Face feature comparison system, face feature comparison method and cloud computing platform

Abstract

The invention provides a face feature comparison system, a face feature comparison method and a cloud computing platform, the face feature comparison system comprises a computer, a processing chip and a plurality of face feature memory banks, the plurality of face feature memory banks are respectively connected with the processing chip, and the processing chip is connected with the computer; the processing chip receives target face feature data sent by the computer, sequentially reads the face feature data of the face feature set from each face feature memory bank, and performs similarity comparison on the face feature data read into the processing chip and the target face feature data; obtaining face feature data of the first M ranks corresponding to each face feature memory bank; and the processing chip sends the face feature data of the first M ranks corresponding to each face feature memory bank to the computer, so that the computer screens the face feature data of the first N ranks in similarity from all the received face feature data to serve as a face feature comparison result.

Description

Technical Field

[0001] The present invention relates to the technical field of face recognition, and in particular to a face feature comparison system, a face feature comparison method, and a cloud computing platform. Background Art

[0002] Face recognition technology has been widely used in recent years, especially in the fields of public security, finance, intelligent security, etc. With the rapid development of big data and artificial intelligence technologies, the amount of data that face recognition systems need to process has increased exponentially. For example, in the field of public security, a large number of face images need to be compared and recognized in real time to achieve fast and accurate target positioning and monitoring.

[0003] Although significant progress has been made in face recognition technology, existing face recognition systems still face many challenges when dealing with large-scale face feature data. Traditional face recognition systems usually rely on clusters of multiple computers or high-performance GPU accelerated computing. Although these methods can improve the comparison speed to a certain extent, they have obvious defects. For example, cluster computing of multiple computers requires a large amount of hardware resources, resulting in high power consumption, high noise, and high demand for computer room area, and the construction cost and operation cost also increase significantly. In addition, although GPUs can provide high-speed computing capabilities, their problems of high power consumption and high heat generation limit their applications in some scenarios.

[0004] In large-scale face feature comparison, how to reduce the hardware power consumption while ensuring the comparison speed is the core problem faced by current face recognition technology. Summary of the Invention

[0005] The present invention provides a face feature comparison system, a face feature comparison method, and a cloud computing platform to solve the technical defects of high hardware resource power consumption and low comparison efficiency in the prior art.

[0006] The present invention provides a face feature comparison system, including: a computer, a processing chip, and multiple face feature storage bodies, where the face feature storage bodies are used to store face feature data of a face feature set; multiple of the face feature storage bodies are respectively connected to the processing chip, and the processing chip is connected to the computer; The processing chip receives the target face feature data sent by the computer, and sequentially reads the face feature data of the face feature set from each face feature storage body. The processing chip compares the face feature data read into the processing chip with the target face feature data to obtain the top M face feature data corresponding to each face feature storage body. The processing chip sends the top M face feature data corresponding to each face feature storage body to the computer, so that the computer filters out the top N face feature data with the highest similarity from all the received face feature data as the face feature comparison result.

[0007] According to the face feature comparison system provided by the present invention, the face feature storage body includes a plurality of storage chips with the same model arranged horizontally.

[0008] According to the face feature comparison system provided by the present invention, the number of the face feature storage bodies is 4, and the data capacity of each face feature storage body is greater than or equal to 16GB.

[0009] According to the face feature comparison system provided by the present invention, the face feature storage body includes 4 storage chips; the data bit positions of each storage chip are the same and are all greater than or equal to the set bit number, where the set bit number is 16bit, 32bit or 64bit.

[0010] The present invention provides a face feature comparison method for the face feature comparison system as described above. The face feature comparison method includes: Receiving, by a processing chip, the target face feature data sent by the computer; Sequentially reading, by the processing chip, the face feature data of the face feature set from each face feature storage body, and respectively comparing the face feature data read into the processing chip with the target face feature data to obtain the top M face feature data corresponding to each face feature storage body; Sending, by the processing chip, the top M face feature data corresponding to each face feature storage body to the computer, so that the computer filters out the top N face feature data with the highest similarity from all the received face feature data as the face feature comparison result.

[0011] According to the face feature comparison method provided by the present invention, the face feature data is stored in the form of vectors; intermediate vectors corresponding to each face feature storage body are pre-stored in the processing chip. The processing chip sequentially reads the face feature data of the face feature set from each face feature storage body, and respectively compares the similarity between the face feature data read into the processing chip and the target face feature data, so as to obtain the top M face feature data corresponding to each face feature storage body. Specifically, it includes: The processing chip sequentially reads the face feature data of the face feature set from each face feature storage body, and sequentially calculates the vector inner product between the face feature data read into the processing chip and the target face feature data; The processing chip sorts the vector inner products corresponding to each face feature storage body from large to small, and stores the face feature data with the top M vector inner products into the intermediate vector corresponding to the face feature storage body.

[0012] According to the face feature comparison method provided by the present invention, the processing chip sends the top M face feature data corresponding to each face feature storage body to the computer, so that the computer screens the face feature data with the top N similarity rankings from all the received face feature data as the face feature comparison result. Specifically, it includes: When the similarity comparison between the face feature data of each face feature storage body and the target face feature data is completed, the processing chip sends the top M face feature data corresponding to each face feature storage body to the computer, so that the computer sorts the face feature data stored in all the intermediate vectors again in the order of the vector inner product from large to small, and obtains the top N face feature data as the face feature comparison result.

[0013] According to the face feature comparison method provided by the present invention, the face feature data of the face feature set is generated in the following way: multiply each of the multiple initial face feature data in the face feature set by 10000 and then take the integer to obtain the corresponding face feature data; wherein, the initial face feature data is represented by P 32-bit floating-point numbers, and the face feature data is represented by P 16-bit integers.

[0014] According to the face feature comparison method provided by the present invention, the number of face feature storage bodies is 4, and the data capacity of each face feature storage body is greater than or equal to 16GB; the number of face feature data stored in the face feature storage body is greater than or equal to 60 million.

[0015] The present invention provides a cloud computing platform, and the cloud computing platform includes a plurality of face feature comparison systems as described above.

[0016] The face feature comparison system, face feature comparison method, and cloud computing platform provided by the present invention, through the collaborative work of a processing chip and multiple face feature storage bodies, the processing chip can parallelly read the face feature data in the multiple face feature storage bodies, and compare the similarity with the target face feature data, and send the top M face feature data corresponding to each face feature storage body to a computer; furthermore, the computer screens the top N face feature data with the highest similarity from all the received face feature data as the face feature comparison result, thereby significantly improving the comparison efficiency; and, the face feature comparison system of the present invention, through an optimized hardware architecture and data processing method, through a multi-storage body architecture and parallel processing, reduces the dependence on multiple computer clusters or high-performance GPUs, and reduces the hardware resource requirements. Brief Description of the Drawings

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0018] Figure 1 It is a schematic structural diagram of a face feature storage body provided by an embodiment of the present invention.

[0019] Figure 2 It is a schematic diagram of a face feature comparison system provided by an embodiment of the present invention.

[0020] Figure 3 It is a schematic flowchart of a face feature comparison method provided by an embodiment of the present invention.

[0021] Figure 4 It is a schematic structural diagram of a cloud computing platform provided by an embodiment of the present invention. Detailed Embodiments

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0023] First, a schematic explanation of the noun terms related to the embodiments of the present invention will be given.

[0024] Face feature storage: A hardware device for storing face feature data of a face feature set, composed of multiple storage chips, capable of storing a large amount of face feature data.

[0025] Storage chip: The basic unit that constitutes the face feature storage, used to store specific face feature data.

[0026] Processing chip: A hardware device responsible for receiving target face feature data and reading face feature data from the face feature storage for similarity comparison.

[0027] Similarity comparison: The process of calculating the similarity between face feature data and target face feature data and screening out the top N face feature data in terms of ranking.

[0028] Vector inner product: A mathematical method for measuring the similarity between two vectors, used in the present invention to calculate the similarity between face feature data and target face feature data.

[0029] Intermediate vector: A data structure used to temporarily store the similarity calculation results of the face feature data read from each storage and the target face feature data.

[0030] The face feature comparison system provided by the embodiments of the present invention has its core in efficiently completing the comparison task of large-scale face feature data through the collaborative work of the processing chip and multiple face feature storages. The face feature storage is used to store the face feature data of the face feature set, and multiple storages are respectively connected to the processing chip. After receiving the target face feature data, the processing chip sequentially reads the face feature data from each face feature storage and conducts similarity comparison, and finally screens out the top N face feature data in terms of similarity as the comparison result.

[0031] To implement the above functions, the face feature comparison system of the embodiments of the present invention includes: A computer, a processing chip, and multiple face feature storages, where the face feature storage is used to store the face feature data of the face feature set; Multiple face feature storages are respectively connected to the processing chip, and the processing chip is connected to the computer, forming a face feature comparison system structure of computing + storage + processing chip integrated computing of computer + storage body + processing chip.

[0032] The processing chip receives the target face feature data sent by the computer, and sequentially reads the face feature data of the face feature set from each face feature storage body. The processing chip compares the face feature data that has been read into the processing chip with the target face feature data to obtain the top M face feature data corresponding to each face feature storage body. The processing chip sends the top M face feature data corresponding to each face feature storage body to the computer, so that the computer can screen out the top N face feature data with the highest similarity from all the received face feature data as the face feature comparison result.

[0033] Among them, the processing chip is the core component of the face feature comparison system, responsible for receiving the target face feature data and sequentially reading the face feature data from the face feature storage body for similarity comparison. The processing chip can be an AI chip or a dedicated FPGA chip.

[0034] The specific working process of the processing chip is as follows: 1) The processing chip receives the externally input target face feature data through a specific interface. This data is usually represented in the form of a vector and contains feature values in multiple dimensions. The target face feature data can be extracted from the face image to be recognized through a face recognition algorithm or obtained from other systems or devices.

[0035] 2) The processing chip sequentially reads the face feature data from each face feature storage body according to a preset order. To improve the reading efficiency, the processing chip can adopt a parallel reading method to obtain data from multiple storage bodies simultaneously. The data in each storage body is arranged in order, and the processing chip sequentially reads the data stored in each storage body through address mapping and data bus control.

[0036] 3) The processing chip compares the face feature data that has been read into the chip with the target face feature data. In this embodiment, the vector inner product is used as the measure of similarity, and the similarity between the two vectors is evaluated by calculating their inner product. The processing chip sequentially calculates the inner product of the read face feature data and the target face feature data and stores the results in the corresponding intermediate vectors.

[0037] 4) After the similarity comparison between the face feature data of each storage body and the target face feature data is completed, the processing chip sorts the face feature data stored in all the intermediate vectors again in descending order of the vector inner product, and finally screens out the top N face feature data as the face feature comparison result. The screening process can adopt efficient sorting algorithms such as quicksort and mergesort to improve the screening efficiency.

[0038] In addition, the face feature storage body is composed of multiple storage chips with the same model arranged horizontally. This layout is conducive to improving the efficiency of data reading and facilitating the integration and expansion of hardware.

[0039] See Figure 1 , Figure 1 shows a schematic structural diagram of a face feature storage body composed of 4 storage chips adopted in an embodiment of the present invention. Among them, the address bits and timing bits of the 4 storage chips are connected together in one-to-one correspondence, while the data bits are independent of each other. The 4 storage chips communicate with external devices (such as a processing chip) through a data bus, an address bus, and a timing control signal.

[0040] The face feature data is distributed and stored in 4 storage chips, and each storage chip stores a part of the data. The 4 storage chips are arranged horizontally and connected by a high-speed data bus to ensure the fast reading and transmission of data. The capacity of each storage chip is 4GB, and the total capacity of the storage body composed of 4 storage chips is 16GB, which can store at least 60 million face feature data. This distributed storage method can improve the efficiency and reliability of data storage.

[0041] For example, each storage chip stores a 32-bit feature value, and a single parallel read / write operation can simultaneously read or write 4 32-bit feature values. If there are 4 storage chips, a single parallel read / write operation can simultaneously read or write 4×4 32-bit feature values, thereby improving the efficiency of data storage and reading.

[0042] Through the above structure and working principle, the face feature storage body composed of 4 storage chips can efficiently and reliably store and manage large-scale face feature data, providing a solid foundation for subsequent face feature comparison and recognition.

[0043] Again, the working process of the computer in the face feature comparison system of this embodiment is as follows: 1) Data preprocessing: Before performing face feature comparison, the computer preprocesses the original face image data. For example, removing noise, normalizing the image size, adjusting the image brightness and contrast, etc., to provide high-quality input data for subsequent feature extraction and comparison.

[0044] 2) Result postprocessing: After the processing chip completes face feature comparison, the computer postprocesses the result. It summarizes, analyzes, and interprets the comparison result to generate an easy-to-understand report or visualization result. For example, converting the comparison similarity score into a specific matching probability and judging whether the face matches according to the set threshold.

[0045] 3) Additionally, software algorithms can run on the computer, such as face detection algorithms, feature extraction algorithms, and comparison algorithms. These algorithms guide the processing chip to perform hardware-accelerated calculations and optimize and adjust the results. In some cases, the computer also participates in the collaborative processing of some computing tasks.

[0046] 4) For a face feature comparison system based on machine learning, the computer is responsible for training the algorithm model. It loads the original face image dataset and uses machine learning frameworks (such as TensorFlow, PyTorch, etc.) to train the neural network model to generate model parameters for face feature extraction and comparison. These model parameters will be transmitted to the processing chip to enable it to perform efficient inference calculations.

[0047] Figure 2 The figure shows a schematic diagram of a face feature comparison system according to an embodiment of the present invention. Figure 2 In the shown face feature comparison system, the number of face feature storage bodies is 4, and the data capacity of each storage body is greater than or equal to 16GB, which can meet the storage requirements of large-scale face feature data. Each storage body contains 4 storage chips, and the data bit positions of each storage chip are the same and are all greater than or equal to the set bit number (such as 16bit, 32bit, or 64bit) to ensure the accuracy and reliability of data storage.

[0048] Among them, 16 storage chips are divided into 4 groups to form 4 face feature storage bodies. DDR1 - DDR4 form the 1st face feature storage body, DDR2 - DDR8 form the 2nd face feature storage body, DDR9 - DDR12 form the 3rd face feature storage body, and DDR13 - DDR16 form the 4th face feature storage body. The hardware board completes large-capacity face feature comparison and sorting processing.

[0049] In the face feature comparison system according to the embodiment of the present invention, through the collaborative work of the processing chip and multiple face feature storage bodies, the processing chip can parallelly read the face feature data in multiple face feature storage bodies and compare the similarity with the target face feature data, and send the top M face feature data corresponding to each face feature storage body to the computer; then, the computer screens the top N face feature data with the highest similarity from all the received face feature data as the face feature comparison result, thereby significantly improving the comparison efficiency; moreover, the face feature comparison system according to the embodiment of the present invention reduces the dependence on multiple computer clusters or high-performance GPUs and reduces the hardware resource requirements through an optimized hardware architecture and data processing method, through a multi-storage body architecture and parallel processing.

[0050] The application scenarios of the face feature comparison system according to the embodiment of the present invention are extensive, for example: Public security field: In the public security field, a face feature comparison system can be used for real-time monitoring and target recognition. For example, by installing cameras in public places, face images are collected in real time and features are extracted, and then compared with the feature data stored in the face feature storage to quickly identify the target person and improve the security guarantee ability of public security.

[0051] Financial field: In the financial field, a face feature comparison system can be used for identity verification and customer recognition. For example, in bank counters or self-service devices, the identity of customers is verified through face recognition technology to ensure the security and legality of transactions.

[0052] Intelligent security field: In the intelligent security field, a face feature comparison system can be used for access control systems and visitor management. For example, through face recognition technology, the opening and closing of the access control system are controlled to ensure that only authorized personnel can enter specific areas.

[0053] Correspondingly, the embodiments of the present invention also disclose a face feature comparison method applied to the face feature comparison system of the above embodiments. The face feature comparison method described below can be mutually corresponding and referred to with the face feature comparison system described above.

[0054] See Figure 3 , the face feature comparison method of the embodiments of the present invention includes: 301. Receive the target face feature data sent by the computer through the processing chip.

[0055] The processing chip receives the target face feature data through a specific interface. This data is usually represented in the form of a vector and contains feature values in multiple dimensions. The target face feature data can be extracted from the face image to be recognized through a face recognition algorithm, or obtained from other systems or devices.

[0056] After receiving the target face feature data, the processing chip will perform preliminary parsing and verification on it to ensure the integrity and accuracy of the data. If the data is incorrect or incomplete, the processing chip will send an error signal to the computer and request to resend the data. In addition, the processing chip may also cache the received target face feature data so that subsequent similarity comparison operations can be performed more efficiently.

[0057] 302. Sequentially read the face feature data of the face feature set from each face feature storage through the processing chip, and respectively compare the face feature data read into the processing chip with the target face feature data to obtain the top M face feature data corresponding to each face feature storage.

[0058] The generation and storage of face feature data are the basic links of a face feature comparison system, directly affecting the accuracy and efficiency of comparison. The face feature data of N integers (N≥60 million) are evenly loaded into 4 memory banks through the PCIE bus, and each memory bank stores ≥16 million integer face feature data. It only needs to be loaded once without power-off.

[0059] In the embodiment of the present invention, the process of generating and storing face feature data is as follows: The N initial face feature data in the face feature set are each multiplied by 10,000 and then rounded to obtain the corresponding face feature data. The initial face feature data are represented by P 32-bit floating-point numbers, and the face feature data are represented by P 16-bit integers. This processing method can convert the face features represented by floating-point numbers into integer representation, reducing the data storage space and improving the calculation efficiency at the same time. For example, assuming the initial face feature data are {f1, f2, …, f M}, then the face feature data after processing are: {[f1×10,000], [f2×10,000], …, [f M ×10,000]} where [.] represents the rounding operation.

[0060] The face feature data are stored in face feature memory banks. The data capacity of each memory bank is greater than or equal to 16GB, which can store a large amount of face feature data. In this embodiment, the number of face feature memory banks is 4, and the number of face feature data stored in each memory bank is greater than or equal to 60 million. The models of the storage chips are the same and are arranged horizontally, facilitating data reading and management. The data bit positions of each storage chip are the same and are all greater than or equal to the set bit number (such as 16bit, 32bit or 64bit) to ensure the accuracy and reliability of data storage.

[0061] Each memory bank is composed of 4 storage chips. The capacity of each storage chip is 4GB, and the data bit length is 64bit. The 4 storage chips are connected through a high-speed data bus to ensure fast data reading and transmission. The total capacity of each memory bank is 16GB, which can store at least 60 million face feature data. The total capacity of the 4 memory banks is 64GB, which can store at least 240 million face feature data, meeting the requirements of large-scale face feature comparison.

[0062] 303. Send the top M face feature data corresponding to each face feature memory bank to the computer through the processing chip, so that the computer screens out the face feature data with the top N similarity rankings from all the received face feature data as the face feature comparison result.

[0063] The processing chip sends the top M face feature data corresponding to each face feature storage to the computer. Specifically, the processing chip sends this data to the computer through a high-speed data bus or a network interface. After receiving this data, the computer aggregates the data from different storage bodies to form a large vector containing all face feature data. Subsequently, the computer re-sorts the data in this large vector and filters out the top N face feature data in terms of similarity as the final face feature comparison result. The computer may also perform further processing on these results, such as generating reports or visual charts, so that users can more intuitively understand the comparison results. In addition, the computer may also store these results in a database for subsequent query and analysis.

[0064] Further, for step 302 above, it specifically includes: First, for each face feature storage, define an intermediate vector corresponding to this storage. The number of intermediate vectors is the same as the number of face feature storages. The intermediate vector is used to temporarily store the similarity calculation results between the face feature data read from each face feature storage and the target face feature data.

[0065] Second, the processing chip sequentially reads the face feature data of the face feature set from each of the face feature storages through the processing chip, and sequentially calculates the vector inner product between the face feature data read into the processing chip and the target face feature data.

[0066] Among them, the calculation formula of the vector inner product is: Among them, a[i] and b[i] respectively represent the feature values of the i-th dimension of the target face feature data and the face feature data in the storage body, and M represents the dimension of the feature vector.

[0067] Third, the processing chip sorts the vector inner products corresponding to each face feature storage from large to small, and stores the face feature data with the top M vector inner product rankings into the intermediate vector corresponding to the face feature storage.

[0068] Take the calculated vector inner product as the similarity, sort it from large to small, and store the face feature data with the top M rankings into the intermediate vector corresponding to this storage.

[0069] Then, when the similarity comparison between the face feature data in each face feature storage and the target face feature data is completed, the processing chip sends the top M face feature data corresponding to each face feature storage to the computer, so that the computer sorts the face feature data stored in all the intermediate vectors again in descending order of the vector inner product, and obtains the top N face feature data as the face feature comparison result.

[0070] Taking two face feature storages as an example, the process of step 302 above will be described schematically.

[0071] Define two groups of intermediate vectors c1

[50] and c2

[50] , each element is a 32-bit integer, and all are initialized to 0. The target face feature data to be compared is b

[512] .

[0072] Each time, read 4 groups of face feature data vectors from two face feature storages respectively, denoted as b11

[512] , b12

[512] , b13

[512] , b14

[512] and b21

[512] , b22

[512] , b23

[512] , b24

[512] , and then continue to read the next group of face feature data vectors. Since reading takes a certain amount of time, the inner products of the vectors of the face feature data that have been read into the processing chip and the vector b

[512] of the target face feature data can be calculated respectively while reading the vectors, obtaining 8 32-bit integer data d11, d12, d13, d14, d21, d22, d23, d24. Sort d11, d12, d13, d14 together with the elements in c1

[50] in descending order, and take out the top 50 and re-store them in c1

[50] ; sort d21, d22, d23, d24 together with the elements in c2

[50] in descending order, and take out the top 50 and re-store them in c2

[50] .

[0073] Because the calculation time is less than the time to read a group of vectors, each time a group of vectors is read, the above steps can be cycled to form a pipeline structure until all 10,000,000 vectors of the two face feature storages are calculated, and the final intermediate vector c1

[50] and the intermediate vector c2

[50] are output to the computer.

[0074] In the computer, sort the intermediate vectors c1

[50] and c2

[50] in descending order, and take out the top 50 face feature data and store them in c

[50] . c

[50] is the final face feature comparison result.

[0075] The face feature comparison method according to the embodiments of the present invention, in addition to improving the comparison efficiency and reducing the hardware resource requirements, also achieves the following beneficial effects: Reducing hardware dependence: Traditional face recognition systems usually rely on clusters of multiple computers or high-performance GPU accelerated computing. These hardware devices are not only costly, but also consume a large amount of power, generate a lot of noise, and require a large computer room area. By optimizing the hardware architecture and data processing methods, the present invention reduces the dependence on these high-performance hardware, reduces the hardware resource requirements, and thus reduces the construction cost and operation cost.

[0076] Reducing power consumption: The face feature comparison system of the present invention significantly reduces the hardware power consumption while ensuring the comparison speed through optimized hardware architecture and data processing methods. Compared with traditional clusters of multiple computers or high-performance GPUs, the system of the present invention has lower power consumption, less noise, and also requires a smaller computer room area, being more environmentally friendly and economical.

[0077] Efficient data reading and management: Face feature data is stored in the form of vectors and generated and stored through specific methods, reducing the data storage space while improving the computing efficiency. The processing chip can efficiently read face feature data from the storage and perform similarity comparison, improving the efficiency of data reading and management.

[0078] The embodiments of the present invention also provide a cloud computing platform, which includes a plurality of the above-mentioned face feature comparison systems. As Figure 4 shown, multiple face feature comparison systems 420 are connected to the face feature comparison server 410 in the cloud through the network 430.

[0079] The cloud computing platform proposed by the embodiments of the present invention places the comparison of billions of face features with large computing amounts at the front end. The face feature comparison is completed at high speed by the computing-in-memory integrated billion-level face feature comparison system, and a cloud computing platform for face feature comparison with more than 1 billion levels is formed by multiple billion-level face feature comparison systems. This structure has advantages in comparison speed, system power consumption, and system price compared with the existing centralized face feature comparison platforms with more than 1 billion in the prior art.

[0080] The cloud computing platform integrates multiple comparison systems together to form a powerful comparison cluster, which can process larger-scale face feature data. Through the cloud computing platform, users can submit face feature comparison tasks to the cloud, and multiple comparison systems in the cloud cooperate to complete the comparison tasks and finally return the comparison results.

[0081] The cloud computing platform consists of multiple nodes, and each node contains a face feature comparison system. These nodes are connected through a high-speed network to form a distributed computing cluster. Users can submit comparison tasks through the cloud service interface. The tasks are assigned to multiple nodes for parallel processing, and the final results are returned to the users through the cloud service interface. The cloud computing platform can dynamically adjust the number of nodes according to the task volume to ensure efficient processing of large-scale comparison tasks.

[0082] The advantages of the cloud computing platform include: High scalability: The cloud computing platform can dynamically expand the number of nodes according to user needs to support larger-scale face feature comparison tasks.

[0083] High availability: Through the distributed architecture, the cloud computing platform can provide high availability to ensure the continuity and stability of comparison tasks.

[0084] High security: The cloud computing platform adopts various security measures such as data encryption and access control to ensure the security and privacy of user data.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A facial feature comparison system, characterized in that: include: A computer, a processing chip and a plurality of facial feature storage bodies, wherein the facial feature storage bodies are used to store facial feature data of a facial feature set; The plurality of facial feature storage bodies are respectively connected to the processing chip, and the processing chip is connected to the computer; The processing chip receives the target facial feature data sent by the computer, and reads the facial feature data of the facial feature set from each facial feature storage body in sequence respectively; the processing chip compares the facial feature data that has been read into the processing chip with the target facial feature data in similarity to obtain the top M facial feature data corresponding to each facial feature storage body; the processing chip sends the top M facial feature data corresponding to each facial feature storage body to the computer, so that the computer selects the facial feature data with the top N similarities from all the received facial feature data as the facial feature comparison result.

2. The facial feature comparison system according to claim 1, characterized in that: The facial feature storage body includes a plurality of storage chips of the same model and arranged in a horizontal direction.

3. The facial feature comparison system according to claim 1, characterized in that: The number of the facial feature storage bodies is 4, and the data capacity of each facial feature storage body is greater than or equal to 16GB.

4. The facial feature comparison system according to claim 2, characterized in that: The facial feature storage body includes 4 storage chips; The data bits of each storage chip are the same as each other and are greater than or equal to a set number of bits, wherein the set number of bits is 16 bits, 32 bits or 64 bits.

5. A facial feature comparison method, characterized in that: Used in the facial feature comparison system according to any one of claims 1 to 4, the facial feature comparison method comprising: receiving the target facial feature data sent by the computer through the processing chip; The facial feature data of the facial feature set are sequentially read from each facial feature storage body by the processing chip, and the facial feature data read into the processing chip are respectively compared with the target facial feature data for similarity, so as to obtain the top M facial feature data corresponding to each facial feature storage body; The processing chip sends the top M facial feature data corresponding to each facial feature storage body to the computer, so that the computer selects the top N facial feature data with the highest similarity from all the received facial feature data as the facial feature comparison result.

6. The facial feature comparison method according to claim 5, characterized in that: The facial feature data is stored in the form of vectors; the processing chip pre-stores an intermediate vector corresponding to each facial feature storage volume; The facial feature data of the facial feature set are sequentially read from each facial feature storage body by the processing chip, and the facial feature data read into the processing chip are respectively compared with the target facial feature data for similarity, so as to obtain the top M facial feature data corresponding to each facial feature storage body, specifically including: The facial feature data of the facial feature set are sequentially read from each of the facial feature storage bodies by the processing chip, and the vector inner products of the facial feature data read into the processing chip and the target facial feature data are sequentially calculated; The processing chip sorts the vector inner products corresponding to each facial feature storage body from large to small, and stores the facial feature data with the top M positions of the vector inner products in the intermediate vectors corresponding to the facial feature storage bodies.

7. The facial feature comparison method according to claim 6, characterized in that: The processing chip sends the top M facial feature data corresponding to each facial feature storage to the computer, so that the computer selects the top N facial feature data with the highest similarity from all the received facial feature data as the facial feature comparison result, specifically including: When the facial feature data of each facial feature storage body is compared with the target facial feature data for similarity, the top M facial feature data corresponding to each facial feature storage body are sent to the computer through the processing chip, so that the computer sorts the facial feature data stored in all intermediate vectors again in the order of vector inner product from large to small, and obtains the top N facial feature data as the facial feature comparison result.

8. The facial feature comparison method according to claim 5, characterized in that: The facial feature data of the facial feature set is generated in the following manner: Multiple initial facial feature data in the facial feature set are multiplied by 10000 respectively, and then rounded to the nearest integer to obtain corresponding facial feature data; wherein the initial facial feature data is represented by P 32-bit floating-point numbers, and the facial feature data is represented by P 16-bit integers.

9. The facial feature comparison method according to claim 5, characterized in that: The number of the facial feature storage bodies is 4, and the data capacity of each facial feature storage body is greater than or equal to 16GB; the number of facial feature data stored in the facial feature storage body is greater than or equal to 60 million.

10. A cloud computing platform, characterized in that: The cloud computing platform includes multiple facial feature comparison systems as described in any one of claims 1-4.

Images