KR101991028B1 - Device and method for finger-vein recognition - Google Patents

Abstract

The present invention relates to an apparatus and method for finger vein recognition, and more particularly, to an apparatus and method for performing finger vein recognition by determining whether a finger vein image is photographed in an infrared environment based on a convolutional neural network. will be. The present invention can perform finger vein recognition more accurately by determining whether the input finger vein image is forged through a deep learning based convolutional neural network and a support vector machine.

Description

DEVICE AND METHOD FOR FINGER-VEIN RECOGNITION}

The present invention relates to a finger vein recognition device and method, and more particularly, to determine whether the forgery of the finger vein image based on deep learning based convolutional neural network and Support Vector Machine (Support Vector Machine) An apparatus and method for performing finger vein recognition.

Finger vein recognition is a technology that reads vein patterns and hemoglobin patterns inside a finger using near-infrared sensors for biometrics. Unlike finger prints, finger vein recognition is not affected by the external environment such as foreign matter, moisture and dryness on the surface of the finger. However, because the forgery or tampering is impossible, there is a problem that the element is no longer available when the personal finger vein information leaked. Thus, the need for a technique for discriminating forgery or modulation of the finger vein is increasing.

Background art of the present invention is disclosed in Republic of Korea Patent Publication No. 10-1354881 (2014.02.05).

The present invention provides a finger vein recognition device and method for performing finger vein recognition by judging whether the input finger vein image through a deep learning based convolutional neural network and a support vector machine.

According to an aspect of the present invention, a finger vein recognition device is provided.

In the finger vein recognition device according to an embodiment of the present invention, the input unit receives a finger vein image generated by photographing a vein pattern inside a finger, sets a region of interest in the input finger vein image, and normalizes it to a predetermined size. Characterization of finger vein image using preprocessing unit that corrects pixel brightness by applying zero-mean normalization to the selected finger vein image, and previously learned convolutional neural network (CNN) model And a forgery determination unit for determining whether the extracted feature is a real feature or a forged feature, and a finger vein recognition unit for performing finger vein recognition when the extracted feature is determined to be an actual feature.

According to another aspect of the present invention, a finger vein recognition method and a computer program for executing the same are provided.

Finger vein recognition method according to an embodiment of the present invention and a computer program for executing the step of receiving a finger vein image generated by photographing the vein pattern inside the finger, and set the region of interest in the input finger vein image, Preprocessing by normalizing the size and pixel brightness of a region of interest, extracting features of the finger vein image using the previously learned CNN model, and optimizing them through principal component analysis Selecting a feature, learning the selected best feature through a support vector machine and classifying it into a real feature class or a forged feature class; and when an optimal feature is classified as a forged feature class, Outputting and performing finger vein recognition when the optimal feature is classified as a real feature class. It can hamhal.

The present invention can perform finger vein recognition more accurately by determining whether the input vein image is forged through a deep learning based convolutional neural network and a support vector machine.

1 to 5 are diagrams for explaining a finger vein recognition device according to an embodiment of the present invention.
6 to 7 are diagrams for explaining a finger vein recognition method for determining whether a forgery using a convolutional neural network (CNN) model for feature extraction of finger vein images according to an embodiment of the present invention.
8 is a diagram illustrating a finger vein recognition method using the CNN model according to an embodiment of the present invention for the purpose of feature extraction and forgery determination of finger vein images.
9 to 13 are diagrams for explaining a database constructed for the performance evaluation result of the trained CNN model according to an embodiment of the present invention.
14 to 16 are diagrams comparing the recognition performance error rates of the finger vein recognition method and the previously studied finger vein recognition method according to an embodiment of the present invention.

As the inventive concept allows for various changes and numerous embodiments, particular embodiments will be illustrated in the drawings and described in detail in the written description. However, this is not intended to limit the present invention to specific embodiments, it should be understood to include all changes, equivalents, and substitutes included in the spirit and scope of the present invention. In describing the present invention, when it is determined that the detailed description of the related known technology may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. Also, the singular forms used in the specification and claims are to be interpreted as generally meaning "one or more", unless stated otherwise.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings, and in the following description with reference to the accompanying drawings, the same or corresponding components will be given the same reference numerals and redundant description thereof will be omitted. Shall be.

1 to 6 are views for explaining a finger vein recognition device according to an embodiment of the present invention.

Referring to FIG. 1, the finger vein recognition device 100 includes a photographing unit 110, an input unit 120, a preprocessor 130, a forgery determination unit 140, and a finger vein recognition unit 150.

The photographing unit 110 generates a finger vein image by photographing a vein pattern inside a finger through a near infrared camera.

The input unit 120 receives the finger vein image photographed by the photographing unit 110.

The preprocessor 130 sets a region of interest in the input finger vein image and normalizes the size and pixel brightness of the region of interest. For example, the preprocessor 130 may receive an input of an actual finger vein image or an image of a fake finger vein image, remove a background region, set a region of interest, and normalize it to a predetermined pixel size. In addition, the preprocessor 130 may correct the pixel brightness of the image by applying zero-mean normalization to the normalized finger vein image. For example, the preprocessor 130 may correct the pixel brightness between 0 and 255 to be within a range between -127 and 127 or between -1 and 1.

The forgery determination unit 140 extracts a feature using a convolutional neural network (CNN) model previously learned from the normalized finger vein image, and whether the extracted feature is a real vein feature or a forged vein feature. Determine.

Referring to FIG. 2, the CNN structure is generally composed of two important parts, convolutional layers and fully-connected layers. Each convolutional layer may extract and extract image feature vectors X = {X_1, X_2, ..., X_n} by manipulating and extracting features of the input image. The extracted image feature vectors are input to the fully connected layer to classify the input images into predefined categories. Finger vein recognition device 100 according to an embodiment of the present invention can determine whether the forgery of the finger vein image by the first CNN model or the second CNN model designed by changing the above-described CNN structure. Here, the first CNN model may be designed based on the Alex network, and the second CNN model may be designed based on the Visual Geometry Group (VGG) network.

Referring to FIG. 3, the first CNN model includes (1) an input layer, (2) a set of five convolution layers, (3) a set of two fully connected layers, and (4) It includes an output layer.

(1) The input layer receives a finger vein image of a preset scale. In this case, the size of the finger vein image may be 87 × 151 × 3 pixels.

(2) Five sets of convolutional layers extract a feature of an image through a convolution operation by applying a plurality of filter masks to an input finger vein image. Each of the five convolutional layer sets is in conjunction with a rectified linear unit (hereinafter referred to as ReLU) layer and may include a normalization layer and a max pooling layer.

In the set of five convolution layers, an image of 87 × 151 × 3 pixels is input to the first convolution layer, and can be convolved at intervals of 2 pixels using 96 filters of 11 × 11 × 3 sizes. In this case, the weight size per filter is 11 × 11 × 3 = 363, and the total number of parameters in the first convolutional layer is (121 + 1) × 363 = 44286, where 1 represents a bias. The size of the feature map in the first convolutional layer is 39 × 71 × 96 and 39 and 71 are the output height and width. The output height (or width) can be calculated using the formula: input height (or width)-filter height (or width) + 2 x padding) / stride +1. The outputs go through the ReLU layer and the normalization layer. Here, the ReLU layer performs a threshold operation that sets all of the values when the value input based on 0 is less than zero. Thus, the ReLU layer helps to speed up the learning of deep networks by simplifying computation. The output map of the normalization layer is processed by the first max pooling layer at intervals of 2 pixels using one filter having a size of 3 × 3 and then downsampled. The output from the first max pooling layer is 39 (= ((87-11 +) by the output height (or width) = input height (or width)-filter height (or width) + 2 × padding) / stride + 1 equation. 2 × 0) / 2 × 1)) × 71 (= ((151−11 + 2 × 0) / 2 + 1)) × 96. Subsequently, the second convolution layer is convolved by 2 paddings at 1 pixel intervals using 128 filters having a size of 5 × 5 × 96, and a second filter having a 3 × 3 filter is passed through a ReLU layer and a normalization layer. Apply the Max pooling layer to reduce the size of the feature map every two pixels. The third convolution layer and the fourth convolution layer are convolved by one padding at intervals of one pixel by using 192 filters having a size of 3 × 3 × 128 and apply a ReLU layer. Next, the fifth convolutional layer applies 128 3 × 3 × 256 filters, each padded at 1 pixel intervals, and a third max pooling layer having 3 × 3 filters through the ReLU layer and the normalization layer. The output map of 4x8x128 pixels is output by applying the pixel interval.

(3) Two Fully Connected Layers reduce the feature by connecting the result of the convolution layer to a plurality of nodes to form a one-dimensional matrix. The first fully connected layer sets may include a fully connected layer and a ReLU layer, and the second fully connected layer set may include a fully connected layer, a ReLU layer, and a drop out layer. In this case, the first fully connected layer may have 768 and 2048 nodes as inputs and outputs, respectively, and a dropout layer may be applied to set the output of each hidden node to 0 based on a randomly preset probability. Here, the preset probability may be 0.50. The first fully connected layer may apply a ReLU layer after applying the dropout. The second fully connected layer may have 2048 and 1024 nodes as input and output, respectively. The second fully connected layer may apply a dropout layer after applying a ReLU layer.

(4) The output layer classifies the features output from the second fully connected layer into two classes through the Softmax layer. For example, each of the two classes may be an actual feature class and a fake feature class.

Referring to FIG. 4, the second CNN model includes (1) an input layer, (2) a set of 13 convolution layers, (3) a set of two fully connected layers, and (4) It includes an output layer.

(1) The input layer receives a predetermined finger vein image. In this case, the size of the finger vein image may be 128 × 256 × 3 pixels.

(2) 13 convolutional layer sets extract a feature of an image through a convolution operation by applying a plurality of filter masks to the input finger vein image. Each of the thirteen convolution layers may include a rectified linear unit (hereinafter referred to as ReLU) layer and a max pooling layer.

In a set of 13 convolution layers, an image of 128 × 256 × 3 pixels is input to a 1-1 convolution layer, and convolutional by 1 padding at 1 pixel intervals using 32 filters of 3 × 3 × 3 size Can be. In this case, the weight size per filter is 3 × 3 × 3 = 27, and the total number of parameters in the first convolutional layer is (9 + 1) × 27 = 270, where 1 represents a bias. The size of the feature map in the 1-1 convolution layer is 128 × 256 × 32, and 128 and 256 are output height and width. The output height (or width) can be calculated using the formula: input height (or width)-filter height (or width) + 2 x padding) / stride + 1. The outputs perform a critical operation that sets all values to 0 when the value input based on 0 is smaller than 0 through the ReLU layer. Thus, the ReLU layer helps to speed up the learning of deep networks by simplifying computation. The output map of the ReLU layer is convolved by 1 padding at 1 pixel intervals using 32 filters of size 3 × 3 × 32 in the 1-2 convolution layer, and passes through the ReLU layer. The output map of the ReLU layer is processed by the first max pooling layer at 2 pixel intervals using one filter having a size of 2 × 2 and then downsampled. The output from the first max pooling layer is 64 (= ((128-2 +) by the output height (or width) = input height (or width) -filter height (or width) + 2 × padding) / stride +1 equation. 2 x 0) / 2 + 1)) x 128 (= ((256-2 + 2 x 0) / 2 + 1)) x 32. An image of 64 × 128 × 64 pixels is input to the 2-1 convolution layer, and convolutionalized by 1 padding at 1 pixel intervals using 64 filters of 3 × 3 × 32 size. The output of the 2-1 convolutional layer is input to the 2-2 convolutional layer via the ReLU layer. The 2-2 convolution layer is convolved by 1 padding at 1 pixel intervals using 64 filters having a size of 3 × 3 × 64. The output of the 2-2 convolutional layer is processed at 2 pixel intervals by one filter of 2x2 size by the second max pooling layer and then downsampled. An image of 32 × 64 × 128 pixels is input to the 3-1 convolution layer, and convolutional by 1 padding is performed at intervals of 1 pixel by using 128 filters having a 3 × 3 × 64 size. The output of the 3-1 convolution layer is input to the 3-2 convolution layer via the ReLU layer. The 3-2 convolution layer is convolved by 1 padding at 1 pixel intervals using 128 filters having a size of 3 × 3 × 128. The output of the 3-2 convolution layer is input to the 3-3 convolution layer via the ReLU layer. The 3-3 convolution layer is convolved by 1 padding at 1 pixel intervals using 128 filters having a size of 3 × 3 × 128. The output of the 3-3 convolutional layer is processed by the third max pooling layer at 2 pixel intervals using one filter of 2 × 2 size and then downsampled.

An image of 16 × 32 × 128 pixels is input to the 4-1 convolution layer, and convolutionalized by one padding at intervals of one pixel using 256 filters having a size of 3 × 3 × 128. The output of the 4-1 convolutional layer is input to the 4-2 convolutional layer via the ReLU layer. The 4-2 convolution layer is convolved by one padding at intervals of one pixel using 256 filters having a size of 3 × 3 × 256. The output of the 4-2 convolution layer is input to the 4-3 convolution layer via the ReLU layer. The 4-3 convolution layer is convolved by one padding at intervals of one pixel using 256 filters having a size of 3 × 3 × 256. The output of the 4-3 convolutional layer is processed down to 2 pixel intervals by one filter of 2x2 size by the fourth max pooling layer and then downsampled. An image of 8 × 16 × 256 pixels is input to the 5-1 convolution layer, and is convolved by 1 padding at intervals of 1 pixel using 256 filters having a size of 3 × 3 × 256. The output of the 5-1 convolutional layer is input to the 5-2 convolutional layer via the ReLU layer. The 5-2 convolution layer is convolved by 1 padding at intervals of 1 pixel using 256 filters having a size of 3 × 3 × 256. The output of the 5-2 convolution layer is input to the 5-3 convolution layer via the ReLU layer. The 5-3 convolution layer is convolved by 1 padding at 1 pixel intervals using 256 filters having a size of 3 × 3 × 256. The output of the fifth-3 convolutional layer is processed by two pixel intervals by one filter of 2x2 size by the fifth max pulling layer and then downsampled.

(3) Two Fully Connected Layers reduce the feature by connecting the result of the convolution layer to a plurality of nodes so as to form a one-dimensional matrix. The first fully connected layer and the second fully connected layer may apply a dropout layer after applying a ReLU layer. In this case, the first fully connected layer may have 768 and 2048 nodes as inputs and outputs, respectively, and a dropout layer may be applied to set the output of each hidden node to 0 based on a randomly preset probability. Here, the preset probability may be 0.50. The second fully connected layer may have 2048 and 1024 nodes as input and output, respectively, and a dropout layer may be applied to set the output of each hidden node to zero based on a randomly preset probability. Here, the preset probability may be 0.50.

(4) The output layer classifies the features output from the second fully connected layer into two classes through the Softmax layer. For example, each of the two classes may be an actual feature class and a fake feature class.

Referring to FIG. 5, the forgery determination unit 140 according to an embodiment of the present invention may include a feature extractor 141, a feature selector 142, and a feature classifier 143.

The feature extractor 141 extracts features of the finger vein image using the previously trained CNN model. For example, the feature extractor 141 may extract a feature of the finger vein image input through the input layer and the convolution layer of the above-described CNN model. The finger vein recognition device 100 according to an embodiment of the present invention is different from conventional feature extraction methods such as LBP, BSIF, windowed dynamic mode decomposition (DMD), pyramid decomposition, Fourier descriptor, wavelet descriptor, etc. By extracting the feature through the generated CNN model, it is possible to extract a feature suitable for determining whether the finger vein image is forged.

The feature selector 142 selects an optimal feature through Principal Component Analysis (PCA) on the extracted feature of the finger vein image. The features of the finger vein images extracted from the CNN model have a very high dimension (about 4,000 components). Accordingly, the feature selector 142 may select the optimal feature by reducing the dimension of the finger vein feature.

The feature classifier 143 learns the selected optimal feature through a support vector machine (SVM) and classifies the selected feature into an actual feature class or a fake feature class. The feature classifier 143 classifies features by applying three kinds of SVM kernel functions to equation (1), as shown in equations (2) to (4). Equation (2) represents a linear kernel, Equation (3) represents a radial basic function (RBF kernel), and Equation (4) represents a polynomial kernel.

The finger vein recognition unit 150 performs finger vein recognition when the extracted feature is determined to be an actual vein feature.

6 to 7 are diagrams for explaining a finger vein recognition method for determining whether a forgery using a convolutional neural network (CNN) model for feature extraction of finger vein images according to an embodiment of the present invention.

Referring to FIG. 6, in operation S710, the finger vein recognition device 100 receives a finger vein image generated by photographing a vein pattern inside a finger.

In operation S720, the finger vein recognition apparatus 100 sets a region of interest in the input finger vein image and normalizes the size and pixel brightness of the set region of interest.

Referring to FIG. 7, the finger vein recognition device 100 receives an image of (a) a real finger vein image or (b) an image of a fake finger vein image, removes a background region, and sets a region of interest to pre-specify the pixel. Can be normalized to size (c) or (d). In addition, the finger vein recognition apparatus 100 may normalize pixel brightness of an image by applying zero mean normalization to a normalized finger vein image having a size. For example, the finger vein recognition device 100 may normalize the pixel brightness between 0 and 255 to be within a range between -127 and 127 or between -1 and 1.

In operation S730, the finger vein recognition device 100 extracts features of the finger vein image using the previously trained CNN model. The CNN model is the same as described above with reference to FIGS. 4 and 5.

In operation S740, the finger vein recognition device 100 selects an optimal feature from a feature of the extracted finger vein image through principal component analysis. For example, the finger vein recognition device 100 may select an optimal feature by reducing the dimension of the finger vein feature.

In step S750, the finger vein recognition apparatus 100 learns the selected optimal feature through a support vector machine and classifies it into an actual feature class or a fake feature class. At this time, the finger vein recognition device 100 classifies an optimal feature into an actual feature or a fake feature class using at least one of a linear kernel, a radial basic function, and a polynomial kernel. can do.

In operation S760, the finger vein recognition device 100 informs that the input finger vein image is forged when the optimal feature is classified as a forged feature class.

In operation S770, the finger vein recognition device 100 performs finger vein recognition when the optimal feature is classified into an actual finger vein image class.

8 is a diagram illustrating a finger vein recognition method using a convolutional neural network (CNN) model according to an embodiment of the present invention for the purpose of feature extraction and forgery determination of finger vein images.

Referring to FIG. 8, in operation S810, the finger vein recognition device 100 receives a finger vein image generated by photographing a vein pattern inside a finger.

In operation S820, the finger vein recognition device 100 sets a region of interest in the input finger vein image, and normalizes the size and pixel brightness of the set region of interest to preprocess.

In operation S830, the finger vein recognition apparatus 100 determines whether the finger vein image is forged using the previously trained CNN model. The CNN model is the same as described above with reference to FIGS. 4 and 5.

In operation S840, the finger vein recognition device 100 outputs an alarm signal when the input finger vein image is classified as a fake finger vein.

In operation S850, the finger vein recognition device 100 performs finger vein recognition when the input finger vein image is classified as an actual finger vein.

9 to 13 are views for explaining a database constructed for performance evaluation of the finger vein recognition device according to an embodiment of the present invention.

The present invention used three databases for learning and verification. Referring to FIG. 9, the first database is an ISPR DB built by itself, and a total of 3,300 pieces (33 people 10 fingers x 10 times) of finger vein image data set generated by photographing real finger vein (Real Access) and It consists of a total of 7,560 printer resolution data sets created by forging a real finger vein image. Referring to FIG. 10, the fake image database is a case in which the actual finger vein image is output on A4 paper, MAT paper, or OHP film by changing the material (Material) of the printing paper, and 300dpi by different printing resolution (Printer Resolution). Sub-databases were constructed based on the output of 1200dpi and 2400dpi.

Referring to FIG. 11, the training set of the ISPR DB expanded the number of data using an artificial pixel shift and cropping method in the original finger vein image. Finally, the present invention performed learning and verification using 112,260 learning data sets and 56,160 verification data sets.

Referring to (a) of FIG. 12, the second database uses an open Idiap Full-DB, and the third database is an image generated by partially cutting out of the finger vein image of the Idiap Full-DB to configure the Idiap Cropped-DB. .

Referring to FIG. 12 (b), the training data sets of Idiap Full-DB and Idiap Cropped-DB have been expanded by using a pixel shifting and cropping method in the original finger vein image. Finally, the present invention performed additional learning and verification by using 14,640 learning data sets and 400 validation data sets.

Referring to FIG. 13, the actual finger vein image in Idiap Full-DB is as shown in (a), and the normalized image is as shown in (b). In Idiap Full-DB, the finger vein image that forged the actual finger vein image is shown in (c), and the normalized image is shown in (d).

Performance evaluation methods refer to standard methods proposed by the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC-30107), such as attack presentation classification error rate (APCER), BPCER as shown in Equations (5) to (7) below. (bona fide presentation classification error rate) and NPCER (normal presentation classification error rate) were used. In Eq. (5) and Eq. (6), N _PA and N _BF mean the number classified as counterfeit data and the number classified as actual data, respectively. RES _i is a value of 1 when the i th image is classified as counterfeit data Take a value of zero if classified as actual data. In equation (7), ACER is the average of APCER and BPCER.

14 to 16 are diagrams comparing the recognition performance error rates of the finger vein recognition method and the previously studied finger vein recognition method according to an embodiment of the present invention.

As a result of comparison, the ACER of the finger vein recognition method of the present invention described above in Figure 6 shows lower value than the ACER of the finger vein recognition method previously studied in all the experimental results, it can be confirmed that the forgery data detection performance is excellent.

Finger vein recognition method according to an embodiment of the present invention is implemented in the form of program instructions that can be executed by various computer means may be recorded on a computer readable medium. Computer-readable media may include, alone or in combination with the program instructions, data files, data structures, and the like. The program instructions recorded on the computer readable medium may be those specially designed and constructed for the present invention, or may be known and available to those skilled in the computer software arts. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tape, optical media such as CD-ROMs, DVDs, and magnetic disks, such as floppy disks. Hardware devices specially configured to store and execute program instructions such as magneto-optical media and ROM, RAM, flash memory and the like. Examples of program instructions include not only machine code generated by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.

So far, the present invention has been described with reference to the embodiments. Those skilled in the art will appreciate that the present invention can be implemented in a modified form without departing from the essential features of the present invention. Therefore, the disclosed embodiments should be considered in descriptive sense only and not for purposes of limitation. The scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the scope will be construed as being included in the present invention.

100: finger vein recognition device
110: the photographing unit
120: input unit
130: preprocessing unit
140: forgery judgment unit
150: finger vein recognition unit

Claims (16)

In the finger vein recognition device,
An input unit for receiving a finger vein image generated by photographing a vein pattern inside the finger;
A preprocessor configured to set a region of interest in the input finger vein image and normalize to a predetermined size, and to correct pixel brightness by applying zero-mean normalization to the normalized finger vein image;
A forgery determination unit that extracts features of the finger vein image using a previously learned convolutional neural network (CNN) model and determines whether the extracted features are non-forged or forged features; And
If the extracted feature is determined to be a non-forgery feature includes a finger vein recognition unit that performs finger vein recognition,
The forgery determination unit,
Finger vein image input using at least one of a first CNN model consisting of five sets of convolutional layers and two fully connected layer sets, or a second CNN model consisting of 13 sets of convolutional layers and two fully connected layer sets To determine whether or not
The first CNN model and the second CNN model further comprises an output layer for classifying features of the finger vein image as a non-forged feature class or a forged feature class.
The method of claim 1,
The forgery determination unit
Feature extraction unit for extracting features of the finger vein image using the previously learned CNN model;
A feature selector which selects a first feature from a feature of the extracted finger vein image through Principal Component Analysis; And
A finger vein recognition device including a feature classifier that learns the selected first feature through a support vector machine and classifies it into a non-forged feature class or a forged feature class.
delete The method of claim 1,
The five convolution layer sets
An image of 87 × 151 × 3 pixels is input, and a first convolution layer, a rectified linear unit layer, and normalization, which are convolved at intervals of 2 pixels using 96 filters of 11 × 11 × 3 size. A first convolution layer set including a (Normalization) layer and a first max pooling layer configured to output a feature map of 19 × 35 × 96 pixels by applying one 3 × 3 size filter at 2 pixel intervals;
A second convolution layer and a calibration linear that convolves 128 filters having a size of 5 × 5 × 96 by 2 padding at intervals of 1 pixel to a feature map of 19 × 35 × 96 pixels output from the first convolution layer set A second convolution layer set including a second max pooling layer configured to output a feature map of 9 × 17 × 128 pixels by applying a unit layer, a normalization layer, and one 3 × 3 filter at 2 pixel intervals;
A third convolutional layer for convolving 2 paddings at 1 pixel intervals using 192 filters having a 3 × 3 × 128 size to a feature map of 9 × 17 × 128 pixels output from the second convolution layer set; A third convolutional layer set composed of calibration linear unit layers;
A fourth convolution layer for convolving the padding by 1 pixel interval by 1 pixel interval using 192 filters having a size of 3 × 3 × 256 to a feature map of 9 × 17 × 192 pixels output from the third convolution layer set; A fourth convolutional layer set composed of calibration linear unit layers; And
A fifth convolutional layer for correcting a 9 × 17 × 192 pixel feature map output from the fourth convolution layer set by condensing 128 filters having a size of 3 × 3 × 256 by 1 padding at intervals of one pixel A convolution layer set comprising a third max pooling layer configured to output a feature map of 4 × 8 × 128 pixels by applying a unit layer, a normalization layer, and one 3 × 3 filter at 2 pixel intervals; Finger vein recognition device.
The method of claim 1,
The two sets of fully connected layers
A first fully connected layer with 768 and 2048 nodes as inputs and outputs, respectively;
A second fully connected layer with 2048 and 1024 nodes as input and output, respectively,
The first fully connected layer and the second fully connected layer
A vein recognition device that applies a dropout layer to set the output of each hidden node to zero based on a 0.50 probability.
The method of claim 1,
The 13 convolution layer sets
Rectified Linear Unit (1-1) Convolutional Layer (128-256 × 3 pixels), 1-1 convolutional layer convoluted by 1 pixel interval using 32 filters of 3 × 3 × 3 size Unit) a first convolution layer set composed of layers;
A 1-2 convolution layer for convoluting 32 filters having a 3 × 3 × 32 size by 1 padding on a feature map of 128 × 256 × 32 pixels output from the first convolution layer set by 1 pixel intervals, A second convolution layer set consisting of a first max pulling layer for outputting a feature map of 64 × 128 × 32 pixels by applying a calibration linear unit layer and one 2 × 2 filter at 2 pixel intervals;
2-1 convolution by convolving 1 padding by 1 pixel interval using 64 filters of 3 × 3 × 32 size to a feature map of 64 × 128 × 64 pixels output from the second convolution layer set A third convolution layer set consisting of a layer, a rectified linear unit layer; And
A second-2 convolution layer for convoluting 64 filters having a 3 × 3 × 64 size by 1 padding on a feature map of 64 × 128 × 64 pixels output from the third convolution layer set by 1 pixel intervals, A designation containing a fourth set of convolutional layers consisting of a calibration linear unit layer and a second max pooling layer that outputs a feature map of 32 × 64 × 64 pixels by applying one 2 × 2 filter at 2-pixel intervals. Mac recognition device.
The method of claim 1,
The 13 convolution layer sets
A 3-1 convolution layer convolving 1 padding every 1 pixel interval using 128 3 × 3 × 64 sized filters on a feature map of 32 × 64 × 64 pixels output from the fourth convolution layer set And a fifth convolutional layer set composed of correction linear unit layers;
A 3-2 convolution layer which convolves one padding by one pixel interval using 128 3 × 128 × 128 filters on a feature map of 32 × 64 × 128 pixels output from the fifth convolution layer And a sixth convolutional layer set composed of correction linear unit layers; And
A third-3 convolution layer that convolves 128 filters having a size of 3 × 3 × 128 by one padding at intervals of one pixel on a feature map of 32 × 64 × 128 pixels output from the sixth convolution layer set; And a seventh convolution layer set consisting of a third max pooling layer that outputs a feature map of 16 × 32 × 128 pixels by applying a calibration linear unit layer and one 2 × 2 filter at 2 pixel intervals. Finger vein recognition device.
The method of claim 1,
The 13 convolution layer sets
A 4-1 convolution layer convolving 1 padding every 1 pixel interval using 256 filters having a size of 3 × 3 × 128 to a feature map of 16 × 32 × 128 pixels output from the seventh convolution layer; and An eighth convolutional layer set composed of calibration linear unit layers;
A 4-2 convolution layer convolving 1 padding every 1 pixel interval using 256 filters having a size of 3 × 3 × 256 to a feature map of the 16 × 32 × 256 pixel output from the eighth convolution layer And a ninth convolution layer set composed of a calibration linear unit layer; And
A 4-3 convolution layer for convoluting 256 filters having a size of 3 × 3 × 256 by 1 padding on a feature map of 16 × 32 × 256 pixels output from the ninth convolution layer set by 1 pixel intervals, And a tenth convolution layer set consisting of a fourth max pooling layer for applying a calibration linear unit layer and one 2 × 2 filter at 2 pixel intervals to output a feature map of 8 × 16 × 256 pixels. Finger vein recognition device.
The method of claim 1,
The 13 convolution layer sets
A 5-1 convolutional layer for convoluting 1 padding every 1 pixel interval using 256 filters having a size of 3 × 3 × 256 to a feature map of 8 × 16 × 256 pixels output from the 10th convolution layer; An eleventh convolution layer set consisting of a calibration linear unit layer;
A 5-2 convolution layer which convolutionizes the feature map of the 8 × 16 × 256 pixel output from the eleventh convolution layer by 1 padding at intervals of 1 pixel by using 256 filters having a size of 3 × 3 × 256 And a twelfth convolution layer set composed of a calibration linear unit layer; And
A fifth-3 convolution layer for convoluting 256 filters having a size of 3 × 3 × 256 by 1 padding on a feature map of the 8 × 16 × 256 pixel output from the twelfth convolution layer set by 1 pixel intervals, And a thirteenth convolution layer set comprising a fifth max pooling layer for outputting a feature map of 4 × 8 × 256 pixels by applying a calibration linear unit layer and one 2 × 2 filter at 2 pixel intervals. Finger vein recognition device.
In the finger vein recognition device recognizes the finger vein,
Receiving a finger vein image generated by photographing a vein pattern inside the finger;
Setting a region of interest in the input finger vein image and normalizing the size and pixel brightness of the region of interest;
Extracting features of the finger vein image using a previously learned convolutional neural network (CNN) model;
Selecting a first feature from a feature of the extracted finger vein image through Principal Component Analysis;
Learning the selected first feature through a support vector machine and classifying the selected first feature into a non-forged feature class or a forged feature class;
Outputting an alarm signal when the first feature is classified as a forged feature class; And
Performing finger vein recognition when the first feature is classified as a feature class that is not forged,
Classifying into the non-forged feature class or the forged feature class,
An output layer that categorizes as a non-forged feature class or a forged feature class.
Finger vein image input using at least one of a first CNN model consisting of five sets of convolutional layers and two fully connected layer sets, or a second CNN model consisting of 13 sets of convolutional layers and two fully connected layer sets Finger vein recognition method further comprising the step of determining whether the forgery.
In the finger vein recognition device recognizes the finger vein,
Receiving a finger vein image generated by photographing a vein pattern inside the finger;
Setting a region of interest in the input finger vein image and normalizing the size and pixel brightness of the region of interest;
Extracting features of the finger vein image based on the previously learned CNN model, and classifying the extracted features into non-forged feature classes or forged feature classes;
Outputting an alarm signal when the first feature is classified as a forged feature class; And
Performing finger vein recognition when the first feature is classified as a feature class that is not forged,
Classifying into the non-forged feature class or the forged feature class,
An output layer that categorizes as a non-forged feature class or a forged feature class.
Finger vein image input using at least one of a first CNN model consisting of five sets of convolutional layers and two fully connected layer sets, or a second CNN model consisting of 13 sets of convolutional layers and two fully connected layer sets Finger vein recognition method further comprising the step of determining whether the forgery.
delete The method according to any one of claims 10 and 11,
The step of setting a region of interest in the input finger vein image, normalizing the size and pixel brightness of the set region of interest, pre-processing,
Set the region of interest to normalize to a predetermined size and apply zero-mean normalization to the normalized finger vein image to reduce the pixel brightness between 0 and 255 to between -127 and 127 or between -1 and 1. Finger vein recognition method that corrects to be in the range of.
The method of claim 10,
Learning the selected first feature through a support vector machine and classifying the selected first feature into a non-forged feature class or a forged feature class,
A finger vein recognition method for classifying the first feature into a non-forged feature class or a forged feature class using at least one of a linear kernel, a radial basic function, and a polynomial kernel.
A computer program which executes the finger vein recognition method of any one of claims 10, 11 and 14 and is recorded on a computer-readable recording medium.
A computer program that executes the finger vein recognition method of claim 13 and is recorded on a computer-readable recording medium.

Images