WO2020187036A1 - 一种基于购物工具的防作弊*** - Google Patents

一种基于购物工具的防作弊*** Download PDF

Info

Publication number
WO2020187036A1
WO2020187036A1 PCT/CN2020/077888 CN2020077888W WO2020187036A1 WO 2020187036 A1 WO2020187036 A1 WO 2020187036A1 CN 2020077888 W CN2020077888 W CN 2020077888W WO 2020187036 A1 WO2020187036 A1 WO 2020187036A1
Authority
WO
WIPO (PCT)
Prior art keywords
shopping
behavior
weight
product
image
Prior art date
Application number
PCT/CN2020/077888
Other languages
English (en)
French (fr)
Inventor
王海轩
王玉奎
徐步兵
Original Assignee
南京亿猫信息技术有限公司
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 南京亿猫信息技术有限公司 filed Critical 南京亿猫信息技术有限公司
Priority to KR1020217032741A priority Critical patent/KR20210135317A/ko
Publication of WO2020187036A1 publication Critical patent/WO2020187036A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • the invention relates to the field of intelligent shopping systems, in particular to an anti-cheating system based on shopping tools.
  • the Chinese invention patent “smart shopping cart, smart shopping system and its use method"
  • the application number is 201510209427.7, which discloses a preset data information comparison system based on the three conditions of weight, image, and scan code, but this The method is only based on the data judgment based on three conditions, so as to achieve the purpose of data association, but it does not set the program logic for the illegal operations that may be involved in the shopping process. Therefore, in the judgment of the shopping operation, it does not No creative design.
  • the technical scheme of the present invention is: an anti-cheating system based on shopping tools, which uses commodity image, commodity quality, commodity scan code as judgment or a logical combination of trigger conditions. Therefore, it can be expressed in hardware as: shopping tools, servers, etc.
  • Shopping tools are mainly smart shopping carts, of course, they can also be used on self-service cashiers and other devices. Such shopping carts or self-service cashier devices usually require a series of sensing devices, such as image capture devices, quality capture devices, and product code scanning devices. Therefore, based on the image sensing and weight sensing of the shopping tool on the purchased goods, the current purchase behavior of the goods is judged, and the final settlement goal is achieved by judging whether the purchase behavior is qualified.
  • the system includes: acquisition module, judgment module, alarm module, database;
  • the collection module collects the judgment conditions involved in the judgment process; such as commodity images, commodity types, commodity barcodes, and so on.
  • the judgment module is to judge data relevance, rationality of behavior, etc. based on preset logical relationships; for example: judgment based on image-related product weight and current product weight, and judgment based on the operation situation formed by the relationship between image and weight and many more;
  • the alarm module is based on the feedback of the judgment result; for example, scan code reminder for wrong operation, administrator reminder for illegal operation, etc.;
  • the database stores all kinds of preset data and the data generated during the current operation.
  • Each product has an independent storage path; for example, it stores the collected product pictures and information features, and generates data during self-learning. Data for storage and so on.
  • the acquisition module includes an image acquisition unit based on image sensing, a weight acquisition unit based on weight sensing, and a scan code acquisition unit based on scan code information.
  • the advantage of the present invention is that the system adopts a self-learning algorithm, and based on the algorithm, all kinds of data associated with commodities can perform intelligent self-learning.
  • the self-learning process is realized by the deep neural network algorithm. Based on this algorithm, the data involved in the acquisition module can become the basic feature points in the deep neural network algorithm, that is, the deep neural network algorithm can realize the self-learning of the entire system process.
  • the self-learning process of the system includes: the neural network relationship formed between the basic features obtained by the acquisition module and the correct judgment result of the system.
  • the basic layer of this neural network includes various basic feature points about each commodity.
  • the top layer of the neural network is the exact product associated with the above-mentioned basic feature points, and the other layers between the basic layer and the top layer are the feature sets composed of the basic feature points.
  • the weight adjustment process can be mutual, that is, the weight can be automatically adjusted based on the change of feature points, or the feature of the adjacent layer can be adjusted based on the weight change.
  • the adjustment direction is two-way, that is, it can be adjusted upwards or downwards.
  • image sensing is based on a deep neural network algorithm to identify the current product in weight sensing.
  • Image sensing performs product recognition based on preset image features. New features captured during the recognition process are automatically updated to the database, and the recognition process is gradually optimized. Based on this function, the result of image sensing will be more accurate.
  • the image sensing result has a corresponding discrimination standard, based on the similarity between the current recognition image and the commodity standard image, and the judgment range is adjusted according to the self-learning situation of the system.
  • the judgment value range is controlled at 70%, that is, to ensure the accuracy of image recognition as much as possible.
  • the self-learning of the system its recognition accuracy is getting higher and higher, and the value range of the judgment can be adjusted to 90% later to increase its output accuracy.
  • the image self-learning process based on the deep neural network algorithm includes: collecting basic feature points of commodities by image sensing to form the base layer of the neural network, the complete image of the corresponding commodity is the top layer of the neural network; other layers between the top layer and the base layer It is a collection composed of the basic feature points in the grassroots layer. These collections are displayed in a graphical manner; based on the changes of any layer, the two-way weight relationship between each other except the top layer is automatically modified upward or downward. .
  • the principle is to make "upstream” or “downstream” modifications to accurately achieve the top-level purpose.
  • the self-learning method is to better realize product identification and system judgment. Based on the above-mentioned image recognition result as the product confirmation criterion, the pre-stored weight information associated with the recognized product is compared with the current sensed weight of the product; the shopping behavior is determined based on the comparison result.
  • Shopping behavior judgments include: qualified behavior judgments where image sensing results match weight sensing results, and suspicious behavior judgments where image sensing results do not match weight sensing results; suspicious behaviors will trigger code scan reminders or alarms.
  • the shopping behavior includes: a shopping operation with weight sensing increment, and a return operation with weight sensing decrement.
  • the system further includes an accounting module; when the behavior performed by the customer is a qualified behavior, the corresponding accounting information of the commodity obtained by the collection module is counted by the accounting module.
  • the acquisition module is mainly based on image sensing, the product information recognized by image sensing is correct information, and the accounting module performs statistics based on the correct information; if the acquisition module is mainly based on scanning code, the product information scanned at this time The accounting information is correct information, and the accounting module performs statistics based on the correct information.
  • the judgment basis for the execution of the judgment module is based on the matching degree of the image sensor and the weight sensor. If the image sensor result and the weight sensor result do not match, the scanning code judgment process is entered. Similarly, the code scanning judgment process is also judged by the weight of the product obtained by scanning the code and the weight obtained by the weight sensor after the scanning action.
  • the current scanned product is screen-tracked, that is, the image sensor not only identifies the product in the shopping cart, but also monitors the product at the scanning position. Make sure that the current scanned product is the corresponding product that triggered the suspicious behavior in the process.
  • shopping behaviors can be divided into shopping operations and return operations based on user operations; shopping behaviors can be divided into qualified behaviors and suspicious behaviors based on system judgment standards.
  • Qualified behavior is the behavior of judging data to match, and suspicious behavior is the behavior of judging data that does not match.
  • suspicious behavior includes: abnormal behavior with image sensing, and foul behavior with image sensing. The corresponding behavior in each operation is specified as follows:
  • ⁇ Qualified behavior judgments based on a single shopping operation include:
  • ⁇ Suspicious behavior judgments based on a single shopping operation include:
  • the image acquisition unit has no results in recognition (the new image acquisition is not timely), the alarm module reminds the scan code or reminds the staff;
  • Foul behavior with image sensing the image acquisition unit recognizes correctly, the behavior is suspicious (there is entrainment behavior), the alarm module reminds the scan code or reminds the staff.
  • ⁇ Qualified behavior judgment based on a single return operation includes:
  • ⁇ Suspicious behavior judgments based on a single return operation include:
  • the alarm module reminds the scan code or reminds the staff.
  • the calculation process is clear: the data calling process is clear, and the judgment is made based on the operation of each commodity, and the information matching degree is high, which reduces the system judgment error.
  • an anti-cheating system based on shopping tools is built in an offline shopping environment and aims to provide an anti-cheating mechanism for self-service shopping.
  • the smart shopping cart as the execution carrier, a series of sensing devices are usually required on the shopping cart, including cameras, weight sensors, electronic code scanning guns, and so on.
  • the internal system includes: acquisition module, judgment module, alarm module, database. among them:
  • the collection module collects the judgment conditions involved in the judgment process; such as commodity images, commodity types, commodity barcodes, etc.; collection modules include image sensor-based image collection units, weight-sensing-based weight collection units, and code scanning Information scan code acquisition unit.
  • the judgment module is to judge data relevance, rationality of behavior, etc. based on preset logical relationships; for example: judgment based on image-related product weight and current product weight, and judgment based on the operation situation formed by the relationship between image and weight Wait, the unqualified results will be displayed through the alarm module,
  • the main function of the collection module is reflected in two aspects: First, for the collection of basic commodity data, in the system design process, the basic information of all commodities on the market will be collected, from the 360° image of the commodity, the commodity Collect data in terms of quality and so on, and store the collected data in the database for easy recall. Second, collect information about commodities in shopping behaviors, including image collection, image weight collection, and even scan code collection in a single shopping behavior.
  • the advantage of this system is that the system adopts a self-learning algorithm, based on this algorithm, all kinds of data associated with commodities can be intelligently self-learned.
  • the self-learning process of the system includes: the neural network relationship formed between the basic features obtained by the acquisition module and the correct judgment result of the system.
  • the basic level of this neural network includes various basic feature points about each commodity, neural network
  • the top layer of is the exact product that is associated with the above-mentioned basic feature points, and the other layers between the base layer and the top layer are the feature sets composed of the basic feature points.
  • the weight adjustment process can be mutual, that is, the weight can be automatically adjusted based on the change of feature points, or the feature of the adjacent layer can be adjusted based on the weight change.
  • the adjustment direction is two-way, that is, it can be adjusted upwards or downwards.
  • the recognition process is based on the above-mentioned basic features.
  • the recognition process is: a to f are the basic layers of the deep neural network, and the features of the second layer are: the combination of a+f, the combination of c+d, the combination of e+f...;
  • the features of the three layers are the combination of the features of the second layer...; the fourth layer, the fifth layer and so on; the top layer is the product.
  • the basic algorithm is implemented between adjacent two sides, and their respective weights are identified in the basic algorithm.
  • the user When the user is shopping for item A, he recognizes it through screen sensing, and the possible situations include five features from a to e are recognized, and one more feature g is added. Then based on the a to e On the premise that the five feature systems can recognize the product, the g feature will be added to the basic feature by the system, and the deep neural network will be updated based on the newly added g feature. Then the characteristic of the neural network algorithm is that the update is not only the update of the feature layer, but also based on the change of the feature layer, the weights in the basic algorithm will also be updated. The goal is to more accurately identify the A product.
  • the self-learning method is used to better realize product image recognition and screen feature system judgment. Then, when the self-learning method is used in the entire judgment process, soy milk such as the weight information of the product, the product scan code information, etc. becomes the feature points in the base layer, and the combination of the feature points forms the "imaging" upper layer feature, and so on , And finally correspond to the target product.
  • the pre-stored weight information associated with the identified commodity in image sensing is compared with the current sensed weight of the commodity; shopping behavior judgment is made based on the comparison result.
  • the shopping behavior of the product is qualified, and the accounting information of the product is recorded.
  • the shopping behavior of the product is suspicious, and the system will remind you to use the scanning code collection, based on the product related information collected by the scanning code and the weight information of the product. If it matches, the accounting information of this product is recorded. If it does not match, an alarm will alert the staff.
  • the judgment of shopping behavior in this system includes:
  • the corresponding accounting information of the commodity obtained by the collection module is counted by the accounting module. If the acquisition module is mainly based on image sensing, the product information recognized by image sensing is correct information, and the accounting module performs statistics based on the correct information; if the acquisition module is mainly based on scanning code, the product information scanned at this time The accounting information is correct information, and the accounting module performs statistics based on the correct information.
  • the accounting module will generate a shopping list after statistics. The shopping list is displayed through the mobile terminal, and the user can confirm the payment after verification.
  • Specific shopping behaviors include: shopping operations with weight sensing increments, and return operations with weight sensing decrements.
  • the judgment of shopping behavior is divided into: qualified behavior and suspicious behavior.
  • Qualified behavior is the behavior of judging data to match
  • suspicious behavior is the behavior of judging data that does not match.
  • suspicious behavior includes: abnormal behavior with image sensing, foul behavior with image sensing, and foul behavior without image sensing. The corresponding behavior in each operation is specified as follows:
  • ⁇ Qualified behavior judgments based on a single shopping operation include:
  • ⁇ Suspicious behavior judgments based on a single shopping operation include:
  • image acquisition unit recognition error image sensing error
  • alarm module reminds to scan code or remind staff
  • Foul behavior with image sensing the image acquisition unit recognizes correctly, the behavior is suspicious (the weight does not match after entrainment), the alarm module reminds the scan code or reminds the staff.
  • ⁇ Qualified behavior judgment based on a single return operation includes:
  • ⁇ Suspicious behavior judgments based on a single return operation include:
  • the alarm module reminds the scan code or reminds the staff.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Accounting & Taxation (AREA)
  • Development Economics (AREA)
  • Economics (AREA)
  • Finance (AREA)
  • Marketing (AREA)
  • Strategic Management (AREA)
  • General Business, Economics & Management (AREA)
  • Image Analysis (AREA)
  • Burglar Alarm Systems (AREA)
  • Cash Registers Or Receiving Machines (AREA)

Abstract

本发明公开了一种基于购物工具的防作弊***，基于购物车对商品的：图像传感、重量传感来判断商品当前的购买行为；根据数据库中各商品预存数据，图像传感基于深度神经网络算法来识别当前重量传感中的商品；以图像识别为商品确认基准，被识别商品所关联的预存重量信息与该商品的当前传感重量来比对；基于比对结果进行购物行为判断和***自学习。购物行为的判断包括：图像传感结果与重量传感结果匹配的合格行为判断、图像传感结果与重量传感结果不匹配的可疑行为判断；***自学习过程包括：采集模块获得的基础特征与***正确判断结果之间形成的神经式网络关系，基于网络关系中单层数据的更新，相邻关系层通过对彼此间的双向权重的调整进行整体更新。

Description

一种基于购物工具的防作弊*** 技术领域
本发明涉及智能购物***领域,特别是一种基于购物工具的防作弊***。
背景技术
目前我国各种大型购物超市商品种类数以万计，伴随超市规模的扩大和消费者购物选择多元化，会遇到诸如无法快捷地在超市里准确找到想要购买的商品、对同类商品不同品牌的商品信息了解不充分、对智能购物车内购买商品的数量和价格明细不了解、购物高峰时期在交款台边排很长的队等待结算等等问题。
因此，研究开发集自助扫码结账和实时监控偷盗行为于一体的智能购物***，具有广阔的市场前景和重大的商业价值。国内外已有很多类似的研究，沃尔玛先在零售业研发了RFID技术，并研发了基于RFID的智能购物车并在部分沃尔玛超市试用，随后大连理工大学、上海交通大学等研究了基于RFID的购物智能终端，该研究使用了RFID标签取代多维条形码，该标签不仅包含了商品各项信息，而且有很好的防伪技术、加密技术和信息控制，然而RFID标签目前还无法长时间取代普通的二维码，故现如今只有沃尔玛极个别超市在推广应用。另外还有中南大学、信息科学与工程学院、电子科技大学等研究了基于嵌入式***的智能购物车，该方案结合普通的条形码和嵌入式操作***的智能终端，实现了室内导航、条码扫描、管理购物列表、生成付款码等功能，但该方案没有涉及到购物车的防盗功能，因此也无法广泛使用。但东北林业大学的刘兵等人根据电子商品防盗***即EAS***对购物车商品进行防盗，利用每个商品的软标签，在商品结算后才可以用EAS***对软标签进行解码，从而实现防盗，这与RFID标签类似，需要对每个商品都做不同的软标签，工作量较大，不能广泛使用。
例如中国发明专利《智能购物车、智能购物***及其使用方法》，申请号为201510209427.7，其公开了一种基于重量、图像、扫码三个条件进行预设数据信息比对的***，但是该方法仅仅是基于三个条件进行的数据判断，从而达到数据关联的目的，但是其并没有针对购物过程中可能涉及到的违规操作进行程序逻辑设定，因此，在对购物操作判断上，其并没有创造性的设计。
再例如中国发明专利《超市购物车监控***及其监控方法》，申请号为201610414195.3，其公开了一种基于重量、图像、扫码三个条件对应至购物操作的逻辑 判断法方法，其图像传感是根据购物操作手势进行的信息搜集，判定的过程也需要对画面上的手势进行分析等操作。因此，其判定方法必然繁琐，而且基于手势识别的问题，其判定精度也会有所影响，总体上达不到高精度、高效率的计算结果。
发明内容
本发明的技术方案是：一种基于购物工具的防作弊***，以商品图像、商品质量、商品扫码为判定、或触发条件进行的逻辑组合。因此，在硬件上可以表现为：购物工具、服务器等。购物工具则是以智能购物车为主，当然也可以用于自助收银台等设备上。此类的购物车或自助收银设备上通常需要具备一系列传感设备，例如：图像采集设备、质量采集设备、商品扫码设备等。因此，基于购物工具对所购商品的：图像传感、重量传感来判断商品当前的购买行为，根据该购买行为是否合格的判断达到最终结算的目的。
软件上，***包括：采集模块、判断模块、警报模块、数据库；
采集模块是对判断过程中所涉及的判定条件收集；例如商品图像、商品种类、商品条码等等。
判断模块则是基于预设的逻辑关系进行数据关联性、行为合理性等等的判断；例如：基于图像关联的商品重量与当前商品重量的判断、基于图像与重量关系所形成的操作情况的判断等等；
警报模块则是基于判断结果进行的反馈；例如，对误操作进行扫码的提醒，对违规操作进行管理员提醒等等；
数据库则是对预设的各类数据和当前运行时生成的数据进行存储，每件商品都有独立的存储路径；例如，对实现收集的商品图片、信息特征的存储、对自学习过程中生成的数据进行存储等等。
优选的是，采集模块包括基于图像传感的图像采集单元、基于重量传感的重量采集单元、基于扫码信息的扫码采集单元。
本发明的优势在于：***采用了一种自学习的算法，基于该算法，与商品关联的各类数据都可以进行智能的自学习。自学习的过程是由深度神经网络算法来实现的，基于该算法，采集模块中所涉及的数据都可成为深度神经网络算法中的基础特征点，即深度神经网络算法可以实现整个***的自学习过程。
优选的是，***自学习的过程包括：采集模块获得的基础特征与***的正确判断结果之间形成的神经式网络关系，这个神经式网络的基层上包括关于每件商品的各类基础 特征点，神经式网络的顶层为对应上述基础特征点来关联的确切该商品，基层和顶层之间的其他层为基础特征点相互组合成的特征集合。当该网络关系中某一单层数据的更新，即所说的对单一一层进行训练，那么除却顶层，其他相邻关系层通过对彼此间的双向权重的调整进行整体更新。在该过程的特点包括两方面：第一、权重的调整过程可以是相互的，即可以基于特征点的变化对权重自动调整，也可以基于权重变化对相邻层的特征进行调整。第二、调整方向是双向的，即可以上行调整，也可以下行调整。
以图像传感为示例：根据数据库中各商品预存数据，图像传感基于深度神经网络算法来识别当前重量传感中的商品。图像传感基于预设的图像特征进行商品识别，在识别过程中所捕捉的新特征会自动更新到数据库，并逐步优化识别过程。基于该功能，图像传感的结果将会更加精准。
优选的是，图像传感结果具有相应的判别标准，根据当前识别图像与商品标准图像的相似度为依据，判断的值域根据***自学习情况进行调整。***在使用之初，因自学习数据较少，判断的值域控制在70％，即尽可能确保图像识别准确率。随着***自学习，其识别精度越来越高，后期可以将判断的值域调整至90％，以增加其输的出准确率。
优选的是，基于深度神经网络算法的图像自学习过程包括：以图像传感采集商品基础特征点构成神经网络的基层，对应商品的完整图像为神经网络的顶层；顶层与基层之间的其他层是以基层中基础特征点组合成的集合，这些集合以图形化方式表现查出来；基于任意层的变化，除顶层的其他网络层之间对相互间的双向权重关系进行上行或下行的自动修改。原则以：准确的实现顶层目的而进行“上游”或“下游”的修改。
综上，该自学习的方法是为了更好的实现商品识别以及***判断。基于上述的图像识别结果为商品确认基准，被识别商品所关联的预存重量信息与该商品的当前传感重量来比对；基于比对结果进行购物行为判断。
购物行为的判断包括：图像传感结果与重量传感结果匹配的合格行为判断、图像传感结果与重量传感结果不匹配的可疑行为判断；可疑行为将触发扫码提醒或警报。
优选的是，购物行为包括：重量传感增量的购物操作、重量传感减量的退货操作。
由于行为的触发条件不同，因此其对应的运行结果也不同，但是其判断原理是基于上述方式实现的。
基于上述方案，本发明中实现完整的购物流程还需要其他功能的模块进行配合。
优选的是，***还包括核算模块；当客户实施的行为是合格行为时，采集模块所得的商品的对应核算信息由核算模块统计。若采集模块是以图像传感为主，此时图像传感 识别出的商品信息为正确信息，核算模块基于该正确信息进行统计；若采集模块以扫码采集为主，此时所扫商品的核算信息为正确信息，核算模块基于该正确信息进行统计。
判断模块的实行的判断依据以图片传感和重量传感的匹配程度为基准，若图片传感结果和重量传感结果不匹配，则进入扫码判断过程。同样，扫码判断过程也是以扫码获得的商品的重量与扫码行为后重量传感获得的重量来判断。
在扫码过程中对当前扫码商品今进行画面追踪，即图像传感不仅对购物车内的商品进行识别，同时对扫码位置上的商品进行画面监控。确保当前扫码商品为该过程中触发可疑行为的对应商品。
购物行为可以根据用户操作分为：购物操作和退货操作；购物行为根据***判断标准分为：合格行为和可疑行为。合格行为即判断数据向匹配的行为，可疑行为即判断数据不匹配的行为。一般，可疑行为包括：有图像传感的异常行为、有图像传感的犯规行为。每个操作中的对应行为具体规定如下：
●基于单个购物操作的合格行为判断包括：
有图像传感的正常行为：画面传感结果与重量传感结果匹配，核算模块添加该商品的核算信息；
●基于单个购物操作的可疑行为判断包括：
有图像传感的异常行为：图像采集单元识别无结果(新款图像采集不及时)，警报模块提醒扫码或提醒工作人员；
有图像传感的犯规行为：图像采集单元识别正确，行为可疑(存在夹带行为)，警报模块提醒扫码或提醒工作人员。
●基于单个退货操作的合格行为判断包括：
有图像传感的正常行为：图像传感结果与重量传感结果匹配，核算模块删除该商品的核算信息；
●基于单个退货操作的可疑行为判断包括：
有图像传感的异常行为：图像传感结果与重量传感结果不匹配，警报模块提醒扫码或提醒工作人员。
本发明的优点是：
1、判断的准确度高：基于图像识别技术，图像识别准确率能够达到90％，关联后台信息后进行其他数据判断，以判断条件的匹配度来达到防损目的。
2、运算过程清晰：数据调用过程明了，基于每件商品的操作进行判断， 信息匹配度高，减少了***判断误差。
3、自学习能力突出：基于深度神经网络算法，以基础数据的更新进行自训练，能够提高***的识别精度和判断准度。
附图说明
下面结合附图及实施例对本发明作进一步描述：
图1购物操作的流程图；
图2退货操作的流程图。
具体实施方式
实施例1：
如附图所示，一种基于购物工具的防作弊***，构架于线下购物环境中，主旨在于提供一种自助购物的防作弊机制。以智能购物车为执行载体，在购物车上通常需要具备一系列传感设备，包括摄像头、重量传感器、电子扫码枪等等。内部***则包括了：采集模块、判断模块、警报模块、数据库。其中：
采集模块是对判断过程中所涉及的判定条件收集；例如商品图像、商品种类、商品条码等等；采集模块包括基于图像传感的图像采集单元、基于重量传感的重量采集单元、基于扫码信息的扫码采集单元。
判断模块则是基于预设的逻辑关系进行数据关联性、行为合理性等等的判断；例如：基于图像关联的商品重量与当前商品重量的判断、基于图像与重量关系所形成的操作情况的判断等等，对于不合格的结果将会通过警报模块表现出来，
采集模块的主要作用体现在了两个方面：第一、对于商品基础数据采集，在***设计过程中，就会对目前市场上的所有商品进行基础信息的收集，从商品360°的图像、商品质量等等方面采集数据，并将采集到的数据存入数据库中便于调用。第二、对购物行为中，商品的信息进行采集，包括单个购物行为中，商品的图像收集、图像重量收集、甚至是扫码收集等等。
该***的优势在于：***采用了一种自学习的算法，基于该算法，与商品关联的各类数据都可以进行智能的自学习。***自学习的过程包括：采集模块获得的基础特征与***的正确判断结果之间形成的神经式网络关系，这个神经式网络的基层上包括关于每件商品的各类基础特征点，神经式网络的顶层为对应上述基础特征点来关联的确切该商 品，基层和顶层之间的其他层为基础特征点相互组合成的特征集合。当该网络关系中某一单层数据的更新，即所说的对单一一层进行训练，那么除却顶层，其他相邻关系层通过对彼此间的双向权重的调整进行整体更新。在该过程的特点包括两方面：第一、权重的调整过程可以是相互的，即可以基于特征点的变化对权重自动调整，也可以基于权重变化对相邻层的特征进行调整。第二、调整方向是双向的，即可以上行调整，也可以下行调整。
以图像传感为示例：A商品在基础数据采集时，收集到了6个画面基础特征：a处商标、b处形状、c处颜色、d处文字、e处条纹、f处图案。以上述的基础特征进行识别，识别过程是：a至f为深度神经网络的基层，第二层的特征为：a+f的组合、c+d的组合、e+f的组合……；第三层的特征为第二层特征的相互组合……；第四层、第五层以此类推；顶层即为该商品。相邻的两侧之间通过基础算法来实现，基础算法中标识了各自权重。当用户在对A商品进行购物操作时，通过画面传感进行了识别，那么可能发生的情况中包括了a至e的5处特征被识别，同时多了一个特征g，那么基于a至e这5个特征***已经可以识别该商品的前提下，g特征将被***加入基础特征中，同时深度神经网络将会基于该新加入的g特征进行更新。那么神经网络算法的特点就在于更新不仅仅是对特征层的更新，同时基于特征层的变化，基础算法中的权重也将会被更新，其目标都是为了更加准确的识别A商品。
因此，使用该自学习的方法是为了更好的实现商品图像识别以及画面特征***判断。那么，当整个判断过程都使用该自学习方法时，商品的重量信息、商品的扫码信息等等豆浆成为基层中的特征点，通过特征点的组合形成“图像化”的上层特征，依次类推，最终对应至目标商品。
实施例2：
基于商品的图像传感和重量传感为基础条件，将图像传感的被识别商品所关联的预存重量信息与该商品的当前传感重量来比对；基于比对结果进行购物行为判断。
若该图像传感的结果与重量传感的结果匹配，那么该商品的购物行为合格，且此商品的核算信息被记录。
若该图像传感的结果与重量传感的结果不匹配，那么该商品的购物行为可疑，***将会提醒使用扫码采集，基于扫码采集到的商品关联信息与该商品的重量信息判断，若匹配则此商品的核算信息被记录，若不匹配则警报提醒工作人员。
因此，基于上述的判断原理，本***中购物行为的判断包括：
图像传感结果与重量传感结果匹配的合格行为判断、图像传感结果与重量传感结果不匹配的可疑行为判断；可疑行为将触发扫码提醒或警报。
当客户实施的行为是合格行为时，采集模块所得的商品的对应核算信息由核算模块统计。若采集模块是以图像传感为主，此时图像传感识别出的商品信息为正确信息，核算模块基于该正确信息进行统计；若采集模块以扫码采集为主，此时所扫商品的核算信息为正确信息，核算模块基于该正确信息进行统计。核算模块统计后将会生成购物清单，购物清单通过移动终端来显示，用户可以经过核对后确认支付。
实施例3：
具体的购物行为包括：重量传感增量的购物操作、重量传感减量的退货操作。
购物行为的判断根据***标准分为：合格行为和可疑行为。合格行为即判断数据向匹配的行为，可疑行为即判断数据不匹配的行为。一般，可疑行为包括：有图像传感的异常行为、有图像传感的犯规行为、无图像传感的犯规行为。每个操作中的对应行为具体规定如下：
●基于单个购物操作的合格行为判断包括：
有图像传感的正常行为：画面传感结果与重量传感结果匹配，核算模块添加该商品的核算信息。
●基于单个购物操作的可疑行为判断包括：
有图像传感的异常行为：图像采集单元识别错误，图像传感错误，警报模块提醒扫码或提醒工作人员；
有图像传感的犯规行为：图像采集单元识别正确，行为可疑(夹带后引起重量不匹配)，警报模块提醒扫码或提醒工作人员。
●基于单个退货操作的合格行为判断包括：
有图像传感的正常行为：图像传感结果与重量传感结果匹配，核算模块删除该商品的核算信息。
●基于单个退货操作的可疑行为判断包括：
有图像传感的异常行为：图像传感结果与重量传感结果不匹配，警报模块提醒扫码或提醒工作人员。
上述实施例仅例示性说明本发明的原理及其功效，而非用于限制本发明的。任何熟悉此技术的人士皆可在不违背本发明的精神及范畴下，对上述实施例进行修饰或改变。因此，举凡所属技术领域中具有通常知识者在未脱离本发明的所揭示的精神与技术思想下所完成的一 切等效修饰或改变，仍应由本发明的权利要求所涵盖。

Claims (14)

  1. 一种基于购物工具的防作弊***，基于购物工具对商品的：图像传感、重量传感来判断商品当前的购买行为；***包括：采集模块、判断模块、警报模块；其特征在于：根据数据库中各商品预存数据，图像传感基于深度神经网络算法来识别当前重量传感中的商品；以图像识别为商品确认基准，被识别商品所关联的预存重量信息与该商品的当前传感重量来比对；基于比对结果进行购物行为判断和***自学习；
    购物行为的判断包括：图像传感结果与重量传感结果匹配的合格行为判断、图像传感结果与重量传感结果不匹配的可疑行为判断；所述可疑行为将触发扫码提醒或警报；
    ***自学习的过程包括：采集模块获得的基础特征与***的正确判断结果之间形成的神经式网络关系，基于网络关系中单层数据的更新，相邻关系层通过对彼此间的双向权重的调整进行整体更新。
  2. 根据权利要求1所述的一种基于购物工具的防作弊***，其特征在于：所述采集模块包括基于图像传感的图像采集单元、基于重量传感的重量采集单元。
  3. 根据权利要求1或2所述的一种基于购物工具的防作弊***，其特征在于：***还包括核算模块；合格行为时采集模块所得的商品的对应核算信息由所述核算模块统计。
  4. 根据权利要求3所述的一种基于购物工具的防作弊***，其特征在于：所述采集模块还包括扫码采集单元，若扫码结果判断合格，所扫商品的核算信息由所述核算模块统计。
  5. 根据权利要求4所述的一种基于购物工具的防作弊***，其特征在于：扫码结果判断包括：扫码获得的商品的重量与扫码行为后重量传感获得的重量判断。
  6. 根据权利要求5所述的一种基于购物工具的防作弊***，其特征在于：基于深度神经网络算法的图像自学习过程包括：以图像传感采集商品基础特征点构成神经网络的基层，对应商品的完整图像为神经网络的顶层；基于任意层的变化，除顶层的其他网络层之间对相互间的双向权重关系进行上行或下行的自动修改。
  7. 根据权利要求6所述的一种基于购物工具的防作弊***，其特征在于：***还包括所述数据库，每款商品的具有独立的数据存储路径。
  8. 根据权利要求7所述的一种基于购物工具的防作弊***，其特征在于：所述可疑行为包括：有图像传感的犯规行为。
  9. 根据权利要求8所述的一种基于购物工具的防作弊***，其特征在于：购物行为包括：重量传感增量的购物操作。
  10. 根据权利要求9所述的一种基于购物工具的防作弊***，其特征在于：购物行为包括：基于单个购物操作的合格行为判断包括：
    有图像传感的正常行为：画面传感结果与重量传感结果匹配，核算模块添加该商品的核算信息。
  11. 根据权利要求10所述的一种基于购物工具的防作弊***，其特征在于：基于单个购物操作的可疑行为判断包括：
    有图像传感的犯规行为：图像采集单元识别正确，行为可疑，警报模块提醒扫码或提醒工作人员。
  12. 根据权利要求1或11所述的一种基于购物工具的防作弊***，其特征在于：购物行为包括：重量传感减量的退货操作。
  13. 根据权利要求12所述的一种基于购物工具的防作弊***，其特征在于：基于单个退货操作的合格行为判断包括：
    有图像传感的正常行为：图像传感结果与重量传感结果匹配，核算模块删除该商品的核算信息。
  14. 根据权利要求13所述的一种基于购物工具的防作弊***，其特征在于：基于单个退货操作的可疑行为判断包括：
    有图像传感的异常行为：图像传感结果与重量传感结果不匹配，警报模块提醒扫码或提醒工作人员。
PCT/CN2020/077888 2019-03-15 2020-03-05 一种基于购物工具的防作弊*** WO2020187036A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
KR1020217032741A KR20210135317A (ko) 2019-03-15 2020-03-05 쇼핑 도구에 기반한 부정 행위 방지 시스템

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910196940.5 2019-03-15
CN201910196940.5A CN109872168A (zh) 2019-03-15 2019-03-15 一种基于购物工具的防作弊***

Publications (1)

Publication Number Publication Date
WO2020187036A1 true WO2020187036A1 (zh) 2020-09-24

Family

ID=66920508

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/077888 WO2020187036A1 (zh) 2019-03-15 2020-03-05 一种基于购物工具的防作弊***

Country Status (3)

Country Link
KR (1) KR20210135317A (zh)
CN (1) CN109872168A (zh)
WO (1) WO2020187036A1 (zh)

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109872168A (zh) * 2019-03-15 2019-06-11 南京亿猫信息技术有限公司 一种基于购物工具的防作弊***
CN113593148A (zh) * 2021-06-24 2021-11-02 皮卫华 一种基于智能购物车防作弊方法及***
KR102419220B1 (ko) * 2021-12-21 2022-07-08 주식회사 인피닉 소량 구매를 위한 자동 결제 방법 및 이를 위한 시스템
KR102419222B1 (ko) * 2021-12-21 2022-07-08 주식회사 인피닉 상품 결제를 위한 고객 추적 방법 및 이를 위한 시스템
KR102525234B1 (ko) * 2022-09-07 2023-04-26 농업협동조합중앙회 매장 시스템에서 사용자 행위 분석을 위한 서버 및 그 방법

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107610379A (zh) * 2017-09-11 2018-01-19 江苏弘冠智能科技有限公司 一种购物识别方法及购物车识别装置
CN108537994A (zh) * 2018-03-12 2018-09-14 深兰科技（上海）有限公司 基于视觉识别及重量感应技术的智能商品结算***及方法
CN109872168A (zh) * 2019-03-15 2019-06-11 南京亿猫信息技术有限公司 一种基于购物工具的防作弊***

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8494909B2 (en) * 2009-02-09 2013-07-23 Datalogic ADC, Inc. Automatic learning in a merchandise checkout system with visual recognition
US9396622B2 (en) * 2012-11-02 2016-07-19 Tyco Fire & Security Gmbh Electronic article surveillance tagged item validation prior to deactivation
CN106408369B (zh) * 2016-08-26 2021-04-06 西安超嗨网络科技有限公司 一种智能鉴别购物车内商品信息的方法
CN108269369A (zh) * 2017-09-27 2018-07-10 中山市宾哥网络科技有限公司 结算箱及其结算方法
CN109003403B (zh) * 2018-06-05 2021-09-14 西安超嗨网络科技有限公司 购物车商品称重和计价方法、装置及***
CN109214806B (zh) * 2018-11-20 2022-01-04 北京京东尚科信息技术有限公司 自助结算方法、装置以及存储介质

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107610379A (zh) * 2017-09-11 2018-01-19 江苏弘冠智能科技有限公司 一种购物识别方法及购物车识别装置
CN108537994A (zh) * 2018-03-12 2018-09-14 深兰科技（上海）有限公司 基于视觉识别及重量感应技术的智能商品结算***及方法
CN109872168A (zh) * 2019-03-15 2019-06-11 南京亿猫信息技术有限公司 一种基于购物工具的防作弊***

Also Published As

Publication number Publication date
CN109872168A (zh) 2019-06-11
KR20210135317A (ko) 2021-11-12

Similar Documents

Publication Publication Date Title
WO2020187036A1 (zh) 一种基于购物工具的防作弊***
CN109089082B (zh) 一种基于热特性图像的图像采集***
CN106408369B (zh) 一种智能鉴别购物车内商品信息的方法
US7780081B1 (en) System and method for security protection, inventory tracking and automated shopping cart checkout
CN108320404A (zh) 基于神经网络的商品识别方法、装置、自助收银台
US20190318417A1 (en) Method and system associated with a smart shopping apparatus
US9396622B2 (en) Electronic article surveillance tagged item validation prior to deactivation
CN107845212B (zh) 基于身份识别的智能购物***
WO2019062018A1 (zh) 商品自动结算方法、装置、自助收银台
CN107103503A (zh) 一种订单信息确定方法和装置
CN107103502A (zh) 一种订单信息确定方法和装置
CN107093269A (zh) 一种智能无人售货***及方法
CN105915857A (zh) 超市购物车监控***及其监控方法
US10503961B2 (en) Object recognition for bottom of basket detection using neural network
WO2019062812A1 (zh) 用于自动结算的人机交互装置及其应用
CN110197240A (zh) 基于图像识别的无人称重式售货柜的控制方法
AU2019100428A4 (en) An intelligent in-store shopping platform for customers and retailers. With this, customers can select, scan, and pay for the products via smartphones and check-out of the store with minimal human intervention. The system uses hi-end technologies such as artificial intelligence for anti-shoplifting, automated decision making, Computer Vision, weighing techniques, electronic circuitry and RFID. The framework uses intricate IoT (Internet of Things) technology and self-learning algorithms, big data analytics, customer engagement and pattern analysis using data extraction and knowledge mining.
CN111222870B (zh) 结算方法、装置和***
US11741447B1 (en) Automated purchasing systems and methods
CN114119001B (zh) 基于人工智能的商品销售货款收银智能管理分析平台
CN106652196A (zh) 工业易耗品存取装置
CN109993891A (zh) 称重式智能货柜的控制方法
CN110197560A (zh) 基于称重模块的自助售货***
TW202032468A (zh) 無人零售設備和退款方法及其設備
US20230147769A1 (en) Verification of items placed in physical shopping cart

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: 20774748

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 20217032741

Country of ref document: KR

Kind code of ref document: A

122 Ep: pct application non-entry in european phase

Ref document number: 20774748

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 20774748

Country of ref document: EP

Kind code of ref document: A1