环”和“外循环”相结合的学习模式来解释，见图4。内循环是指对于每一个待检测样本，系统利用离线训练好的模型对未知样本进行处理，这个过程没有外部人工参与。而外循环是指验图人员以适当的方式参与到检测系统中来，对酒类货物检测模型进行有监督的在线学习，更新检测模型，外部人员直接参与了检测过程。这种方式可以应对特殊场合下的检测情况。在线学习时，可以采用Online-SVM的方式直接更新酒类货物检测模型，也可以作为原模型的一个组件。后者可以用一个正样本加上多个负样本用SVM来训练。
外循环的具体操作过程：当验图人员对比检测结果与输入样本，如果检测结果正确，可以不做任何监督处理，如果检测结果错误，需要把结果反馈给系统，具体反馈机制见图5和图6。如图5所示，在步骤S51，一幅有酒的透视图像输入系统，在步骤52使用检测模型执行检测处理过程。在步骤S53，如果没有检出结果且酒瓶比较典型，则可以用一个正样本和多个负样本训练一个组件加入到原模型中，并且在步骤S54对检测模型进行更新，提高召回率。如果检出结果正确则在步骤S55对货物进行标记并发出警报。如果检出结果错误，在步骤S56结合支持向量机训练模型检测高精度低召回的特点，可以把该检出区域增加到负样本库中，使得检测边界收缩，减少误报。在步骤S57，对模型进行更新。如图6所示，在步骤S61一幅没有酒的透视图像输入系统并且在步骤S62使用检测模型进行检测处理过程。如果在步骤S63有检出结果，则在步骤S64把该区域图像片加入到负样本库中，并且在步骤S65对模型进行更新。
在对货物的检测过程中，可以使用滑动窗口(Sliding Window)来实现，即采用遍历方式，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。这种方法在多尺度上处理，时间代价比较高。根据透视图像的特点：图像中物体的尺度变化很小，故在一些实施例中，可以在两个尺度上滑动，这样保证检测精度的前提下实现了加速。图7示出了在本公开实施例的检查系统中得到的酒检测结果的示意图。如图7所示，根据本实施例的方案不仅能够检出结果，还能指出检出结果是那种姿态的酒。
上述实施例的方案能够对货物特别是集装箱扫描图像进行酒检测，可以对判 图人员以智能辅助。
以上的详细描述通过使用示意图、流程图和/或示例，已经阐述了检查方法和系统的众多实施例。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，本公开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公开，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。由于本公开能够以多种形式具体实施而不脱离公开的精神或实质，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 (11)

  1. 一种检查方法，包括：
    获取被检查货物的透射图像；
    对所述透射图像进行处理，得到感兴趣区域；
    利用酒类货物检测模型对所述感兴趣区域进行检查，以确定所述透视图像的感兴趣区域是否包含酒类货物。
  2. 如权利要求1所述的方法，其中利用酒类货物检测模型对所述感兴趣区域进行检查的步骤包括：
    对所述感兴趣区域提取局部目标形状信息和纹理信息；
    利用基于局部目标形状特征和纹理特征而创建的酒类货物检测模型对所述感兴趣区域的局部目标形状信息和纹理信息进行分类，以确定所述透射图像的感兴趣区域是否包含酒类货物。
  3. 如权利要求1所述的方法，其中利用酒类货物检测模型在多个尺度上对感兴趣区域进行检查。
  4. 如权利要求1所述的方法，其中对所述透射图像进行处理的步骤包括：
    检测所述透射图像中的空气区域和穿不透区域；
    将空气区域和穿不透区域排除在所述感兴趣区域之外。
  5. 如权利要求1所述的方法，还包括步骤：
    通过一些已知货物类别的扫描图像，其中必须包含酒类货物，以及一些跟酒类货物图像类似，但不是酒的货物的扫描图像，通过人工方式进行标注，训练所述酒类货物检测模型。
  6. 如权利要求5所述的方法，其中人工标注的过程包括标注图像中的酒类货物的位置和摆放姿态。
  7. 如权利要求5所述的方法，其中所述酒类货物检测模型是通过对样本的局部目标形状特征和纹理特征加权求和而建立的，所述加权过程中的权重与区域条件熵相关。
  8. 如权利要求5所述的方法，其中所述酒类货物检测模型是针对酒类货物的不同摆放姿态而建立的。
  9. 如权利要求1所述的方法，还包括步骤：
    对于未能够通过所述酒类货物检测模型检测到的酒类货物，以及通过酒类货 物检测模型检测出为酒类货物，但实际上不是的货物，通过人工方式进行标注，重新训练所述酒类货物检测模型。
  10. 一种检查系统，包括：
    扫描成像系统，对被检查货物进行扫描，以获取被检查货物的透射图像；
    数据处理装置，对所述透射图像进行处理，得到感兴趣区域，并利用酒类货物检测模型对所述感兴趣区域进行检查，以确定所述透视图像的感兴趣区域是否包含酒类货物。
  11. 如权利要求10所述的检查系统，其中数据处理装置对所述感兴趣区域提取局部目标形状信息和纹理信息，并利用基于局部目标形状特征和纹理特征而创建的酒类货物检测模型对所述感兴趣区域的局部目标形状信息和纹理信息进行分类，以确定所述透射图像的感兴趣区域是否包含酒类货物。
PCT/CN2015/098476 2014-12-30 2015-12-23 检查方法和系统 Ceased WO2016107481A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
SG11201607836WA SG11201607836WA (en) 2014-12-30 2015-12-23 Inspection methods and systems
US15/129,389 US10229336B2 (en) 2014-12-30 2015-12-23 Inspection methods and systems
MYPI2016703474A MY189002A (en) 2014-12-30 2015-12-23 Inspection methods and systems
US16/255,738 US10509979B2 (en) 2014-12-30 2019-01-23 Inspection methods and systems

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201410844405.3A CN105809091B (zh) 2014-12-30 2014-12-30 检查方法和系统
CN201410844405.3 2014-12-30

Related Child Applications (2)

Application Number Title Priority Date Filing Date
US15/129,389 A-371-Of-International US10229336B2 (en) 2014-12-30 2015-12-23 Inspection methods and systems
US16/255,738 Continuation US10509979B2 (en) 2014-12-30 2019-01-23 Inspection methods and systems

Publications (1)

Publication Number Publication Date
WO2016107481A1 true WO2016107481A1 (zh) 2016-07-07

Family

ID=55068875

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/098476 Ceased WO2016107481A1 (zh) 2014-12-30 2015-12-23 检查方法和系统

Country Status (6)

Country Link
US (2) US10229336B2 (zh)
EP (1) EP3040743B1 (zh)
CN (1) CN105809091B (zh)
MY (1) MY189002A (zh)
SG (1) SG11201607836WA (zh)
WO (1) WO2016107481A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107463945A (zh) * 2017-07-12 2017-12-12 浙江大学 一种基于深度匹配网络的商品种类识别方法
CN107463946A (zh) * 2017-07-12 2017-12-12 浙江大学 一种结合模板匹配与深度学习的商品种类检测方法

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105809091B (zh) * 2014-12-30 2020-10-16 同方威视技术股份有限公司 检查方法和系统
CN108109455A (zh) * 2016-11-24 2018-06-01 同方威视技术股份有限公司 安检模拟培训装置、方法及系统
CN109993767B (zh) * 2017-12-28 2021-10-12 北京京东尚科信息技术有限公司 图像处理方法和系统
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
CN109001833A (zh) * 2018-06-22 2018-12-14 天和防务技术（北京）有限公司 一种基于深度学习的太赫兹危险品检测方法
CN109001834A (zh) * 2018-06-22 2018-12-14 天和防务技术（北京）有限公司 一种基于主动式太赫兹安检方法
CN110335250A (zh) * 2019-05-31 2019-10-15 上海联影智能医疗科技有限公司 网络训练方法、装置、检测方法、计算机设备和存储介质
CN111539920B (zh) * 2020-04-16 2022-03-15 泸州老窖酿酒有限责任公司 一种白酒酿造过程中糟醅质量的自动检测方法
CN113505807A (zh) * 2021-06-02 2021-10-15 上海物影科技有限公司 安检设备的训练方法、装置、系统及存储介质
US11836909B2 (en) * 2022-01-26 2023-12-05 Mellanox Technologies, Ltd. Active learning of product inspection engine
CN118279578B (zh) * 2022-12-30 2025-05-27 同方威视科技江苏有限公司 Ct图像处理方法及装置、以及国际快件的查验方法及装置
JP2024129324A (ja) * 2023-03-13 2024-09-27 株式会社日立ソリューションズ Ｘ線検査装置及びｘ線検査方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1802676A (zh) * 2003-06-11 2006-07-12 昆腾磁性元件公司 乘客和行李的检查站
CN101558327A (zh) * 2006-08-31 2009-10-14 Ge本国防护股份有限公司 用于集成爆炸物检测系统的系统和方法
US20090290757A1 (en) * 2008-05-22 2009-11-26 Mian Zahid F Inspection using three-dimensional profile information
CN101606084A (zh) * 2007-02-05 2009-12-16 通用电气公司 检查系统和方法

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6111974A (en) 1998-02-11 2000-08-29 Analogic Corporation Apparatus and method for detecting sheet objects in computed tomography data
US20050058242A1 (en) * 2003-09-15 2005-03-17 Peschmann Kristian R. Methods and systems for the rapid detection of concealed objects
US20050117700A1 (en) * 2003-08-08 2005-06-02 Peschmann Kristian R. Methods and systems for the rapid detection of concealed objects
GB2441551B (en) 2006-09-05 2008-11-19 Nuctech Co Ltd Method and device for security inspection of liquid articles using radiation
WO2008034232A1 (en) 2006-09-18 2008-03-27 Optosecurity Inc. Method and apparatus for assessing characteristics of liquids
JP5885661B2 (ja) * 2009-05-26 2016-03-15 ラピスカン システムズ、インコーポレイテッド 特定の対象物品の識別のためのｘ線断層撮影検査システム
EP2375243B1 (de) * 2010-04-08 2023-06-07 Institut Dr. Foerster Gmbh & Co. Kg Thermografisches Prüfverfahren und Prüfvorrichtung zur Durchführung des Prüfverfahrens
CN103903297B (zh) * 2012-12-27 2016-12-28 同方威视技术股份有限公司 三维数据处理和识别方法
CN104237262A (zh) * 2013-06-09 2014-12-24 北京华力兴科技发展有限责任公司 一种用于检查海关进口“三废”的设备和方法
CN103679195B (zh) * 2013-12-02 2016-08-17 北京工商大学 基于局部边缘模式的纹理图像分类方法及系统
CN105809091B (zh) * 2014-12-30 2020-10-16 同方威视技术股份有限公司 检查方法和系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1802676A (zh) * 2003-06-11 2006-07-12 昆腾磁性元件公司 乘客和行李的检查站
CN101558327A (zh) * 2006-08-31 2009-10-14 Ge本国防护股份有限公司 用于集成爆炸物检测系统的系统和方法
CN101606084A (zh) * 2007-02-05 2009-12-16 通用电气公司 检查系统和方法
US20090290757A1 (en) * 2008-05-22 2009-11-26 Mian Zahid F Inspection using three-dimensional profile information

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107463945A (zh) * 2017-07-12 2017-12-12 浙江大学 一种基于深度匹配网络的商品种类识别方法
CN107463946A (zh) * 2017-07-12 2017-12-12 浙江大学 一种结合模板匹配与深度学习的商品种类检测方法
CN107463946B (zh) * 2017-07-12 2020-06-23 浙江大学 一种结合模板匹配与深度学习的商品种类检测方法

Also Published As

Publication number Publication date
EP3040743B1 (en) 2023-11-01
CN105809091B (zh) 2020-10-16
SG11201607836WA (en) 2016-11-29
US20190156139A1 (en) 2019-05-23
US10229336B2 (en) 2019-03-12
EP3040743A1 (en) 2016-07-06
CN105809091A (zh) 2016-07-27
MY189002A (en) 2022-01-17
US20170308766A1 (en) 2017-10-26
US10509979B2 (en) 2019-12-17

Similar Documents

Publication Publication Date Title
CN105809091B (zh) 检查方法和系统
EP3349048B1 (en) Inspection devices and methods for detecting a firearm in a luggage
US10013615B2 (en) Inspection methods and devices
EP3349050B1 (en) Inspection devices and methods for detecting a firearm
Jaccard et al. Tackling the X-ray cargo inspection challenge using machine learning
WO2019052561A1 (zh) 检查方法和检查设备以及计算机可读介质
WO2017084408A1 (zh) 检查货物的方法和系统
EP3349049B1 (en) Inspection devices and methods for inspecting a container
US12417524B2 (en) Method and system of inspecting vehicle
US20170178312A1 (en) Methods, systems, and apparatuses for inspecting goods
Sun et al. Image recognition technology in texture identification of marine sediment sonar image
Lu et al. Automatic visual inspection of a missing split pin in the China railway high-speed
CN105784732B (zh) 检查方法和检查系统
Liu et al. Directional-aware automatic defect detection in high-speed railway catenary system
CN120411493A (zh) 基于先验信息的图像分层和语义分割方法
CN112581467B (zh) 一种基于疑似危险品评价的智能安检方法
CN114299337A (zh) 一种物品检测方法及检测系统
CN111582367A (zh) 一种小金属威胁检测的方法
Sribua et al. Impact of Background Color on PET Preform Defect Detection Using a YOLO Based Deep Learning Model
KR20250055850A (ko) Bic 코드 인식 방법 및 장치
HK40004246B (zh) 检查方法和检查设备以及计算机可读介质
HK40004246A (zh) 检查方法和检查设备以及计算机可读介质

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: 15875159

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 15129389

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 15875159

Country of ref document: EP

Kind code of ref document: A1