Face recognition system and method with parallel convolutional neural network for mode selection
By using parallel convolutional neural networks to process facial images with different facial features, extracting feature vectors and binding them with IDs, the problem of insufficient accuracy in existing facial recognition technologies is solved, and higher recognition accuracy is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GENKITEK TECH CO LTD
- Filing Date
- 2021-12-02
- Publication Date
- 2026-07-21
AI Technical Summary
Existing convolutional neural networks struggle to accurately match feature vectors in face recognition due to the variability of face images, thus affecting recognition performance.
Parallel convolutional neural networks are used to extract feature vectors for different facial features (such as front, left, right, bottom, gender, and age) through multiple convolutional neural network units, and bind them to specific IDs to establish parallel convolutional neural network modules and database modules for pattern selection and ID judgment.
It improves the accuracy of face recognition, reduces noise interference, enhances the similarity of feature vector extraction, and improves the precision of recognition.
Smart Images
Figure CN116229529B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a system and method for face recognition, and more particularly to a system and method for face recognition that uses several parallel convolutional neural networks (CNNs) for pattern selection. Background Technology
[0002] In interpersonal communication, the basic judgment for getting to know and interact with each other is based on a person's appearance and personality traits. Regarding facial recognition, the human brain has a complex neural network. With training from a young age, it becomes remarkably easy to distinguish different people by their faces as one reaches a certain age. Therefore, this ability is used by humans to handle certain tasks, such as security guarding or access control.
[0003] Inspired by human abilities, machine learning has expanded from simple image learning to facial recognition and is applied to automated management. Convolutional Neural Networks (CNNs) are a type of deep learning algorithm commonly used for image recognition. Although the complete CNN algorithm is relatively new, its effectiveness and the large amount of research and development invested in it have led to its wide range of applications. Generally, CNNs in facial recognition learn from a large number of facial images to recognize and distinguish different people. When a new facial image is received, it can use existing data to determine whether the person in the image has been previously recognized by the CNN, thus achieving facial recognition. However, due to differences in head direction, posture, age, facial expressions, hair color, background, and accessories, even for the same person, the large dataset used for learning will contain significant variations between images. Therefore, even in the feature extraction stage for the same person within a CNN system, the extracted feature vectors will differ greatly, making the feature vector matching stage ineffective and limiting the power of the CNN.
[0004] Pre-classifying a training dataset of facial images of the same person into different subgroups with similar settings can reduce noise from other subgroups and improve similarity during feature vector extraction. When all subgroups point to the same person, the accuracy of the resulting facial recognition increases. However, there are no relevant technological inspirations in the existing field. Summary of the Invention
[0005] This paragraph extracts and compiles certain features of the invention. Other features will be disclosed in subsequent paragraphs. Its purpose is to cover various modifications and similar arrangements within the spirit and scope of the appended claims.
[0006] To meet the above requirements, this invention discloses a face recognition system that uses parallel convolutional neural networks for pattern selection, which is installed and run on a computer. It includes: a parallel convolutional neural network module comprising multiple convolutional neural network units, each convolutional neural network unit extracting a feature vector for one or more received face images with a facial feature, and binding the feature vector to a specific ID; a mode selection module receiving multiple frames of face images from external input, determining the facial feature of each face image, and transmitting each face image to at least one convolutional neural network unit corresponding to its facial feature; a database module storing the feature vectors generated by different convolutional neural network units and the bound IDs; and an ID determination module performing the following operations in a testing phase: a. receiving at least one test feature vector extracted from at least one convolutional neural network unit for a test face image, and comparing the at least one test feature vector with the feature vector of the corresponding convolutional neural network unit stored in the database module; b. selecting the ID bound to the feature vector closest to the at least one test feature vector as the voting ID; and c. calculating the most frequently occurring among all voting IDs. In a learning phase, the pattern selection module is for each specific person with a corresponding ID. At least one frame of facial image of the specific person is input into at least one convolutional neural network unit according to facial features to build the content of the database module.
[0007] This invention also discloses a face recognition method using parallel convolutional neural networks for pattern selection, which can be applied to a computer. The method includes the following steps: establishing a parallel convolutional neural network module, the parallel convolutional neural network module including multiple convolutional neural network units, each convolutional neural network unit extracting a feature vector for one or more received face images with a facial feature, and binding the feature vector to a specific ID; in a learning phase, a pattern selection module, for each specific person with a corresponding ID, according to facial feature judgment, feeds at least one frame of the received external input face image of the specific person into at least one of the multiple related convolutional neural network units, so as to establish the generated feature vector and the bound ID in a database module; and in a testing phase, the pattern selection module judges the facial feature of a test face image, transmits the test face image to at least one convolutional neural network unit corresponding to its facial feature, and the ID judgment module performs the following operations: a. receiving at least one test feature vector extracted from at least one convolutional neural network unit of a test face image, and comparing the at least one test feature vector with the feature vector of the corresponding convolutional neural network unit stored in the database module; b. Select the ID bound to the feature vector that is closest to the at least one test feature vector as the voting ID; and c. Calculate the most frequently occurring voting ID among all voting IDs.
[0008] According to the present invention, the facial features can be the front of the face, the left side of the face, the right side of the face, the bottom of the face, gender, or age group.
[0009] In one embodiment, the mode selection module can further adjust the weight value of each relevant convolutional neural network unit for the received test face image, so that the ID determination module uses these weight values to adjust the weight of each vote ID respectively, so that the vote ID from the relevant convolutional neural network unit may not be calculated once.
[0010] In another embodiment, if every corresponding element between a convolutional neural network unit's feature vectors generated and stored in the database module and the test feature vector is different, or the distance between them is higher than a preset value, then the ID determination module determines that no voted ID comes from the convolutional neural network unit.
[0011] Ideally, the face image can be a frame from an image file, a frame from a video file, a frame from a streaming video, or a captured image from any of the aforementioned parts.
[0012] By training a face image using several convolutional neural network units targeting different facial features of the same person, noise from other convolutional neural network units can be reduced, and the similarity during the feature vector extraction stage can be improved. This results in increased accuracy in face recognition. Attached Figure Description
[0013] Figure 1 This is a component diagram of a face recognition system using a parallel convolutional neural network for pattern selection according to an embodiment of the present invention.
[0014] Figure 2 This is a schematic diagram of the operation of the face recognition system that uses parallel convolutional neural networks for pattern selection during a learning phase.
[0015] Figure 3 This is a schematic diagram of the operation of the face recognition system that uses parallel convolutional neural networks for pattern selection during a test phase.
[0016] Figure 4 The process of calculating the vote ID;
[0017] Figure 5 This is a flowchart illustrating a face recognition method using a parallel convolutional neural network for pattern selection, as described in an embodiment of the present invention.
[0018] Explanation of reference numerals in the attached figures
[0019] 1. Computer
[0020] 2. Face Image Learning Resource Database
[0021] 3. Test face image source
[0022] 4 Output devices
[0023] 10. Input / Output Interface
[0024] 11 Mode Selection Module
[0025] 12 Parallel Convolutional Neural Network Modules
[0026] 121 First Convolutional Neural Network Unit
[0027] 1211 First eigenvector
[0028] 1212 First eigenvector
[0029] 121A First Test Feature Vector
[0030] 122 Second Convolutional Neural Network Unit
[0031] 1221 Second eigenvector
[0032] 1222 Second eigenvector
[0033] 123 Third Convolutional Neural Network Unit
[0034] 1231 Third eigenvector
[0035] 1233 Third eigenvector
[0036] 123A Third Test Feature Vector
[0037] 124 Fourth Convolutional Neural Network Unit
[0038] 1241 Fourth eigenvector
[0039] 125 Fifth Convolutional Neural Network Unit
[0040] 1251 Fifth eigenvector
[0041] 1252 Fifth eigenvector
[0042] 125A Fifth Test Feature Vector
[0043] 13 Database Module
[0044] 14 ID judgment module
[0045] A 45-degree frontal photo
[0046] B. Frontal photo with head down
[0047] C. Frontal view (eye-level photo)
[0048] D. Right side view photo
[0049] E Left side view photo
[0050] F Hair dyed front view
[0051] G's youthful photos
[0052] H test face image
[0053] Steps S01~S03. Detailed Implementation
[0054] The present invention will be described in more detail with reference to the following embodiments.
[0055] Please see Figure 1 This figure is a component diagram of a face recognition system using parallel convolutional neural networks for pattern selection according to an embodiment of the present invention (hereinafter referred to as the "system"). The system can be installed in a computer 1 and operates through the hardware architecture of the computer 1. The computer 1 on which the system is installed can be, but is not limited to, a server, industrial computer, desktop computer, or laptop computer; its hardware architecture may include a central processing unit, memory, storage devices (such as hard disks), input / output units, etc., but preferably with a larger amount of memory, such as 32GB, and the central processing unit should preferably also be able to perform multitasking and parallel processing with high processing speed. These hardware components are not shown in the figure. Figure 1 However, this is an architecture that those skilled in the art of computer science should understand. Furthermore, one of the important hardware components of computer 1 is the input / output interface 10, which is the assembly of crucial software and firmware for connecting computer 1 to external hardware via a network or data bus. This assembly may include network cards, connection cables, wireless communication modules, and other hardware. The various modules described below are technical elements of this system that operate using or in conjunction with the aforementioned existing computer 1 equipment. Therefore, they can be software, including specific program code and data, running under an operating system within at least a portion of the hardware architecture (e.g., program code and related data files are stored in a storage device, temporarily stored in memory under the operation of the operating system, and dynamically invoked and executed by the central processing unit). On the other hand, these modules can also be specialized hardware, such as application-specific integrated circuits (ASICs) or external cards, used to perform the functions assigned to these modules. Moreover, these technical elements can be partly software and partly hardware, effectively integrated according to the needs of product designers, all within the scope of the technology claimed in this patent.
[0056] This system includes a pattern selection module 11, a parallel convolutional neural network module 12, a database module 13, and an ID determination module 14. First, let's introduce the parallel convolutional neural network module 12. The parallel convolutional neural network module 12 includes several independently operating convolutional neural network units. Similar to the architecture of a typical convolutional neural network, each convolutional neural network unit has at least one convolutional layer to extract image features. After dimensionality reduction through linear rectified layers and pooling layers, the learning or testing results are output through fully connected layers. A feature of this invention is that each convolutional neural network unit learns or tests face images for a specific facial feature; the parallel convolutional neural network module 12 is not a parallel computational structure processing face images with the same features. Here, facial features refer to the different classifications of face images learned by the convolutional neural network units, based on the image capture orientation and facial features. According to this invention, facial features can be frontal face, left side of face, right side of face, lower face, gender, or age group. Not all facial features need to be applied to this system; generally, only the frontal, left, and right sides of the face are required. Each convolutional neural network unit (CNN) can extract a feature vector for one or more received face images with a single facial feature (whether for learning or testing), and bind this feature vector to a specific ID. It should be noted that each CNN can extract feature vectors from a single face image (for learning or testing) or from multiple face images of the same person (typically for learning). Binding the feature vector to a specific ID is used as the output of the learning result, and this specific ID belongs to a specific person. In other words, after each CNN has completed learning for all face images of a specific person with a single facial feature, the resulting unique feature vector will be bound to all IDs of that specific person. In this embodiment, the parallel convolutional neural network module 12 includes five convolutional neural network units: the first convolutional neural network unit 121 processes frontal face images, the second convolutional neural network unit 122 processes right-side face images, the third convolutional neural network unit 123 processes left-side face images, the fourth convolutional neural network unit 124 processes adolescent face images, and the fifth convolutional neural network unit 125 processes male-feature face images. The operation of the parallel convolutional neural network module 12 will be described in detail later.
[0057] The pattern selection module 11 is a selector that receives multiple frames of face images from an external source (such as a face image learning database 2) and determines the facial features of each face image. In other words, the pattern selection module 11 determines which facial feature (frontal, left-side, etc.) each face image has and sends each face image to at least one convolutional neural network unit corresponding to its facial feature. More specifically, each face image does not necessarily have only one facial feature; for example, a 45-degree frontal photo is also a 45-degree left-side photo. Therefore, the file of this photo can be simultaneously input into the first convolutional neural network unit 121 and the third convolutional neural network unit 123 for learning or testing. It is important to note that, regardless of the file format, the face image input to the convolutional neural network unit can be a color image, a grayscale image, or even a sketch, as long as the face image can represent a specific person and present the facial features of that specific person, allowing an ordinary person to recognize that specific person at a glance.
[0058] Database module 13 is the technical component used by this system to store learning results. Database module 13 can store the feature vectors generated by different convolutional neural network units and their associated IDs.
[0059] The system operates in two phases: a learning phase and a testing phase. The learning phase involves studying multiple facial images of multiple individuals, while the testing phase uses the learning results to determine whether a subsequent test facial image belongs to a person with a specific ID in database module 13. For the assignments in the learning phase, please refer to [link to assignment]. Figure 2 This diagram illustrates the system's operation during the learning phase. In short, during the learning phase, the pattern selection module 11 inputs at least one frame of a face image of each specific person with a corresponding ID into at least one associated convolutional neural network unit based on facial features, to build the content of the database module 13. For example... Figure 2As illustrated, the facial image learning database 2 (which could also be a photo collection stored on a storage device, but categorized according to different people) provides multiple facial images of many people. Taking person A as an example, database module 13 contains his 45-degree frontal photo A, frontal photo with head down B, frontal photo with eye level C, right side photo D, left side photo E, frontal photo with dyed hair F, and frontal photo from his youth G. The pattern selection module 11 categorizes these photos according to facial features. The 45-degree frontal photo A is input into the third convolutional neural network unit 123; the frontal photo with the head down (B) is input into the first convolutional neural network unit 121; the frontal photo looking straight ahead (C) is input into the first convolutional neural network unit 121 and the fifth convolutional neural network unit 125; the right-side photo D is input into the second convolutional neural network unit 122; the left-side photo E is input into the third convolutional neural network unit 123; the frontal photo with dyed hair (F) is input into the fifth convolutional neural network unit 125; and the frontal photo from youth (G) is input into the first convolutional neural network unit 121, the fourth convolutional neural network unit 124, and the fifth convolutional neural network unit 125. The five convolutional neural network units learn from the input data and output the learning results as the first feature vector 1211, the second feature vector 1221, the third feature vector 1231, the fourth feature vector 1241, and the fifth feature vector 1251, respectively. According to the present invention, the form of the feature vector is not limited; it can be a series of binary, octal, decimal, or hexadecimal numbers representing the features, and the number varies depending on the level of precision. In this embodiment, please see the illustration of the process for calculating the vote ID. Figure 4 For simplicity, each feature vector is represented by 8 bits of 0s and 1s. The aforementioned five feature vectors are each bound to ID 00001, representing person A. Next, the parallel convolutional neural network module 12 establishes the feature vectors generated from the learning process and their bound IDs in the database module 13. Afterwards, the system continues learning from other people's facial images during the learning phase. However, not every ID is bound to all five of its respective feature vectors. For example, ID 00002 does not have a facial image from adolescence provided for learning; therefore, ID 00002 only has its bound first feature vector 1212, second feature vector 1222, third feature vector 1232, and fifth feature vector 1252. These learned results can be applied to access control systems for people entering and exiting the premises.
[0060] For the tasks in the testing phase of this system, please refer to Figure 3This diagram illustrates the system's operation during the testing phase. During testing, an external test face image source 3, such as a streaming video monitor, transmits a test face image H. This test face image H is then used by the system to determine whether it belongs to a specific person with an ID in the database module 13. This process begins with the pattern selection module 11, which determines the facial features of the test face image H and transmits it to at least one convolutional neural network unit corresponding to those features. In this embodiment, the test face image H is a 15-degree frontal head-up image, selected by the pattern selection module 11 and input to the first convolutional neural network unit 121, the third convolutional neural network unit 123, and the fifth convolutional neural network unit 125. Next, the aforementioned three convolutional neural network units calculate the first test feature vector 121A, the third test feature vector 123A, and the fifth test feature vector 125A, respectively, based on the test face image H. Then, the ID determination module 14 performs the following operations. Assignment a: Receive at least one test feature vector extracted from at least one convolutional neural network unit about the test face image H, and compare this at least one test feature vector with the feature vector of the corresponding convolutional neural network unit stored in database module 13. For an explanation of this step, please see [link to documentation]. Figure 4 . Figure 4 It uses a tabular format to explain the process of calculating the vote ID, showing how feature vectors are compared. Figure 4 In the first test feature vector 121A, which is [1, 0, 1, 1, 0, 0, 0, 1], it is compared one by one with all related first feature vectors 1211, 1212, etc. Similarly, the third test feature vector 123A and the fifth test feature vector 125A do the same. Assignment b: Select the ID bound to the feature vector closest to the at least one test feature vector as the voting ID. According to the present invention, the closeness between two feature vectors can be represented by the distance between them. In this embodiment, the distance is displayed by the difference in the corresponding content elements of the two feature vectors. For example, if the first test feature vector 121A and the first feature vector 1211 are different at the 4th bit, then the distance is 1; if the fifth test feature vector 125A and the fifth feature vector 1211 are different at the 1st and 8th bits, then the distance is 2. However, since other carry methods can be used to represent feature vectors, the distance can also use Euclidean distance in multi-space. According to... Figure 4 As a result, the ID judgment module 14 determined that the first test feature vector 121A was closest to the first feature vector 1211, and selected ID00001 as the voting ID; the third test feature vector 123A was closest to a certain third feature vector (not shown in...). Figure 4In the table, the closest match is found to ID 00008, which is bound to the third feature vector. Therefore, ID 00001 was selected as the voting ID. The fifth test feature vector 125A is closest to the fifth feature vector 1251, so ID 00001 was selected as the voting ID. Finally, step c: calculate the most frequently occurring voting ID among all voting IDs. Figure 4 It is known that ID00001 was used as the voting ID twice, more than ID00008. Therefore, the test face image H belongs to the person represented by ID00001. The test results of this system can be output to an output device 4, such as an LCD screen, through the input / output interface 10. If this system is applied to access control, the person is accepted by the system and can be allowed to pass.
[0061] The above examples illustrate that the operational results of each convolutional neural network unit are equivalent, and therefore their voting IDs are also equivalent. However, in some applications, such as when someone likes to wear makeup, the accuracy of the calculation results of the test face image obtained from them may differ across different convolutional neural network units. Therefore, this problem can be solved by assigning different weight values to the voting IDs of different convolutional neural network units; the more accurate the convolutional neural network unit's voting ID, the higher the weight value it receives. According to the present invention, the pattern selection module 11 can further adjust the weight value of each relevant convolutional neural network unit for the received test face image, so that the ID judgment module 14 uses these weight values to adjust the weight of each voting ID respectively, ensuring that the number of times the voting ID from the relevant convolutional neural network unit is calculated is not necessarily 1. Figure 4 To illustrate with an example, if the first convolutional neural network unit 121 is given a voting ID weight of 0.8, the third convolutional neural network unit 123 is given a voting ID weight of 1.5, and the fifth convolutional neural network unit 125 is given a voting ID weight of 0.6, then the number of times DI00001 is calculated as the voting ID becomes 1.4, while the number of times DI00008 is calculated as the voting ID becomes 1.5. The result is that the test face image H belongs to the person represented by ID00008.
[0062] Furthermore, not every operation result of a convolutional neural network unit needs to be calculated as a vote. This is because if the content elements of the two feature vectors being compared are different from each other, or if the difference between them exceeds a preset tolerance range, it is easy to cause misjudgment (that is, forcing the selection of the closest face among a group of dissimilar face images). The system can handle this by determining that if every corresponding element between the feature vectors generated by a convolutional neural network unit and the test feature vector H is different from all other feature vectors in the database module 13, or if the distance between them is greater than a preset value, then the ID determination module 14 determines that no voted ID comes from that convolutional neural network unit.
[0063] It should be noted that the face image (including test face image) mentioned in this invention can be a frame of an image file, an audio-visual file, a frame of a streaming video, or a partial image captured from any of the aforementioned sources.
[0064] Based on the foregoing description of this system, a method for face recognition using parallel convolutional neural networks for pattern selection (hereinafter referred to as this method) can be derived. This method can be applied to a computer. Please see... Figure 5 , Figure 5 This is a flowchart of the method. As shown in the flowchart, the method includes the following steps: establishing a parallel convolutional neural network module, which includes multiple convolutional neural network units. Each convolutional neural network unit extracts a feature vector for one or more received face images with a facial feature, and binds the feature vector to a specific ID; in a learning phase, a pattern selection module, for each specific person with a corresponding ID, determines the facial feature and feeds at least one frame of the received external input face image of that specific person to at least one of the multiple related convolutional neural network units, based on the facial feature judgment, to establish the generated feature vector and the bound ID in a database module; and in a testing phase, the pattern selection module determines the facial feature of a test face image, transmits the test face image to at least one convolutional neural network unit corresponding to its facial feature, and performs the following operation using the ID judgment module: a. a. Receive at least one test feature vector extracted from at least one convolutional neural network unit about a test face image, and compare the at least one test feature vector with the feature vector of the corresponding convolutional neural network unit stored in the database module; b. Select the ID bound to the feature vector that is closest to the at least one test feature vector as the voting ID; and c. Calculate the most frequently occurring voting ID among all voting IDs.
[0065] As disclosed in this system, the facial features can be the front of the face, the left side of the face, the right side of the face, the bottom of the face, gender, or age group; the mode selection module can further adjust the weight value of each relevant convolutional neural network unit for the received test face image, so that the ID judgment module uses these weight values to adjust the weight of each vote ID respectively, so that the vote ID from the relevant convolutional neural network unit may be calculated multiple times; if every corresponding element between all feature vectors generated and stored in the database module by a convolutional neural network unit and the test feature vector is different, or the distance between them is higher than a preset value, then the ID judgment module determines that no vote ID comes from the convolutional neural network unit; the face image is a frame of the image portion of an image file, a frame of a streaming image, or a partial image captured from any of the above.
[0066] Although the present invention has been disclosed above with reference to embodiments, it is not intended to limit the present invention. Anyone skilled in the art can make some modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.
Claims
1. A face recognition system using parallel convolutional neural networks for pattern selection, characterized in that, Installed and running on a computer, including: A parallel convolutional neural network module includes multiple convolutional neural network units. Each convolutional neural network unit extracts a feature vector for one or more received face images with one face feature and binds the feature vector to a specific ID. A mode selection module receives multiple frames of face images from the outside, determines the facial features of each face image, and transmits each face image to at least one convolutional neural network unit corresponding to its facial features. A database module stores the feature vectors generated by different convolutional neural network units and their associated IDs; and An ID recognition module performs the following tasks during a test phase: a. Receive at least one test feature vector extracted from at least one convolutional neural network unit about a test face image, and compare the at least one test feature vector with the feature vector of the corresponding convolutional neural network unit stored in the database module; b. Select the ID bound to the feature vector that is closest to the at least one test feature vector as the voting ID; and c. Calculate the most frequently occurring ID among all voting IDs. In one learning phase, the pattern selection module is for each specific person with a corresponding ID, and at least one frame of facial image of that specific person is input into at least one related convolutional neural network unit according to facial features to build the content of the database module.
2. The face recognition system using parallel convolutional neural networks for pattern selection according to claim 1, characterized in that, This facial feature refers to the front of the face, the left side of the face, the right side of the face, the bottom of the face, and the gender or age group.
3. The face recognition system using parallel convolutional neural networks for pattern selection according to claim 1, characterized in that, The mode selection module further adjusts the weight value of each relevant convolutional neural network unit for the received test face image, so that the ID determination module uses these weight values to adjust the weight of each vote ID respectively, so that the vote ID from the relevant convolutional neural network unit may be calculated multiple times other than 1.
4. The face recognition system using parallel convolutional neural networks for pattern selection according to claim 1, characterized in that, If every corresponding element between a convolutional neural network unit's generated and stored feature vectors in the database module and the test feature vector is different, or the distance between them is higher than a preset value, then the ID judgment module determines that no voted ID comes from the convolutional neural network unit.
5. The face recognition system using parallel convolutional neural networks for pattern selection according to claim 1, characterized in that, The face image is a frame from an image file, a video file, a streaming video, or a captured image from any of the aforementioned parts.
6. A method for face recognition using parallel convolutional neural networks for pattern selection, characterized in that, Applied to a computer, including the following steps: A parallel convolutional neural network module is established, which includes multiple convolutional neural network units. Each convolutional neural network unit extracts a feature vector for one or more received face images with one face feature and binds the feature vector to a specific ID. In a learning phase, a pattern selection module, for each specific person with a corresponding ID, determines facial features and feeds at least one frame of the received external input facial image of that specific person into at least one of multiple related convolutional neural network units, so as to establish the generated feature vector and the bound ID in a database module; and In a testing phase, the mode selection module determines the facial features of a test face image, transmits the test face image to at least one convolutional neural network unit corresponding to its facial features, and the ID determination module performs the following operations: a. Receive at least one test feature vector extracted from at least one convolutional neural network unit about a test face image, and compare the at least one test feature vector with the feature vector of the corresponding convolutional neural network unit stored in the database module; b. Select the ID bound to the feature vector that is closest to the at least one test feature vector as the voting ID; and c. Calculate the most frequently occurring ID among all voting IDs.
7. The face recognition method using parallel convolutional neural networks for pattern selection according to claim 6, characterized in that, This facial feature refers to the front of the face, the left side of the face, the right side of the face, the bottom of the face, and the gender or age group.
8. The face recognition method using parallel convolutional neural networks for pattern selection according to claim 6, characterized in that, The mode selection module further adjusts the weight value of each relevant convolutional neural network unit for the received test face image, so that the ID determination module uses these weight values to adjust the weight of each vote ID respectively, so that the vote ID from the relevant convolutional neural network unit may be calculated multiple times other than 1.
9. The face recognition method using parallel convolutional neural networks for pattern selection according to claim 6, characterized in that, If every corresponding element between a convolutional neural network unit's generated and stored feature vectors in the database module and the test feature vector is different, or the distance between them is higher than a preset value, then the ID judgment module determines that no voted ID comes from the convolutional neural network unit.
10. The face recognition method using parallel convolutional neural networks for pattern selection according to claim 6, characterized in that, The face image is a frame from an image file, a video file, a streaming video, or a captured image from any of the aforementioned parts.