WO2023070843A1 - 多模态货币识别系统、方法、存储介质和终端 - Google Patents

多模态货币识别系统、方法、存储介质和终端 Download PDF

Info

Publication number
WO2023070843A1
WO2023070843A1 PCT/CN2021/136410 CN2021136410W WO2023070843A1 WO 2023070843 A1 WO2023070843 A1 WO 2023070843A1 CN 2021136410 W CN2021136410 W CN 2021136410W WO 2023070843 A1 WO2023070843 A1 WO 2023070843A1
Authority
WO
WIPO (PCT)
Prior art keywords
currency
recognition
features
module
image
Prior art date
Application number
PCT/CN2021/136410
Other languages
English (en)
French (fr)
Inventor
王晓亮
陈�光
Original Assignee
广州广电运通金融电子股份有限公司
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 广州广电运通金融电子股份有限公司 filed Critical 广州广电运通金融电子股份有限公司
Publication of WO2023070843A1 publication Critical patent/WO2023070843A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/15—Correlation function computation including computation of convolution operations
    • G06F17/153—Multidimensional correlation or convolution
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent

Definitions

  • the invention belongs to the technical field of currency recognition, and in particular relates to a multimodal currency recognition system, method, storage medium and terminal based on texture perception and open set recognition.
  • the usual banknote recognition method often adopts a closed-set recognition method, that is, the ID and face value of the currency to be recognized are known, and the corresponding ID and face value are also used during training.
  • this closed-set recognition method is unable to deal with Currency with new currency ID and new denomination.
  • the purpose of the present invention is to provide a multi-modal currency identification system, method, storage medium and terminal, which can solve the above problems.
  • a multi-modal currency recognition system based on texture perception and open-set recognition includes a telecommunications-connected image acquisition module, an image processing module, a currency identification module, and a result output module, wherein the image acquisition module uses a multi-modal sensor , for generating a multimodal feature map of the currency to be identified; the image processing module includes an image preprocessing unit and a feature extraction unit, and the image preprocessing unit performs unified preprocessing on the received multimodal feature map, The feature extraction unit collects the global features of the image through the backbone network of the image processing algorithm, and collects the local features of the image through the texture perception module of the image processing algorithm; the currency recognition module includes a storage unit and a comparison judgment unit, wherein the The storage unit stores a template library and a test library, and the comparison and determination unit compares the received feature information with the currency characteristics of the template library to determine the classification and face value of the currency to be tested; the result output module displays the currency identification module through the display unit. The recognition results are visualized.
  • the multi-modal sensor includes a visible light sensor, an infrared light sensor and an ultraviolet light sensor, which are used to collect multi-modal feature maps of the front and back sides of the currency.
  • the backbone network of the image processing module adopts a 10-channel recognition neural network, and the received image features include visible light RGB three-channel image features, infrared light image features and ultraviolet light image features corresponding to the front and back of the currency, a total of 10 channels image feature information.
  • the texture perception module of the image processing module is composed of:
  • n ⁇ [1,2,3,4] represents the branch number
  • Formula (1) represents the convolution operation and normalization operation with a convolution kernel size of (1,2n-1);
  • Formula (2) represents the convolution operation and normalization operation with a convolution kernel size of (2n-1,1);
  • Formula (3) represents the convolution operation and normalization operation with a convolution kernel size of (3,3) and an expansion rate of 2n-1;
  • Formula (4) indicates that the input feature X is cut into x n ;
  • Equation (5) represents the branch network in the texture perception module
  • Equation (6) represents a texture-aware module that introduces a residual mechanism.
  • the algorithm model parameters corresponding to the image processing module include number of iterations, learning rate, loss function and training optimization function, wherein the loss function adopts Triplet loss function, the training optimization function adopts SGD method, and the model is updated by backpropagation internal parameters.
  • the currency identification module adopts an open set identification algorithm including a cosine similarity calculation unit, a comparison unit and a threshold determination unit, the cosine similarity calculation unit is used to calculate and sort the cosine similarity of the currency features to be tested, and the comparison unit It is used to extract the highest cosine similarity score and match it with template tags in the template library, and the threshold determination unit is used to set a threshold and extract valid matching results.
  • the present invention also provides a currency identification method, the method comprising:
  • S1 data collection, collect different currency IDs, different denominations and real or fake currency data sets through multi-modal sensors, so each currency corresponds to multi-modal images under visible light, ultraviolet light, and red light;
  • Model application collecting data on the currency to be tested, importing the collected multi-modal features into the currency recognition model, and visually presenting the recognition results.
  • model application in step S5 includes:
  • S53 Collect multimodal features of the currency to be detected through a multimodal sensor, and input the multimodal features into an algorithm to obtain identification features;
  • the discriminant features are compared with the features in the template library, and the template label in the template library corresponding to the maximum score is the prediction result of the test currency;
  • Steps S53-S55 are repeated to update and manage the currency identification module.
  • the present invention also provides a computer-readable storage medium, on which computer instructions are stored, and the steps of the aforementioned method are executed when the computer instructions are executed.
  • the present invention also provides a terminal, including a memory and a processor, the memory stores a template library and computer instructions that can run on the processor, and the processor executes the aforementioned method when running the computer instructions step.
  • the beneficial effect of the present invention lies in that the system and method of the present application adopt texture perception and open set recognition technology, and through multi-modal feature fusion, the recognition ability of depth features is improved, and the multi-currency of texture perception
  • the recognition algorithm improves the recognition performance.
  • the open-set recognition algorithm can realize the recognition of new currency and new denomination banknotes with a small number of samples. The overall development cost is low and the recognition accuracy is high.
  • Figure 1 is a schematic diagram of an existing recognition algorithm
  • Fig. 2 is a schematic diagram of a multimodal currency identification system
  • Fig. 3 is a schematic diagram of multimodal feature fusion
  • Fig. 4 is a texture perception algorithm process diagram
  • Fig. 5 is a structural diagram of a texture perception module
  • Fig. 6 is a process diagram of an open set recognition algorithm based on texture perception
  • Fig. 7 is a flow chart of the currency identification method.
  • a multi-modal currency recognition system based on texture perception and open set recognition see Figure 2, the system includes a telecommunications-connected image acquisition module, an image processing module, a currency recognition module and a result output module.
  • the image acquisition module adopts a multimodal sensor to generate a multimodal feature map of the currency to be identified.
  • the multimodal sensor includes a visible light sensor, an infrared light sensor and an ultraviolet light sensor, which are used to collect multimodal characteristic maps of the front and back sides of the currency.
  • the image processing module includes an image preprocessing unit and a feature extraction unit, the image preprocessing unit performs unified preprocessing on the received multimodal feature map, and the feature extraction unit passes through the backbone network of the image processing algorithm The global features of the image are collected, and the local features of the image are collected through the texture perception module of the image processing algorithm.
  • the backbone network of the image processing module adopts a 10-channel recognition neural network, and the received image features include visible light RGB three-channel image features, infrared light image features and ultraviolet light image features corresponding to the front and back of the currency, a total of 10 channels image feature information.
  • multimodal feature fusion After currency passes through multimodal sensors (including visible light sensors, infrared light sensors, and ultraviolet light sensors), it can obtain multimodal feature maps of visible light, ultraviolet light, and infrared light regions (currency vs. Ultraviolet light and infrared light produce fluorescence reflection, so the multimodal characteristics of ultraviolet light and infrared light have a certain ability to distinguish true and false).
  • the multimodal feature map uses feature cascading (splicing) for feature fusion, as shown in Figure 3, a schematic diagram of multimodal feature fusion.
  • the fusion of ultraviolet light, visible light, and infrared light multi-modal feature maps can be obtained with 10 channels (five channels on the front and back sides, and the five channels include: visible light RGB three-channel image, one channel image)
  • the multimodal feature map is extracted by common convolution method after being input into the network.
  • the convolution kernel is extracted by weight sharing.
  • the features of each mode in the same receptive field area of different channels are then integrated, that is, the multi-modal features of ultraviolet and infrared light are mapped to the depth features.
  • the introduction of multi-modal feature maps of ultraviolet and infrared light can improve the ability of the algorithm to distinguish true from false.
  • this application designs a recognition algorithm based on texture perception, specifically the embedding of the aforementioned texture perception module.
  • the conventional recognition algorithms used in daily life often use deep neural networks with a large number of parameters such as VGG and ResNet.
  • CNN in the convolutional neural network often obtains global features, so it lacks the ability to mine local features.
  • texture features one of the local features
  • this application mines the texture features of currency by designing a texture perception module to improve the recognition performance of multiple currencies.
  • the process diagram of the texture perception algorithm is shown in Figure 4.
  • the texture perception module in the human visual system, a set of receptive fields of different sizes helps to highlight the area close to the fovea, which is very sensitive to small spatial changes (such as texture changes), which is beneficial to Target Recognition. Therefore, this application simulates human visual senses through convolution with different convolution kernel sizes, and designs a texture perception module.
  • this application simulates human visual senses through convolution with different convolution kernel sizes, and designs a texture perception module.
  • high-dimensional mapping is not conducive to feature expression and although there are subtle differences in the features carried by different channels in the same feature layer, the statistical characteristics between multi-channel features are consistent, so this paper aims to ease the transition from high-dimensional features to low-latitude features.
  • the texture perception module is composed of:
  • n ⁇ [1,2,3,4] represents the branch number
  • formula (1) represents the convolution operation and normalization operation with the convolution kernel size (1,2n-1)
  • formula (2) represents The convolution operation and normalization operation with a convolution kernel size of (2n-1,1)
  • the formula (3) indicates that the convolution kernel size is (3,3)
  • the expansion rate is 2n-1 convolution operation and normalization operation One operation.
  • the texture perception module can be expressed as:
  • the texture perception module corresponding to the above algorithm process can also refer to FIG. 5 .
  • the algorithm model parameters corresponding to the image processing module include number of iterations, learning rate, loss function and training optimization function, wherein the loss function adopts Triplet loss function, the training optimization function adopts SGD method, and the model is updated by backpropagation internal parameters.
  • the currency identification module includes a storage unit and a comparison and determination unit, wherein a template library and a test library are stored in the storage unit, and the comparison and determination unit compares the received characteristic information with the currency characteristics of the template library, and determines whether Classification and denomination of currencies.
  • the currency identification module adopts an open set identification algorithm including a cosine similarity calculation unit, a comparison unit and a threshold determination unit, the cosine similarity calculation unit is used to calculate and sort the cosine similarity of the currency features to be tested, and the comparison unit It is used to extract the highest cosine similarity score and match it with template tags in the template library, and the threshold determination unit is used to set a threshold and extract valid matching results.
  • the currency recognition module adopts an open-set recognition solution as a whole.
  • the usual banknote recognition method often adopts a closed-set recognition method, that is, the ID and face value of the currency to be recognized are known, and the corresponding ID and face value are also used during training.
  • this method of closed-set identification cannot cope with new currency IDs and currencies with new denominations. Therefore, this paper adopts the method of open set identification, that is, the currency ID to be identified can allow the existence of unseen IDs or denominations.
  • the trained neural network algorithm is only used as a feature extraction network and does not participate in classification and recognition. See Figure 6.
  • Each currency in the test set is compared with the currency in the template library, and the comparison score is the largest.
  • the template label in the template library corresponding to the score is the prediction result of the currency to be recognized.
  • the result output module visually presents the recognition result of the currency recognition module through the display unit.
  • the display unit adopts a display such as a liquid crystal screen, an LED screen, etc., and receives and displays the prediction result of the currency identification module.
  • a currency identification method based on the system described in the first embodiment, the method includes:
  • each currency corresponds to multi-modal images under visible light, ultraviolet light, and red light.
  • a currency recognition model with 10 input channels and a texture-aware module is obtained through data input, model construction and model parameter setting.
  • this article uses multi-modal image input, so when inputting the image, it is not only the currency image in the visible light area that is input to the network , and the currency image in the infrared and ultraviolet regions is used as input, that is, the RGB three-channel of the currency image is feature stitched with the infrared image of one channel and the currency image of the ultraviolet region of one channel during input (include currency positive In the opposite case, a total of 10 channels of image features) are then fed into the network.
  • the network model is mainly divided into two parts: the backbone network and the texture perception module.
  • image preprocessing is also included before feature extraction.
  • the parameters of the model setting include: number of iterations, learning rate, loss function, and training optimization function.
  • the number of update iterations is 100
  • the learning rate is set to 1e-3
  • the learning rate plan is set.
  • the iteration reaches 80 times and 90 times, the learning rate is multiplied by 0.1 to facilitate the convergence of the model; using Triplet loss (ternary Group loss) function is used as the loss function; SGD (stochastic gradient descent method) is used as the training optimization function. Update the parameters inside the model through backpropagation, and save the model every 20 iterations.
  • test data can be currency of known currency and face value known to be true or false in the training set, or currency of known currency and face value known to be true or false in the training set (open set recognition method ). Divide the dataset into two parts, one is the test library and the other is the template library. The labels should be correct for every image in the test gallery and the template gallery.
  • Feature extraction The test images and templates first need to extract features through the neural network.
  • the specific process of feature extraction is: the depth features are expanded into 1-dimensional features after passing through the texture perception module, and then normalized by 1D Batch Normalization operation. One treatment.
  • (c) Feature comparison The features of the test image and the features in the template library are compared to obtain scores through cosine similarity, and the comparison scores are sorted.
  • the template label in the template library corresponding to the maximum score in the comparison score is the prediction result of the test currency .
  • thresholds can be set, such as when the comparison score is greater than 0.9, the comparison is considered correct, and the performance of the algorithm can be evaluated by calculating EER, FAR, TAR and other indicators. When the algorithm performance indicators meet the requirements, enter the application process of the algorithm.
  • Model application collecting data on the currency to be tested, importing the collected multi-modal features into the currency recognition model, and visually presenting the recognition results. Specifically, model application includes the following steps:
  • S53 Collect multimodal features of the currency to be detected through a multimodal sensor, and input the multimodal features into an algorithm to obtain identification features;
  • the discriminant features are compared with the features in the template library, and the template label in the template library corresponding to the maximum score is the prediction result of the test currency;
  • Steps S53-S55 are repeated to update and manage the currency identification module.
  • the currency in the above scheme is more suitable for banknotes.
  • the present invention also provides a computer-readable storage medium, on which computer instructions are stored, and the steps of the aforementioned method are executed when the computer instructions are executed.
  • a computer-readable storage medium on which computer instructions are stored, and the steps of the aforementioned method are executed when the computer instructions are executed.
  • Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Flash memory or other memory technology, Compact Disc Read-Only Memory (CD-ROM), Digital Versatile Disc (DVD) or other optical storage, Magnetic tape cartridge, tape magnetic disk storage or other magnetic storage device or any other non-transmission medium that can be used to store information that can be accessed by a computing device.
  • computer-readable media excludes transitory computer-readable media, such as modulated data signals and carrier waves.
  • the computer program codes required for the operation of each part of this application can be written in any one or more programming languages, including object-oriented programming languages such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python etc., conventional procedural programming languages such as C language, VisualBasic, Fortran2003, Perl, COBOL2002, PHP, ABAP, dynamic programming languages such as Python, Ruby and Groovy, or other programming languages.
  • the program code may run entirely on the user's computer, or as a stand-alone software package, or run partly on the user's computer and partly on a remote computer, or entirely on the remote computer or processing device.
  • the remote computer can be connected to the user computer through any form of network, such as a local area network (LAN) or wide area network (WAN), or to an external computer (such as through the Internet), or in a cloud computing environment, or as a service Use software as a service (SaaS).
  • LAN local area network
  • WAN wide area network
  • SaaS service Use software as a service
  • the present invention also provides a terminal, including a memory and a processor, the memory stores a template library and computer instructions that can run on the processor, and the processor executes the aforementioned method when running the computer instructions step.
  • a terminal including a memory and a processor
  • the memory stores a template library and computer instructions that can run on the processor
  • the processor executes the aforementioned method when running the computer instructions step.
  • this application adopts the method of multi-modal feature fusion, which integrates image features under different modalities, and improves the recognition and expression ability of deep features;
  • this paper proposes a multi-currency recognition algorithm based on texture perception.
  • the module extracts a large amount of texture information in banknotes to make up for the local features lost when the convolutional neural network extracts deep features, so as to improve the recognition performance;
  • this paper adopts the open set recognition mode, which can realize new currency types with a small number of samples. , New denomination currency recognition.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Pure & Applied Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Mathematical Optimization (AREA)
  • Mathematical Analysis (AREA)
  • Computational Mathematics (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Image Analysis (AREA)
  • Inspection Of Paper Currency And Valuable Securities (AREA)

Abstract

一种多模态货币识别系统、方法、存储介质和终端，属于货币识别技术领域，所述系统及方法采用了纹理感知和开集识别技术，通过多模态特征融合，提高了深度特征的识别能力，纹理感知的多币种识别算法提高识别性能，开集识别算法通过少量样本即可实现新币种、新面值钞票的识别，整体开发成本低、识别精度高。

Description

多模态货币识别系统、方法、存储介质和终端 技术领域
本发明属于货币识别技术领域，具体涉及一种基于纹理感知和开集识别的多模态货币识别系统、方法、存储介质和终端。
背景技术
本部分的陈述仅仅是提供了与本发明相关的背景技术信息，不必然构成在先技术或先有技术。
随着经济全球化的发展，许多国家、地区或者组织的市面上往往流通着多个币种，然而国家之间，地区之间的币种种类繁多，甚至某些币种之间具有一定的相似性，所以这就导致了无法通过人工识别的方式有效的识别大量的币种及其面值，因此如何高效的清分多币种、币种面值和真假是企业和银行等金融组织亟待解决的问题。然而在当前的币种识别系统在使用过程中依然存在一些问题，原因如下：
1、多币种、多面值识别性能有待提高：参见图1，市场上的识别算法通常采用的是ResNet、VGGNet等拥有大量参数的深度卷积神经网络，然而深度卷积神经网络采用的卷积操作通过权值共享获得的是全局特征，因此在货币样本识别过程中随着网络结构深度的加深，逐渐丢失了货币中的纹理特征(局部特征)，这就导致了当货币与货币之间具有一定相似度时(尤其是颜色特征相似度较高时)，识别算法的错误识别率较高。此外，通常的纸币识别方式往往采用闭集识别的方式，即已知待识别币种ID和面值，并且在训练时也是采用对应的ID和面值，然而这种采用闭集识别的方式是无法应对新币种ID和新面值的货币。
2、开发时间较长、开发成本高：为了获得较高的识别率，市场上往往是一种算法对应一个币种，因此导致了在多币种识别的市场中往往需要开发多个算法，多算法识别系统的开发不仅仅需要人力资源的投入，而且当新币种、新版本的货币出现时往往需要重新采集大量的数据集对模型进行训练，甚至需要重新设计算法，这就导致了识别系统的开发时间长、开发的成本高。因此需要一种能够解决在面临新币种、新面值货币时，算法不再需要重新训练的技术方案。
3、币种的采集和存储问题。通常的卷积神经网络算法往往采用数据驱动的模式，即通过采集大量的样本数据对模型进行训练，然而货币数据的采集往往会面临着诸多的问题，比如汇率的问题，大量货币存储时的安全问题等等，因此需要一种能够通过少量样本就能实现新币种、新面值货币识别的技术方案。
综上所诉，现在的多币种识别技术在应用过程中还存在诸多不足。
发明内容
为了克服现有技术的不足，本发明的目的在于提供一种多模态货币识别系统、方法、存储介质和终端，其能解决上述问题。
具体方案如下：
一种基于纹理感知和开集识别的多模态货币识别系统，系统包括电信连接的图像采集模块、图像处理模块、货币识别模块和结果输出模块，其中，所述图像采集模块采用多模态传感器，用于生成待识别货币的多模态特征图；所述图像处理模块包括图像预处理单元和特征提取单元，所述图像预处理单元对接收到的多模态特征图进行统一化预处理，所述特征提取单元通过图像处理算法的主干网络采集图像的全局特征，并通过图像处理 算法的纹理感知模块采集图像的局部特征；所述货币识别模块包括存储单元和对比判定单元，其中在所述存储单元中存储模板库和测试库，所述对比判定单元将接收的特征信息与模板库的货币特征进行比较，判定待测货币的分类和面值；所述结果输出模块通过显示单元将货币识别模块的识别结果可视化呈现。
进一步的，所述多模态传感器包括可见光传感器、红外光传感器和紫外光传感器，用于采集货币正反两面的多模态特征图。
进一步的，所述图像处理模块的主干网络采用10通道的识别类神经网络，接收的图像特征包括货币正反两面对应的可见光RGB三通道图像特征、红外光图像特征和紫外光图像特征共计10通道的图像特征信息。
进一步的，所述图像处理模块的纹理感知模块构成为：
f 1(x,n)＝BConv size＝(1,2n-1)(x)        (1)
f 2(x,n)＝BConv size＝(2n-1,1)(x)       (2)
f 3(x,n)＝BConv dilate＝2n-1(x)      (3)
x 1,x 2,x 3,x 4＝Splited(X)       (4)
Figure PCTCN2021136410-appb-000001
RF＝f 1(X,1)+Cat n＝1,2,3,4F n       (6)
式中，n∈[1,2,3,4]表示分支序号，
公式(1)表示卷积核尺寸为(1,2n-1)的卷积操作以及归一化操作；
公式(2)表示卷积核尺寸为(2n-1,1)的卷积操作以及归一化操作；
公式(3)表示卷积核尺寸为(3,3),膨胀率为2n-1的卷积操作以及归一化操作；
公式(4)表示输入特征X被切割为x n；
公式(5)表示纹理感知模块中的分支网络；
公式(6)表示引入残差机制的纹理感知模块。
进一步的，所述图像处理模块对应的算法模型参数包括迭代次数、学习率、损失函数和训练优化函数，其中，损失函数采用Triplet loss函数，训练优化函数采用SGD法，并通过反向传播更新模型内部的参数。
进一步的，所述货币识别模块采用包括余弦相似计算单元、比较单元和阈值判定单元开集识别算法，所述余弦相似计算单元用于计算待测货币特征的余弦相似度并排序，所述比较单元用于提取余弦相似度得分最高值并与模板库中的模板标签匹配，所述阈值判定单元用于设定阈值并提取有效匹配结果。
本发明还提供了一种货币识别方法，方法包括：
S1、数据采集，通过多模态传感器采集不同币种ID、不同面值且有真有假的货币数据集，因此每张货币对应的是可见光、紫外光、红光下的多模态图像；
S2、数据集的标注,标注采集货币图像的币种ID、面值和真假，使得每张货币包含有币种ID、面值大小和真假三个标签；
S3、模型的训练,通过数据输入、模型构建和模型参数设置来训练获得具有10个输入通道的且具有纹理感知模块的货币识别模型；
S4、模型测试，通过数据划分、特征提取和特征对比来测试评估模型；
S5、模型应用，对待测货币进行数据采集，并将采集的多模态特征导入货币识别模型，并将识别结果可视化呈现。
进一步的，步骤S5的模型应用包括：
S51、将带有算法、算法参数、模型权重的货币识别模型嵌入至图像处理设备；
S52、建立拥有大量币种ID、不同面值包含真假的模板库，将模板库中的图像输入处理设备并提取模板库中的图像的特征，存储特征及其标签；
S53、将待检测货币，通过多模态传感器采集多模态特征，并将多模态特征输入算法，获得鉴别特征；
S54、鉴别特征通过与模板库中的特征进行得分比对，得分最大分值对应的模板库中的模板标签为测试货币的预测结果；
S55、当出现新币种、新面值的货币时，采集少量的货币样本，对货币进行标注，并将新货币特征添加至模板库；
S56、重复步骤S53-S55，对货币识别模进行更新和管理。
本发明还提供了一种计算机可读存储介质，其上存储有计算机指令，所述计算机指令运行时执行前述方法的步骤。
本发明还提供了一种终端，包括存储器和处理器，所述存储器上储存有模板库和能够在所述处理器上运行的计算机指令，所述处理器运行所述计算机指令时执行前述方法的步骤。
相比现有技术，本发明的有益效果在于：本申请的系统及方法采用了纹理感知和开集识别技术，通过多模态特征融合，提高了深度特征的识别能力，纹理感知的多币种识别算法提高识别性能，开集识别算法通过少量样本即可实现新币种、新面值钞票的识别，整体开发成本低、识别精度高。
附图说明
图1为现有识别算法示意图；
图2为多模态货币识别系统示意图；
图3为多模态特征融合示意图；
图4为纹理感知算法过程图；
图5为纹理感知模块结构图；
图6为基于纹理感知的开集识别算法过程图；
图7为货币识别方法流程图。
具体实施方式
为使本发明实施例的目的、技术方案和优点更加清楚，下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本发明的一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动的前提下所获得的所有其他实施例，都属于本发明保护的范围。
本说明书中使用了流程图用来说明根据本说明书的实施例的系统所执行的操作。应当理解的是，前面或后面操作不一定按照顺序来精确地执行。相反，可以按照倒序或同时处理各个步骤。同时，也可以将其他操作添加到这些过程中，或从这些过程移除某一步或数步操作。
第一实施例
一种基于纹理感知和开集识别的多模态货币识别系统，参见图2，系统包括电信连接的图像采集模块、图像处理模块、货币识别模块和结果输出模块。
其中，图像采集模块采用多模态传感器，用于生成待识别货币的多模态特征图。具体的，所述多模态传感器包括可见光传感器、红外光传感器 和紫外光传感器，用于采集货币正反两面的多模态特征图。
其中，所述图像处理模块包括图像预处理单元和特征提取单元，所述图像预处理单元对接收到的多模态特征图进行统一化预处理，所述特征提取单元通过图像处理算法的主干网络采集图像的全局特征，并通过图像处理算法的纹理感知模块采集图像的局部特征。
进一步的，所述图像处理模块的主干网络采用10通道的识别类神经网络，接收的图像特征包括货币正反两面对应的可见光RGB三通道图像特征、红外光图像特征和紫外光图像特征共计10通道的图像特征信息。
具体的多模态特征融合，货币在经过多模态传感器(包括可见光传感器、红外光传感器、紫外光传感器)之后，能够得到可见光、紫外光、红外光区域的多模态特征图(货币对紫外光、红外光产生荧光反映，因此紫外光和红外光的多模态特征具备一定的鉴别真假的能力)。多模态特征图采用特征级联(拼接)的方式进行特征融合，如图3，多模态特征融合示意图。
具体来说在货币识别中，紫外光、可见光、红外光多模态特征图的融合能够获具备10通道(正、反面各五个通道，五个通道包括：可见光RGB三通道图像、一个通道的红外光图像、一个通道的紫外光图像)的多模态特征图，多模态特征图在输入网络以后采用常用的卷积方式提取深度特征，与此同时卷积核通过权值共享的方式提取不同通道同一感受野区域的各个模态的特征，随后进行特征整合，即将紫外光和红外光多模态特征中映射到深度特征中。紫外光和红外光多模态特征图的引入可以提高算法鉴别真假的能力。
进一步的，为了提高多币种识别算法的性能，本申请设计了基于纹理 感知的识别算法，具体为前述纹理感知模块的嵌入。在日常生活中采用的常规的识别算法为了获得更高的识别率，往往采用的是VGG、ResNet等拥有大量参数的深度神经网络。然而卷积神经网络中CNN往往是获得全局特征的，因此缺乏挖掘局部特征的能力。然而在货币中往往存在大量的纹理特征(局部特征中的一种)，因此本申请通过设计纹理感知模块挖掘货币的纹理特征，提高多币种的识别性能，纹理感知算法过程图参见图4。
进一步的，对于纹理感知模块，在人类的视觉系统中，一组不同大小的感受野有助于突出靠近视网膜中央凹的区域，该区域对微小的空间变化(比如纹理变化)非常敏感，有利于目标识别。因此本申请通过不同卷积核尺寸的卷积方式模拟人类视觉感官，设计了纹理感知模块。同时由于高维度映射不利于特征表达而同一特征层中虽然在不同通道携带的特征存在细微区别，但是多通道特征之间的统计特性是一致的，因此本文为了缓解高维度特征向低纬度特征过渡时存在的特征丢失问题，提出了基于通道切分的纹理感知模块。所述纹理感知模块构成为：
f 1(x,n)＝BConv size＝(1,2n-1)(x)      (1)
f 2(x,n)＝BConv size＝(2n-1,1)(x)       (2)
f 3(x,n)＝BConv dilate＝2n-1(x)       (3)
式中，n∈[1,2,3,4]表示分支序号，公式(1)表示卷积核尺寸为(1,2n-1)的卷积操作以及归一化操作，公式(2)表示卷积核尺寸为(2n-1,1)的卷积操作以及归一化操作，公式(3)表示卷积核尺寸为(3,3),膨胀率为2n-1的卷积操作以及归一化操作。
而输入特征X可以被切割x n可以由公式(4)表达,而纹理感知模块中的分支网络可以由公式(5)表达:
x 1,x 2,x 3,x 4＝Splited(X)        (4)
Figure PCTCN2021136410-appb-000002
同时，为了优化纹理感知模块，还可以引入残差机制，总体来说纹理感知模块可以表示为：
RF＝f 1(X,1)+Cat n＝1,2,3,4F n       (6)
与上述算法过程对应的纹理感知模块还可参见图5。
进一步的，所述图像处理模块对应的算法模型参数包括迭代次数、学习率、损失函数和训练优化函数，其中，损失函数采用Triplet loss函数，训练优化函数采用SGD法，并通过反向传播更新模型内部的参数。
其中，所述货币识别模块包括存储单元和对比判定单元，其中在所述存储单元中存储模板库和测试库，所述对比判定单元将接收的特征信息与模板库的货币特征进行比较，判定待测货币的分类和面值。
进一步的，所述货币识别模块采用包括余弦相似计算单元、比较单元和阈值判定单元开集识别算法，所述余弦相似计算单元用于计算待测货币特征的余弦相似度并排序，所述比较单元用于提取余弦相似度得分最高值并与模板库中的模板标签匹配，所述阈值判定单元用于设定阈值并提取有效匹配结果。
具体的，货币识别模块整体采用开集识别的解决方案，通常的纸币识别方式往往采用闭集识别的方式，即已知待识别币种ID和面值，并且在训练时也是采用对应的ID和面值，然而这种采用闭集识别的方式是无法应对新币种ID和新面值的货币。因此文采用了开集识别的方式，即待识别的币种ID可以允许存在未见过的ID或面值。具体而言，在识别的过程 中，训练好的神经网络算法只作为特征提取网络而不参与分类识别，参见图6，测试集中的每张货币与模板库中的货币进行比较，比较得分中最大分值对应的模板库中的模板标签为待识别货币的预测结果。
其中，所述结果输出模块通过显示单元将货币识别模块的识别结果可视化呈现。
具体的，显示单元采用如液晶屏、LED屏等显示器，接收货币识别模块的预测结果并显示。
第二实施例
一种基于第一实施例所述系统的货币识别方法，方法包括：
S1、数据采集
通过多模态传感器采集不同币种ID、不同面值且有真有假的货币数据集，因此每张货币对应的是可见光、紫外光、红光下的多模态图像。
S2、数据集的标注
标注采集货币图像的币种ID、面值和真假，使得每张货币包含有币种ID、面值大小和真假三个标签。
S3、模型的训练
通过数据输入、模型构建和模型参数设置来训练获得具有10个输入通道的且具有纹理感知模块的货币识别模型。
S31、数据输入
在图像数据的输入过程中，首先需要调整网络的输入：如前所述，本文中采用的是基于多模态的图像输入，因此在输入图像时不仅仅只是将可见光区域的货币图像输入到网络，而且将红外光和紫外光区域的货币图像 作为输入，即在输入时货币图像的RGB三通道与一个通道的红外光图像、一个通道的紫外光区域的货币图像进行特征拼接(在包含货币正反的情况下，总共是10通道的图像特征)然后输入到网络中。
S32、模型的构建
网络模型主要分为两个部分：主干网络，纹理感知模块。当然在特征提取前还包括图像的预处理。
(a)主干网络，本文采用常用的识别网络作为主干网络(比如VGG、ResNet等)进行特征的初步提取，，在训练过程中加载经过ImageNet数据集训练之后的预训练模型，在本文中需要调整网络的输入通道，设置为10。
(b)纹理感知模块，为了进一步提高模型的识别率，本文采用纹理感知模块，通过获得货币中的纹理特征辅助货币识别，模块的具体设计过程参见前述中的式(1)-(6)以及图5。
S33、模型参数的设置和更新
模型设置的参数包括：迭代次数，学习率，损失函数，训练优化函数。一个示例中，更新迭代次数为100，学习率设置为1e-3，设置学习率计划，在迭代到80次，90次时，学习率乘以0.1，便于模型的收敛；采用Triplet loss(三元组损失)函数作为损失函数；采用SGD(随机梯度下降法)作为训练优化函数。通过反向传播更新模型内部的参数，每迭代更新20次保存一次模型。
S4、模型测试，通过数据划分、特征提取和特征对比来测试评估模型；
(a)数据划分：测试数据可以是训练集中已知真假的已知币种和面值的货币，也可以不是训练集中已知真假的已知币种和面值的货币(开集 识别的方式)。将数据集细分为两个部分，一个是测试库，一个是模板库。测试库和模板库中的每张图像的标签均应该是正确的。
(b)特征提取：测试图像和模板首先需要经过神经网络提取特征，特征提取的具体过程为：深度特征在经过纹理感知模块以后被展开成为1维的特征，随后经过1D的Batch Normalization操作进行归一化处理。
(c)特征对比：测试图像的特征和模板库中的特征经过余弦相似度进行比较获得得分，对比得分进行排序，比较得分中最大分值对应的模板库中的模板标签为测试货币的预测结果。与此同时，为了在应用过程中获得更好的结果，可以设置阈值，如对比得分大于0.9时，才认为对比正确，而且可以通过计算EER，FAR，TAR等指标评估算法性能。当算法性能指标均满足要求时进入算法的应用过程。
S5、模型应用，对待测货币进行数据采集，并将采集的多模态特征导入货币识别模型，并将识别结果可视化呈现。具体的，模型应用包括以下步骤：
S51、将带有算法、算法参数、模型权重的货币识别模型嵌入至图像处理设备；
S52、建立拥有大量币种ID、不同面值包含真假的模板库，将模板库中的图像输入处理设备并提取模板库中的图像的特征，存储特征及其标签；
S53、将待检测货币，通过多模态传感器采集多模态特征，并将多模态特征输入算法，获得鉴别特征；
S54、鉴别特征通过与模板库中的特征进行得分比对，得分最大分值对应的模板库中的模板标签为测试货币的预测结果；
S55、当出现新币种、新面值的货币时，采集少量的货币样本，对货 币进行标注，并将新货币特征添加至模板库；
S56、重复步骤S53-S55，对货币识别模进行更新和管理。
上述方案中的货币，更多的适应于钞票。
第三实施例
本发明还提供了一种计算机可读存储介质，其上存储有计算机指令，所述计算机指令运行时执行前述方法的步骤。其中，所述方法请参见前述部分的详细介绍，此处不再赘述。
本领域普通技术人员可以理解上述实施例的各种方法中的全部或部分步骤是可以通过程序来指令相关的硬件来完成，该程序可以存储于计算机可读存储介质中，计算机可读介质包括永久性和非永久性、可移动和非可移动媒体可以由任何方法或技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机的存储介质的例子包括，但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)或其他光学存储、磁盒式磁带，磁带磁磁盘存储或其他磁性存储设备或任何其他非传输介质，可用于存储可以被计算设备访问的信息。按照本文中的界定，计算机可读介质不包括暂存电脑可读媒体(transitory media)，如调制的数据信号和载波。
本申请各部分操作所需的计算机程序编码可以用任意一种或多种程序语言编写，包括面向对象编程语言如Java、Scala、Smalltalk、Eiffel、JADE、Emerald、C++、C#、VB.NET、Python等，常规程序化编程语言如C 语言、VisualBasic、Fortran2003、Perl、COBOL2002、PHP、ABAP，动态编程语言如Python、Ruby和Groovy，或其他编程语言等。该程序编码可以完全在用户计算机上运行、或作为独立的软件包在用户计算机上运行、或部分在用户计算机上运行部分在远程计算机运行、或完全在远程计算机或处理设备上运行。在后种情况下，远程计算机可以通过任何网络形式与用户计算机连接，比如局域网(LAN)或广域网(WAN)，或连接至外部计算机(例如通过因特网)，或在云计算环境中，或作为服务使用如软件即服务(SaaS)。
第四实施例
本发明还提供了一种终端，包括存储器和处理器，所述存储器上储存有模板库和能够在所述处理器上运行的计算机指令，所述处理器运行所述计算机指令时执行前述方法的步骤。其中，所述方法请参见前述部分的详细介绍，此处不再赘述。
综上，本申请采用多模态特征融合的方式，融合了不同模态下的图像特征，提高了深度特征的识别表达能力；其次本文提出了基于纹理感知的多币种识别算法，通过纹理感知模块提取纸币中存在的大量纹理信息弥补卷积神经网络在提取深度特征时丢失的局部特征，达到提高识别性能的目的；最后本文采用开集识别的模式，能够通过少量样本就能实现新币种、新面值货币识别。
最后应说明的是：以上实施例仅用以说明本发明的技术方案，而非对其限制；尽管参照前述实施例对本发明进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修 改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本发明各实施例技术方案的精神和范围。

Claims (10)

  1. 一种基于纹理感知和开集识别的多模态货币识别系统，其特征在于：系统包括电信连接的图像采集模块、图像处理模块、货币识别模块和结果输出模块，其中，
    所述图像采集模块采用多模态传感器，用于生成待识别货币的多模态特征图；
    所述图像处理模块包括图像预处理单元和特征提取单元，所述图像预处理单元对接收到的多模态特征图进行统一化预处理，所述特征提取单元通过图像处理算法的主干网络采集图像的全局特征，并通过图像处理算法的纹理感知模块采集图像的局部特征；
    所述货币识别模块包括存储单元和对比判定单元，其中在所述存储单元中存储模板库和测试库，所述对比判定单元将接收的特征信息与模板库的货币特征进行比较，判定待测货币的分类和面值；
    所述结果输出模块通过显示单元将货币识别模块的识别结果可视化呈现。
  2. 根据权利要求1所述的系统，其特征在于：所述多模态传感器包括可见光传感器、红外光传感器和紫外光传感器，用于采集货币正反两面的多模态特征图。
  3. 根据权利要求2所述的系统，其特征在于：所述图像处理模块的主干网络采用10通道的识别类神经网络，接收的图像特征包括货币正反两面对应的可见光RGB三通道图像特征、红外光图像特征和紫外光图像特征共计10通道的图像特征信息。
  4. 根据权利要求3所述的系统，其特征在于：所述图像处理模块的纹 理感知模块构成为：
    f 1(x,n)＝BConv size＝(1,2n-1)(x)  (1)
    f 2(x,n)＝BConv size＝(2n-1,1)(x)  (2)
    f 3(x,n)＝BConv dilate＝2n-1(x)  (3)
    x 1,x 2,x 3,x 4＝Splited(X)  (4)
    Figure PCTCN2021136410-appb-100001
    RF＝f 1(X,1)+Cat n＝1,2,3,4F n  (6)
    式中，n∈[1,2,3,4]表示分支序号，
    公式(1)表示卷积核尺寸为(1,2n-1)的卷积操作以及归一化操作；
    公式(2)表示卷积核尺寸为(2n-1,1)的卷积操作以及归一化操作；
    公式(3)表示卷积核尺寸为(3,3),膨胀率为2n-1的卷积操作以及归一化操作；
    公式(4)表示输入特征X被切割为x n；
    公式(5)表示纹理感知模块中的分支网络；
    公式(6)表示引入残差机制的纹理感知模块。
  5. 根据权利要求4所述的系统，其特征在于：所述图像处理模块对应的算法模型参数包括迭代次数、学习率、损失函数和训练优化函数，其中，损失函数采用Triplet loss函数，训练优化函数采用SGD法，并通过反向传播更新模型内部的参数。
  6. 根据权利要求4所述的系统，其特征在于：所述货币识别模块采用包括余弦相似计算单元、比较单元和阈值判定单元开集识别算法，所述余 弦相似计算单元用于计算待测货币特征的余弦相似度并排序，所述比较单元用于提取余弦相似度得分最高值并与模板库中的模板标签匹配，所述阈值判定单元用于设定阈值并提取有效匹配结果。
  7. 一种基于权利要求1-6任一项所述系统的货币识别方法，其特征在于，方法包括：
    S1、数据采集，通过多模态传感器采集不同币种ID、不同面值且有真有假的货币数据集，因此每张货币对应的是可见光、紫外光、红光下的多模态图像；
    S2、数据集的标注,标注采集货币图像的币种ID、面值和真假，使得每张货币包含有币种ID、面值大小和真假三个标签；
    S3、模型的训练,通过数据输入、模型构建和模型参数设置来训练获得具有10个输入通道的且具有纹理感知模块的货币识别模型；
    S4、模型测试，通过数据划分、特征提取和特征对比来测试评估模型；
    S5、模型应用，对待测货币进行数据采集，并将采集的多模态特征导入货币识别模型，并将识别结果可视化呈现。
  8. 根据权利要求7所述的方法，其特征在于，模型应用包括：
    S51、将带有算法、算法参数、模型权重的货币识别模型嵌入至图像处理设备；
    S52、建立拥有大量币种ID、不同面值包含真假的模板库，将模板库中的图像输入处理设备并提取模板库中的图像的特征，存储特征及其标签；
    S53、将待检测货币，通过多模态传感器采集多模态特征，并将多模态特征输入算法，获得鉴别特征；
    S54、鉴别特征通过与模板库中的特征进行得分比对，得分最大分值对应的模板库中的模板标签为测试货币的预测结果；
    S55、当出现新币种、新面值的货币时，采集少量的货币样本，对货币进行标注，并将新货币特征添加至模板库；
    S56、重复步骤S53-S55，对货币识别模进行更新和管理。
  9. 一种计算机可读存储介质，其上存储有计算机指令，其特征在于：所述计算机指令运行时执行权利要求7或8所述方法的步骤。
  10. 一种终端，包括存储器和处理器，所述存储器上储存有模板库和能够在所述处理器上运行的计算机指令，其特征在于：所述处理器运行所述计算机指令时执行权利要求7或8所述方法的步骤。
PCT/CN2021/136410 2021-10-25 2021-12-08 多模态货币识别系统、方法、存储介质和终端 WO2023070843A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111238612.0A CN113901942A (zh) 2021-10-25 2021-10-25 多模态货币识别系统、方法、存储介质和终端
CN202111238612.0 2021-10-25

Publications (1)

Publication Number Publication Date
WO2023070843A1 true WO2023070843A1 (zh) 2023-05-04

Family

ID=79026436

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/136410 WO2023070843A1 (zh) 2021-10-25 2021-12-08 多模态货币识别系统、方法、存储介质和终端

Country Status (2)

Country Link
CN (1) CN113901942A (zh)
WO (1) WO2023070843A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115081554B (zh) * 2022-08-16 2023-04-07 山东省齐鲁大数据研究院 一种实现货币类数据智能转换方法、系统及终端机

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040213448A1 (en) * 2003-04-28 2004-10-28 Asn Technology Corp. Apparatus for recognizing counterfeit currency and method thereof
CN108665603A (zh) * 2018-04-11 2018-10-16 深圳怡化电脑股份有限公司 识别纸币币种的方法、装置及电子设备
CN111275879A (zh) * 2020-02-20 2020-06-12 Oppo广东移动通信有限公司 货币识别方法、装置及移动终端
CN111652318A (zh) * 2020-06-05 2020-09-11 中国银行股份有限公司 一种币种识别方法、识别装置、以及电子设备
CN111667002A (zh) * 2020-06-05 2020-09-15 中国银行股份有限公司 一种币种识别方法、识别装置、以及电子设备

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040213448A1 (en) * 2003-04-28 2004-10-28 Asn Technology Corp. Apparatus for recognizing counterfeit currency and method thereof
CN108665603A (zh) * 2018-04-11 2018-10-16 深圳怡化电脑股份有限公司 识别纸币币种的方法、装置及电子设备
CN111275879A (zh) * 2020-02-20 2020-06-12 Oppo广东移动通信有限公司 货币识别方法、装置及移动终端
CN111652318A (zh) * 2020-06-05 2020-09-11 中国银行股份有限公司 一种币种识别方法、识别装置、以及电子设备
CN111667002A (zh) * 2020-06-05 2020-09-15 中国银行股份有限公司 一种币种识别方法、识别装置、以及电子设备

Also Published As

Publication number Publication date
CN113901942A (zh) 2022-01-07

Similar Documents

Publication Publication Date Title
Liu et al. Joint optic disc and cup segmentation using semi-supervised conditional GANs
RU2760471C1 (ru) Способы и системы идентификации полей в документе
CN102324038B (zh) 一种基于数字图像的植物种类识别方法
CN109190698B (zh) 一种网络数字虚拟资产的分类识别系统及方法
CN114329034B (zh) 基于细粒度语义特征差异的图像文本匹配判别方法及系统
WO2022052459A1 (zh) 一种图像类别检测方法、系统、电子设备及存储介质
CN109086794B (zh) 一种基于t-lda主题模型的驾驶行为模式识方法
CN116821698B (zh) 基于半监督学习的小麦赤霉病孢子检测方法
CN109726918A (zh) 基于生成式对抗网络和半监督学习的个人信用确定方法
CN107220598A (zh) 基于深度学习特征和Fisher Vector编码模型的虹膜图像分类方法
WO2023070843A1 (zh) 多模态货币识别系统、方法、存储介质和终端
CN114881775B (zh) 一种基于半监督集成学习的欺诈检测方法及系统
CN106875543B (zh) 一种基于rgb-d相机的视障人士纸币识别系统及识别方法
CN114971294A (zh) 数据采集方法、装置、设备及存储介质
Younis et al. A new parallel bat algorithm for musical note recognition.
CN109583423A (zh) 一种手写数字识别的方法、装置及相关组件
WO2021196477A1 (zh) 基于声纹特征与关联图谱数据的风险用户识别方法、装置
CN106846608B (zh) 一种基于rgb-d相机的视障人士纸币识别眼镜
Jadhav et al. Currency Recognition using Machine Learning
CN115713399B (zh) 一种结合第三方数据源的用户信用评估系统
Mahmood Currency Detection for Visually Impaired Iraqi Banknote as a Study Case
Ng et al. An intelligent banknote recognition system by using machine learning with assistive technology for visually impaired people
Sun et al. The recognition of new and old banknotes based on SVM
Sundravadivelu et al. Extensive Analysis of IoT Assisted Fake Currency Detection using Novel Learning Scheme
Liu et al. Implementation of credit scoring card model based on logistic regression and lightgbm

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE