Face recognition engine multi-algorithm evaluation and preferential scheduling system and method

Abstract

The invention discloses a face recognition engine multi-algorithm evaluation and preferential scheduling system and a face recognition engine multi-algorithm evaluation and preferential scheduling method, which integrate key technologies of algorithm evaluation, preferential scheduling, cyclic feedback and the like and construct a set of comprehensive algorithm evaluation, weighted scheduling and feedback cyclic mechanism. Firstly, performance indexes of different face recognition engines are evaluated and analyzed through algorithms, then weights selected by the algorithms are dynamically adjusted based on algorithm evaluation results, it is ensured that the optimal face recognition engine can be selected according to service requirements, in addition, the performance indexes of the face recognition engines can be deeply analyzed and evaluated regularly, and the face recognition efficiency is improved. According to the method, the detailed face recognition engine performance report is output, a continuous optimization direction is provided for the face recognition engine, so that a benign competition environment can be constructed, continuous improvement of the face recognition engine is promoted, the integrating degree of the face recognition engine and service requirements is improved, and finally optimization of user experience is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of face recognition, and in particular to a multi-algorithm evaluation and optimal scheduling system and method for a face recognition engine. Background Art

[0002] Driven by big data and artificial intelligence (AI) technology, information processing capabilities have experienced a qualitative leap, and many innovative face recognition algorithms have been born, greatly enhancing the recognition and verification process of facial features, making it more convenient and safe. The "Internet +" Trusted Identity Authentication Platform (CTID Platform) has introduced advanced face recognition technology to provide users with efficient and accurate face matching capabilities. However, given the unique advantages and most suitable scenarios of different face recognition algorithms, how to combine the business scenarios of the CTID platform and continuously improve the fit between the algorithm and the business has become an urgent problem that needs to be solved.

[0003] The existing feedback loop technology solution generally includes three stages: data collection, data analysis, and implementation improvement. The first is the data collection stage. The goal of this stage is to collect as much relevant data as possible, including customer feedback, product performance data, and environmental adaptability. Next, enter the data analysis stage. At this time, statistical tools and analysis methods are used to conduct in-depth analysis of the collected data, identify key problem points and improvement opportunities, and provide data basis for the next step of action. Finally, it is the implementation improvement stage. According to the results of data analysis, specific improvement measures are formulated, which may include product adjustments, service optimization, or process improvements. After implementing improvements, continue this cycle to ensure continuous optimization and progress.

[0004] Although the above feedback loop solution can enable continuous product optimization and improvement, it still has the following two shortcomings. First, there is a lack of a horizontal comparison and evaluation mechanism. It only relies on existing face recognition engines for improvement. There is a lack of a horizontal comparison process among the same type of face recognition engines. It is unable to effectively identify and introduce more advanced solutions in the industry, limiting the possibility of technological innovation and performance improvement. Second, there is a lack of a competitive mechanism for preferential scheduling. Face recognition engines that do not meet business requirements will still be selected. It is unable to adapt to changes in market and user needs in a timely manner, which is not conducive to resource allocation and promoting the continuous competition and progress of face recognition engines. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention aims to provide a multi-algorithm evaluation and optimal scheduling system and method for a face recognition engine.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A multi-algorithm evaluation and optimal scheduling system for a face recognition engine includes a face recognition engine subsystem, a comprehensive evaluation subsystem, an optimal scheduling subsystem, and a business subsystem; the comprehensive evaluation subsystem includes a data acquisition module, a data analysis module, and a scoring module; the optimal scheduling subsystem includes a scheduling task generation module, a scheduling task selection module, and a scheduling execution module;

[0008] The face recognition engine in the face recognition engine subsystem will generate multi-dimensional running status data in real time during the face recognition process. The data acquisition module of the comprehensive evaluation subsystem is used to collect the running status data of all face recognition engines and transmit it to the data analysis module of the comprehensive evaluation subsystem;

[0009] The data analysis module is used to classify and categorize the operating status data of all face recognition engines, and then generate statistical data of scoring index items in multiple dimensions based on the results of the classification and classification, and pass them to the scoring module of the comprehensive evaluation subsystem; data classification is to divide the operating status data into multiple levels according to the degree of impact on the CTID platform business, and assign different classification weights; data classification is to divide the operating status data into multiple categories according to different properties, and classify data of different categories into corresponding levels; after data classification and classification, statistical data of scoring index items in multiple dimensions are generated, and the classification weight of each scoring index item is determined according to the level and category to which each scoring index item belongs;

[0010] The scoring module calculates the score of the face recognition engine on each scoring indicator item according to the preset scoring rules of each scoring indicator item. i , combined with the grading weight W of each scoring indicator item i Calculate the comprehensive score of the face recognition engine. The higher the comprehensive score, the better the service quality of the face recognition engine. The scoring module outputs the calculated comprehensive scores of each face recognition engine to the optimal scheduling subsystem.

[0011] After receiving the comprehensive scores of each face recognition engine from the comprehensive evaluation subsystem, the scheduling task generation module will calculate the scheduling weight of each face recognition engine based on the comprehensive score of the face recognition engine and the call success rate of the business subsystem, and then pre-generate the scheduling task according to the scheduling weight and put it into the work bucket; one scheduling task contains multiple face recognition engines, and one business request selects one scheduling task to process;

[0012] The business subsystem completes the call to the face recognition engine by initiating a request to the preferential scheduling subsystem; the scheduling execution module of the preferential scheduling subsystem requests a scheduling task from the scheduling task selection module; the scheduling task selection module selects a suitable scheduling task from the scheduling work bucket and returns it to the scheduling execution module; the scheduling execution module executes the various face recognition engines of the scheduling task in parallel, and integrates the results returned by each face recognition engine to finally generate a comprehensive face recognition result, and then returns this result to the business subsystem.

[0013] Furthermore, the scheduling weight calculation process is:

[0014] Set the comprehensive score of face recognition engine i to S i , the success rate T of business subsystem y calling face recognition engine i yi , the business impact coefficient is U, then the scheduling weight of face recognition engine i is W i =S i ×(1-U)+T yi ×U×100.

[0015] Furthermore, the scheduling task generation rule is:

[0016] Assume that the face recognition engine subsystem integrates n face recognition engines, and selects m face recognition engines for a scheduling task, 1<m<n, n, m are both positive integers, and the scheduling weight of the i-th face recognition engine is W i , in the initial state of a cycle, the cycle weight R of the i-th face recognition engine i =mW i , the whole cycle will be ∑R i The face recognition engine selection can generate ∑W i The probability of selecting the face recognition engine for each scheduling task When selecting the execution of a scheduling task, first select the R i Calculate the selection space K = [1, ∑R i ], set the random value function V(P i )=Random(K,1), random(K,1) means to randomly get an integer between 1 and K. If 1≤V(P i )≤R1, then face recognition engine 1 is selected. Then face recognition engine y is selected. When face recognition engine y is selected, the selection of the next face recognition engine of the scheduling task will exclude face recognition engine y, that is, the selection space K = [1, ∑R i -R y ], and R y =R y-1, at this time, the random selection of the second face recognition engine is executed again, and the m face recognition engines of this scheduling task are selected one by one and put into the work bucket, and then the generation of the next scheduling task is started, until the cycle weights of all face recognitions are returned to 0, and the execution of this cycle is completed.

[0017] Furthermore, the scoring module is also used to output an evaluation result report of the face recognition engine by horizontally comparing the scores of various indicator items of different face recognition engines, and to feed back the report to the maintainer of the face recognition engine. The maintainer will perform targeted optimization and upgrade the face recognition engine based on the feedback results.

[0018] The present invention also provides a multi-algorithm evaluation and optimal scheduling method for a face recognition engine using the above system, the specific process is as follows:

[0019] 1. Horizontal comparison and evaluation of face recognition engines:

[0020] The face recognition engine in the face recognition engine subsystem will generate running status data in real time during the face recognition process. The data collection module of the comprehensive evaluation subsystem collects the running status data of all face recognition engines and transmits it to the data analysis module of the comprehensive evaluation subsystem; the data analysis module performs classification and classification processing on the running status data of all face recognition engines, generates statistical data of multiple dimensional scoring index items according to the classification and classification processing results, and transmits them to the scoring module of the comprehensive evaluation subsystem; the scoring module calculates the scores of each scoring index item of the face recognition engine according to the preset scoring rules of each scoring index item, and then calculates the comprehensive score of the face recognition engine in combination with the classification weights of the evaluation index items, and then outputs the comprehensive score of the face recognition engine to the preferential scheduling subsystem;

[0021] 2. Face recognition engine optimal scheduling:

[0022] The scheduling task generation module of the optimal scheduling subsystem calculates the scheduling weight of the face recognition engine according to the comprehensive score of the face recognition engine, and then pre-generates the scheduling task according to the scheduling weight and puts it into the work bucket. One scheduling task contains multiple face recognition engines, and one business request selects one scheduling task to process;

[0023] When the business subsystem needs face recognition service, it initiates a request to the preferential scheduling subsystem to complete the call to the face recognition engine; the scheduling execution module of the preferential scheduling subsystem requests a scheduling task from the scheduling task selection module; the scheduling task selection module selects a suitable scheduling task from the scheduling work bucket and returns it to the scheduling execution module; the scheduling execution module executes each face recognition engine of the scheduling task in parallel, and integrates the results returned by each face recognition engine, and finally generates a comprehensive face recognition result, which is then returned to the business subsystem.

[0024] The beneficial effect of the present invention is that the present invention integrates key technologies such as algorithm evaluation, optimal scheduling and cyclic feedback to construct a comprehensive algorithm evaluation, optimal scheduling and feedback cyclic mechanism. First, the performance indicators of different face recognition engines are analyzed through algorithm evaluation, and then the weights of each algorithm selected are dynamically adjusted based on the algorithm evaluation results to ensure that the optimal face recognition engine can be selected according to business requirements. In addition, the performance indicators of each face recognition engine are regularly analyzed and evaluated, and a detailed face recognition engine performance report is output, which provides a direction for continuous optimization of the face recognition engine, thereby building a benign competition environment, promoting the continuous improvement of the face recognition engine, and improving the fit between the face recognition engine and business needs, and ultimately optimizing the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a schematic diagram of the overall framework of the system in Example 1 of the present invention;

[0026] Figure 2 This is a flow chart of horizontal comparison and evaluation of a face recognition engine in Embodiment 2 of the present invention;

[0027] Figure 3 This is a flowchart of the optimal scheduling of the face recognition engine in Example 2 of the present invention. DETAILED DESCRIPTION

[0028] The present invention will be further described below in conjunction with the accompanying drawings. It should be noted that this embodiment is based on the technical solution and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to this embodiment.

[0029] Example 1

[0030] This embodiment provides a multi-algorithm evaluation and optimization scheduling system for face recognition engines, such as Figure 1As shown, it includes a face recognition engine subsystem, a comprehensive evaluation subsystem, a preferential scheduling subsystem and a business subsystem in the CTID platform; the comprehensive evaluation subsystem includes a data acquisition module, a data analysis module and a scoring module; the preferential scheduling subsystem includes a scheduling task generation module, a scheduling task selection module and a scheduling execution module.

[0031] The face recognition engine in the face recognition engine subsystem will generate multi-dimensional operating status data such as algorithm performance, stability, and security in real time during the face recognition process. The data acquisition module of the comprehensive evaluation subsystem is used to collect the operating status data of all face recognition engines and pass it to the data analysis module of the comprehensive evaluation subsystem.

[0032] The data analysis module is used to grade and classify the operating status data of all face recognition engines, and then generate statistical data of scoring index items in multiple dimensions based on the results of the grading and classification processing, and pass them to the scoring module of the comprehensive evaluation subsystem.

[0033] Specifically, in this embodiment, data classification mainly divides the operating status data into three levels according to the degree of impact on the CTID platform business, from high to low, namely level one, level two, and level three, and assigns different classification weights. Data classification mainly divides the operating status data into multiple categories according to different properties. In this embodiment, it is specifically divided into four categories: algorithm performance, cost-effectiveness, stability and maintainability, and security, and data of different categories are classified into corresponding levels. After data classification and classification, statistical data of scoring index items in multiple dimensions are generated, and the classification weight of each scoring index item is determined according to the level and category to which each scoring index item belongs.

[0034] The detailed classification is shown in Table 1 below:

[0035] Table 1

[0036]

[0037]

[0038] The scoring module will calculate the score of the face recognition engine on each scoring indicator item according to the preset scoring rules of each scoring indicator item. i , combined with the grading weight w of each scoring indicator item i Calculate the overall score of the face recognition engine n represents the number of scoring index items of the corresponding level. The higher the comprehensive score S, the better the service quality of the face recognition engine. The scoring module outputs the calculated comprehensive score S of each face recognition engine to the optimal scheduling subsystem.

[0039] The scoring rules for the scoring index items take the face recognition accuracy index in the algorithm performance as an example. The formula is as follows:

[0040] If R < 99.99%, s = 0, otherwise:

[0041] s=(1-10000×(1-R)) 3 ×100

[0042] Among them, s represents the indicator score, and R represents the face recognition accuracy.

[0043] In this embodiment, the scoring module is also used to output the evaluation result report of each face recognition engine by comparing the scores of different face recognition engine indicators horizontally, and feedback the report to the maintainer of the face recognition engine. The maintainer performs targeted optimization and upgrades of the face recognition engine based on the feedback results. For example, for the response time indicator, assuming that the response time indicator scores of face recognition engines A and B are 80 points and 90 points respectively, through horizontal comparison, it can be seen that face recognition engine A has a lower score, indicating that the algorithm response time of face recognition engine A is relatively long. The comparison result will be reflected in the evaluation result report of face recognition engine A, and the maintainer can optimize this indicator item.

[0044] After receiving the comprehensive scores of each face recognition engine provided by the comprehensive evaluation subsystem, the scheduling task generation module will calculate the scheduling weight of each face recognition engine based on the comprehensive score of the face recognition engine and the call success rate of the business subsystem, and then pre-generate the scheduling task according to the scheduling weight and put it into the work bucket; a scheduling task contains multiple face recognition engines, and one business request selects one scheduling task to process.

[0045] In this embodiment, the scheduling weight calculation process is:

[0046] Set the comprehensive score of face recognition engine i to S i , the success rate T of business subsystem y calling face recognition engine i yi , the business impact coefficient is U, then the scheduling weight of face recognition engine i is W i =S i ×(1-U)+T yi ×U×100, considering the differences in the actual usage of each face recognition engine in different business scenarios, the business impact coefficient U is set to 0.5, and T yi =0.98, S i =90, the scheduling weight W of face recognition engine i i =90(1-0.5)+0.98×0.5×100=94.

[0047] In this embodiment, the scheduling task generation rule is:

[0048] Assume that the face recognition engine subsystem integrates n face recognition engines, and selects m face recognition engines for one scheduling task, 1<m<n (n, m are both positive integers), and the scheduling weight of the i-th face recognition engine is W i , in the initial state of a cycle, the cycle weight R of the i-th face recognition engine i =mW i , the whole cycle will be ∑R i The face recognition engine selection can generate ∑W i The probability of selecting the face recognition engine for each scheduling task When selecting the execution of a scheduling task, first select the R i Calculate the selection space K = [1, ∑R i ], set the random value function V(P i )=Random(K,1), random(K,1) means to randomly get an integer between 1 and K. If 1≤V(P i )≤R1, then face recognition engine 1 is selected. Then face recognition engine y is selected. When face recognition engine y is selected, the selection of the next face recognition engine of the scheduling task will exclude face recognition engine y, that is, the selection space K = [1, ∑R i -R y ], and R y =R y -1, at this time, the random selection of the second face recognition engine is executed again, and the m face recognition engines of this scheduling task are selected one by one and put into the work bucket, and then the generation of the next scheduling task is started, until the cycle weights of all face recognitions are returned to 0, and the execution of this cycle is completed.

[0049] For example, the face recognition engine subsystem integrates three face recognition engines, namely A1, A2, and A3. A scheduling task contains two face recognition engines. The scheduling weights of the face recognition engines are W1=2, W2=3, and W3=5, respectively. The period weights are R1=4, R2=6, and R3=10, respectively. When the scheduling task selects the first face recognition engine, K=[1,20]. If the random value is 12, the first face recognition engine selected is A3. At this time, R3=10-1=9, and K=[1,10]. At this time, the second face recognition engine is randomly selected in the same way to complete the generation of this scheduling task.

[0050] The business subsystem completes the face recognition engine call by initiating a request to the optimal scheduling subsystem. The scheduling execution module of the optimal scheduling subsystem initiates a request to the scheduling task selection module to obtain a scheduling task. The scheduling task selection module is used to select a scheduling task from the work bucket according to the first-in-first-out rule and return the selected scheduling task to the scheduling execution module.

[0051] The scheduling execution module will execute each face recognition engine in the scheduling task in parallel, and fuse the results returned by each face recognition engine to obtain the final face recognition result, and finally return the face recognition result to the business subsystem. In this embodiment, the specific process of fusion is: considering the differences in recognition performance of different face recognition engines, the recognition results of each face recognition engine are weighted and fused to obtain the final face recognition result. Set the comprehensive score of face recognition engine i to S i , the face recognition score is Z i , then the weighted weight The final weighted fusion score S = ∑(w i '×Z i ).

[0052] Example 2

[0053] This embodiment provides a multi-algorithm evaluation and optimal scheduling method for a face recognition engine using the system described in Embodiment 1. The specific process is as follows:

[0054] 1. Horizontal comparison and evaluation of face recognition engines (such as Figure 2 shown):

[0055] The face recognition engine in the face recognition engine subsystem will generate running status data in real time during the face recognition process. The data acquisition module of the comprehensive evaluation subsystem collects the running status data of all face recognition engines and transmits it to the data analysis module of the comprehensive evaluation subsystem; the data analysis module performs classification and classification processing on the running status data of all face recognition engines, generates statistical data of multiple dimensional scoring index items based on the classification and classification processing results, and transmits them to the scoring module of the comprehensive evaluation subsystem. The scoring module calculates the scores of each scoring index item of the face recognition engine according to the preset scoring rules of each scoring index item, and then calculates the comprehensive score of the face recognition engine in combination with the grading weight of the evaluation index item, and then outputs the comprehensive score of the face recognition engine to the preferential scheduling subsystem.

[0056] 2. Face recognition engine optimal scheduling (such as Figure 3 shown):

[0057] The scheduling task generation module of the optimal scheduling subsystem calculates the scheduling weight of the face recognition engine according to the comprehensive score of the face recognition engine, and then pre-generates the scheduling task according to the scheduling weight and puts it into the work bucket. One scheduling task contains multiple face recognition engines, and one business request selects one scheduling task to process;

[0058] When the business subsystem needs face recognition service, it initiates a request to the optimal scheduling subsystem to complete the call to the face recognition engine. The scheduling execution module of the optimal scheduling subsystem requests a scheduling task from the scheduling task selection module. The scheduling task selection module selects a suitable scheduling task from the scheduling work bucket and returns it to the scheduling execution module. The scheduling execution module executes each face recognition engine of the scheduling task in parallel, and integrates the results returned by each face recognition engine, and finally generates a comprehensive face recognition result, which is then returned to the business subsystem.

[0059] In the above embodiments 1 and 2, based on the optimal scheduling subsystem combined with the business impact coefficient to calculate the weight of the face recognition engine index item, the business subsystem in the CTID platform can select the face recognition engine with the highest degree of fit with the current business subsystem through the optimal scheduling subsystem to complete the face comparison business. At the same time, the comprehensive evaluation subsystem will collect the situation of each face recognition engine operation, and output the algorithm evaluation results after data analysis and horizontal comparison evaluation. The evaluation results can be used as improvement suggestions to guide the continuous optimization of the face recognition engine and the weight data support of the optimal scheduling subsystem.

[0060] Furthermore, the optimal scheduling subsystem adopts the "work bucket" technology to first calculate the weight ratio of each face recognition engine based on the algorithm evaluation results. Then, the system initializes a bucket with a fixed capacity and continuously allocates work tasks to the bucket according to the weight ratio of each engine. When the business subsystem needs to call the face recognition engine, it extracts the corresponding scheduling task from the bucket. This strategy not only ensures the fairness and transparency of the face recognition engine, but also can dynamically adjust the weight ratio of the face recognition engine according to real-time business needs and face recognition engine performance, thereby achieving more accurate matching of business needs.

[0061] The comprehensive evaluation subsystem will collect the operating data of the face recognition engine in terms of performance, stability, security and other dimensions. Then, the data will be graded and classified to generate more detailed scoring indicators. These indicators will be scored according to the preset data analysis model. Finally, by comparing the indicator scores of different face recognition engines horizontally, the system will output a detailed engine evaluation result report. This report can provide improvement suggestions for each engine, and also provide data support for the weight ratio of each face recognition engine in the preferential scheduling subsystem, thereby promoting the continuous optimization of the face recognition engine.

[0062] For those skilled in the art, various corresponding changes and modifications can be made according to the above technical solutions and concepts, and all of these changes and modifications should be included in the protection scope of the claims of the present invention.

Claims

1. A multi-algorithm evaluation and optimization scheduling system for face recognition engines, characterized in that: It includes face recognition engine subsystem, comprehensive evaluation subsystem, optimal scheduling subsystem and business subsystem; the comprehensive evaluation subsystem includes data collection module, data analysis module and scoring module; the optimal scheduling subsystem includes scheduling task generation module, scheduling task selection module and scheduling execution module; The face recognition engine in the face recognition engine subsystem will generate multi-dimensional running status data in real time during the face recognition process. The data acquisition module of the comprehensive evaluation subsystem is used to collect the running status data of all face recognition engines and transmit it to the data analysis module of the comprehensive evaluation subsystem; The data analysis module is used to classify and categorize the operating status data of all face recognition engines, and then generate statistical data of scoring index items in multiple dimensions based on the results of the classification and classification, and pass them to the scoring module of the comprehensive evaluation subsystem; data classification is to divide the operating status data into multiple levels according to the degree of impact on the CTID platform business, and assign different classification weights; data classification is to divide the operating status data into multiple categories according to different properties, and classify data of different categories into corresponding levels; after data classification and classification, statistical data of scoring index items in multiple dimensions are generated, and the classification weight of each scoring index item is determined according to the level and category to which each scoring index item belongs; The scoring module calculates the score of the face recognition engine on each scoring indicator item according to the preset scoring rules of each scoring indicator item. i , combined with the grading weight W of each scoring indicator item i Calculate the comprehensive score of the face recognition engine. The higher the comprehensive score, the better the service quality of the face recognition engine. The scoring module outputs the calculated comprehensive scores of each face recognition engine to the optimal scheduling subsystem; After receiving the comprehensive scores of each face recognition engine from the comprehensive evaluation subsystem, the scheduling task generation module of the optimal scheduling subsystem will calculate the scheduling weight of each face recognition engine according to the comprehensive score of the face recognition engine and the call success rate of the business subsystem, and then pre-generate the scheduling task according to the scheduling weight and put it into the work bucket; A scheduling task contains multiple face recognition engines, and a business request selects a scheduling task to process; The business subsystem completes the call to the face recognition engine by initiating a request to the preferential scheduling subsystem; the scheduling execution module of the preferential scheduling subsystem requests a scheduling task from the scheduling task selection module; the scheduling task selection module selects a suitable scheduling task from the scheduling work bucket and returns it to the scheduling execution module; the scheduling execution module executes the various face recognition engines of the scheduling task in parallel, and integrates the results returned by each face recognition engine to finally generate a comprehensive face recognition result, and then returns this result to the business subsystem.

2. The system according to claim 1, characterized in that The scheduling weight calculation process is: Set the comprehensive score of face recognition engine i to S i , the success rate T of business subsystem y calling face recognition engine i yi , the business impact coefficient is U, then the scheduling weight of face recognition engine i is W i =S i ×(1-U)+T yi ×U×100.

3. The system according to claim 1, characterized in that The scheduling task generation rules are: Assume that the face recognition engine subsystem integrates n face recognition engines, and selects m face recognition engines for a scheduling task, 1<m<n, n, m are both positive integers, and the scheduling weight of the i-th face recognition engine is W i , in the initial state of a cycle, the cycle weight R of the i-th face recognition engine i =mW i , the whole cycle will be ∑R i The face recognition engine selection can generate ∑W i The probability of selecting the face recognition engine for each scheduling task When selecting the execution of a scheduling task, first select the R i Calculate the selection space K = [1, ∑R i ], set the random value function V(P i )=Random(K,1), random(K,1) means to randomly get an integer between 1 and K. If 1≤V(P i )≤R1, then face recognition engine 1 is selected. Then face recognition engine y is selected. When face recognition engine y is selected, the selection of the next face recognition engine of the scheduling task will exclude face recognition engine y, that is, the selection space K = [1, ∑R i -R y ], and R y =R y -1, at this time, the random selection of the second face recognition engine is executed again, and the m face recognition engines of this scheduling task are selected one by one and put into the work bucket, and then the generation of the next scheduling task is started, until the cycle weights of all face recognitions are returned to 0, and the execution of this cycle is completed.

4. The system according to claim 1, characterized in that The scoring module is also used to output an evaluation result report of the face recognition engine by horizontally comparing the scores of various indicators of different face recognition engines, and to feed back the report to the maintainer of the face recognition engine. The maintainer will perform targeted optimization and upgrade the face recognition engine based on the feedback results.

5. A method for evaluating and optimizing multiple algorithms of a face recognition engine using the system of any one of claims 1 to 4, characterized in that: The specific process is:

1. Horizontal comparison and evaluation of face recognition engines: The face recognition engine in the face recognition engine subsystem will generate running status data in real time during the face recognition process. The data collection module of the comprehensive evaluation subsystem collects the running status data of all face recognition engines and transmits it to the data analysis module of the comprehensive evaluation subsystem. The data analysis module performs classification and classification processing on the operating status data of all face recognition engines, generates statistical data of multiple dimensional scoring index items based on the classification and classification processing results, and transmits them to the scoring module of the comprehensive evaluation subsystem; The scoring module calculates the scores of each scoring index item of the face recognition engine according to the preset scoring rules of each scoring index item, and then calculates the comprehensive score of the face recognition engine in combination with the hierarchical weights of the evaluation index items, and then outputs the comprehensive score of the face recognition engine to the optimal scheduling subsystem; 2. Face recognition engine optimal scheduling: The scheduling task generation module of the optimal scheduling subsystem calculates the scheduling weight of the face recognition engine according to the comprehensive score of the face recognition engine, and then pre-generates the scheduling task according to the scheduling weight and puts it into the work bucket. One scheduling task contains multiple face recognition engines, and one business request selects one scheduling task to process; When the business subsystem needs face recognition service, it initiates a request to the preferential scheduling subsystem to complete the call to the face recognition engine; the scheduling execution module of the preferential scheduling subsystem requests a scheduling task from the scheduling task selection module; the scheduling task selection module selects a suitable scheduling task from the scheduling work bucket and returns it to the scheduling execution module; the scheduling execution module executes each face recognition engine of the scheduling task in parallel, and integrates the results returned by each face recognition engine, and finally generates a comprehensive face recognition result, which is then returned to the business subsystem.

Images