A distributed face recognition method and system and a face recognition terminal

Abstract

The application discloses a kind of distributed face recognition method and system and face recognition terminal, specifically related to face recognition technical field, based on face recognition failure rate and identification failure close occurrence ratio carry out weighted summation calculation identification failure bad evaluation value, for quantization monitoring agent department to the performance and stability of face detection function;Through identification failure bad evaluation value, feature extraction comprehensive value and face feature extraction difference evaluation value calculation agent performance stability comprehensive evaluation coefficient, to comprehensively quantified evaluation agent department performance;And when generating agent department operation bad operation signal, can automatically suspend the operation of agent department and notify relevant maintenance personnel;Through the ratio of the occupation time and time length of the generating agent department operation acceptable signal in the comprehensive performance monitoring interval, even without generating clear operation bad signal case identification agent department performance poor trend, help to find potential problems in advance.

Description

Technical Field

[0001] The present invention relates to the field of face recognition technology, and more particularly to a distributed face recognition method and system and a face recognition terminal. Background Art

[0002] Traditional centralized face recognition methods have shortcomings in terms of time efficiency and scalability, and can no longer meet the needs of large-scale real-time face recognition. There is currently a distributed face recognition method: the model consists of multiple agents and a server. The agents can simultaneously perform face detection, face tracking, and feature extraction on people in multiple videos, while the server performs face recognition operations on pedestrians in the video.

[0003] In reality, there are facial recognition devices that are suitable for entry and exit management of enterprises, schools, and various government departments. In practice, multiple facial recognition devices will be set up in the facial recognition access channel according to the actual situation of enterprises, schools, and various government departments; in the facial recognition access channel, usually one facial recognition access channel corresponds to one facial recognition device, and one facial recognition device corresponds to one agent department, and the agents are connected to each other by wire or wirelessly, and the facial recognition devices can switch with each other between the agent departments; under normal circumstances, the facial recognition task of the facial recognition access channel usually uses the agent department corresponding to the facial recognition device itself to recognize and process the face, and then executes the facial recognition operation (such as calling up the identity information corresponding to the face and executing the access command) through the server.

[0004] The agent unit is crucial to the completion of the face recognition processing task, but the existing monitoring of the operating status of the agent unit corresponding to the face recognition device during the face recognition process is inadequate, resulting in the agent unit corresponding to the face recognition device switching to other agent units only when poor performance or failure occurs. The agent unit cannot be warned when poor performance or failure occurs, which will cause face recognition to become unstable, resulting in reduced recognition accuracy, prolonged response time, and even failure to work normally; and it is impossible to formulate a more reasonable strategy based on the real-time operating performance of the agent units corresponding to multiple face recognition devices.

[0005] In order to solve the above problems, a technical solution is now provided. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a distributed face recognition method and system and a face recognition terminal to solve the problems raised in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A distributed face recognition method comprises the following steps:

[0009] Step S1: Collecting proxy detection error information, including recognition error badness evaluation values, calculating face recognition error rate and recognition error close occurrence ratio, and weighted summing the face recognition error rate and recognition error close occurrence ratio to calculate the recognition error badness evaluation value;

[0010] By comparing the recognition error bad evaluation value with the recognition error judgment threshold, an agent part available bad signal or an agent part available normal signal is generated;

[0011] Step S2: collecting feature extraction performance stability information, the feature extraction performance stability information including the feature extraction comprehensive value and the face feature extraction difference evaluation value;

[0012] Step S3: When a normal signal is generated for the agent unit, the recognition error poor evaluation value, the feature extraction comprehensive value, and the facial feature extraction difference evaluation value are normalized to calculate the agent performance stability comprehensive evaluation coefficient; the agent performance stability comprehensive evaluation coefficient is compared with the first agent performance judgment threshold and the second agent performance judgment threshold to generate an agent unit excellent operation signal, an agent unit acceptable operation signal, or an agent unit poor operation signal;

[0013] Step S4: When the agent part poor operation signal is not generated, the comprehensive poor ratio is calculated, and the agent comprehensive performance poor signal or the agent comprehensive performance normal signal is generated by comparing the comprehensive poor ratio with the comprehensive poor judgment threshold.

[0014] In a preferred embodiment, in step S1, the specific logic for obtaining the bad evaluation value of the identification error is:

[0015] Obtaining a recognition error monitoring interval; obtaining the number of all targets recognized as faces by the agent within the recognition error monitoring interval, and obtaining the number of non-face targets mistakenly recognized as faces by the agent within the recognition error monitoring interval;

[0016] Calculate the ratio of the number of non-face targets erroneously identified as faces by the agent within the recognition error monitoring interval to the number of all targets identified as faces by the agent within the recognition error monitoring interval; mark the ratio of the number of non-face targets erroneously identified as faces by the agent within the recognition error monitoring interval to the number of all targets identified as faces by the agent within the recognition error monitoring interval as the face recognition error rate;

[0017] Obtaining time points at which the agent unit mistakenly identifies a non-face target as a face within a recognition error monitoring interval, marking the time points at which the agent unit mistakenly identifies a non-face target as a face within the recognition error monitoring interval as mistaken recognition points, and obtaining the time interval between every two adjacent mistaken recognition points within the recognition error monitoring interval;

[0018] The number of times the time interval between adjacent erroneous recognition points in the recognition error monitoring interval is less than the error interval threshold is counted; the recognition error close occurrence ratio is calculated, which is the ratio of the number of times the time interval between adjacent erroneous recognition points in the recognition error monitoring interval is less than the error interval threshold to the time length corresponding to the recognition error monitoring interval.

[0019] In a preferred embodiment, the face recognition error rate and the recognition error close occurrence ratio are subjected to a unit-removing process, and the face recognition error rate and the recognition error close occurrence ratio after the unit-removing process are weighted and summed to calculate the recognition error bad evaluation value, which is expressed as follows: sszp=μ1*rssv+μ2*ssjb, wherein sszp, rssv, and ssjb are the recognition error bad evaluation value, the face recognition error rate, and the recognition error close occurrence ratio, respectively, and μ1 and μ2 are the weights of the face recognition error rate and the recognition error close occurrence ratio, respectively, and μ1 and μ2 are both greater than 0 and μ1 and μ2 are fixed values;

[0020] When the recognition error bad evaluation value is greater than the recognition error judgment threshold, the agent part is generated. When the recognition error bad evaluation value is less than or equal to the recognition error judgment threshold, the agent part is generated.

[0021] In a preferred embodiment, in step S2, the specific logic for obtaining the feature extraction comprehensive value is:

[0022] Obtain a feature extraction monitoring set, where the feature extraction monitoring set includes n feature extraction monitoring points;

[0023] Obtain the facial feature extraction time corresponding to each feature extraction monitoring point in the feature extraction monitoring set;

[0024] Calculate the feature extraction comprehensive value, which is the ratio of the sum of the facial feature extraction times corresponding to all feature extraction monitoring points in the feature extraction monitoring set to the number of feature extraction monitoring points.

[0025] In a preferred embodiment, the specific logic for obtaining the facial feature extraction difference evaluation value is: performing discrete analysis on the facial feature extraction time corresponding to the feature extraction monitoring points in the feature extraction monitoring set, and calculating the facial feature extraction difference evaluation value, which is expressed as follows: Wherein, n, q are the number of feature extraction monitoring points in the feature extraction monitoring set and the number of feature extraction monitoring points in the feature extraction monitoring set respectively, q = 1, 2, 3, 4, …, n, n, q are positive integers greater than or equal to 1; rtcp, rtts q , ttzz are the face feature extraction difference evaluation value, the face feature extraction time corresponding to the qth feature extraction monitoring point in the feature extraction monitoring set, and the feature extraction comprehensive value respectively.

[0026] In a preferred embodiment, in step S3, the recognition failure bad evaluation value, the feature extraction comprehensive value, and the face feature extraction difference evaluation value are normalized, and the agent performance stability comprehensive evaluation coefficient is calculated through the normalized recognition failure bad evaluation value, the feature extraction comprehensive value, and the face feature extraction difference evaluation value;

[0027] The agent performance stability comprehensive evaluation coefficient is compared with the agent performance judgment first threshold and the agent performance judgment second threshold:

[0028] When the agent performance stability comprehensive evaluation coefficient is less than the agent performance judgment first threshold, an agent department running excellent signal is generated; when the agent performance stability comprehensive evaluation coefficient is greater than or equal to the agent performance judgment first threshold and less than or equal to the agent performance judgment second threshold, an agent department running acceptable signal is generated; when the agent performance stability comprehensive evaluation coefficient is greater than the agent performance judgment second threshold, an agent department running bad signal is generated;

[0029] Wherein, the agent performance judgment first threshold is less than the agent performance judgment second threshold.

[0030] In a preferred embodiment, in step S4, the comprehensive performance monitoring interval is obtained; when the agent department running bad signal is not generated in the comprehensive performance monitoring interval, the occupancy time of generating the agent department running acceptable signal in the comprehensive performance monitoring interval is obtained, and the ratio of the occupancy time of generating the agent department running acceptable signal in the comprehensive performance monitoring interval to the time length corresponding to the comprehensive performance monitoring interval is marked as a comprehensive bad ratio;

[0031] The comprehensive bad ratio is compared with a comprehensive bad judgment threshold:

[0032] When the comprehensive bad ratio is greater than the comprehensive bad judgment threshold, an agent comprehensive performance poor signal is generated;

[0033] When the comprehensive bad ratio is less than or equal to the comprehensive bad judgment threshold, an agent comprehensive performance normal signal is generated, and no measures need to be taken.

[0034] In a preferred embodiment, a distributed face recognition system includes a data processing module and an information acquisition module in communication with the data processing module, a detection error judgment module, an agent operation evaluation module, and a comprehensive status judgment module;

[0035] The information collection module collects agent detection error information, sends the agent detection error information to the data processing module, and calculates the recognition error bad evaluation value;

[0036] The detection error judgment module generates an agent part available bad signal or an agent part available normal signal by comparing the recognition error bad evaluation value with the recognition error judgment threshold;

[0037] The information collection module collects the feature extraction performance stability information, sends the feature extraction performance stability information to the data processing module, and calculates the feature extraction comprehensive value and the face feature extraction difference evaluation value;

[0038] When the detection error judgment module generates a normal signal for the agent part, the data processing module normalizes the recognition error poor evaluation value, the feature extraction comprehensive value, and the face feature extraction difference evaluation value to calculate the agent performance stability comprehensive evaluation coefficient;

[0039] The agent operation evaluation module compares the agent performance stability comprehensive evaluation coefficient with the agent performance judgment first threshold and the agent performance judgment second threshold to generate an agent department operation excellent signal, an agent department operation acceptable signal, or an agent department operation poor signal;

[0040] When the agent operation evaluation module does not generate a poor operation signal for the agent part, the comprehensive bad ratio is calculated by the data processing module, and the comprehensive status judgment module generates a poor agent comprehensive performance signal or a normal agent comprehensive performance signal by comparing the comprehensive bad ratio and the comprehensive bad judgment threshold.

[0041] In a preferred embodiment, the face recognition terminal consists of multiple agent units and a server. The agent units can simultaneously perform face detection, face tracking, and feature extraction on people in multiple videos, and the server performs face recognition operations on pedestrians in the video.

[0042] The technical effects and advantages of the distributed face recognition method and system and face recognition terminal of the present invention are as follows:

[0043] 1. The recognition error worst evaluation value is calculated based on the weighted sum of the face recognition error rate and the recognition error occurrence ratio. It is used to quantitatively monitor the performance and stability of the face detection function of the agent department, detect problems in advance and take measures to improve the reliability of the system, thereby reducing the risk of agent department failure.

[0044] 2. The performance stability comprehensive evaluation coefficient of the agent is calculated by comprehensively analyzing the identification failure bad evaluation value, the feature extraction comprehensive value and the face feature extraction difference evaluation value, so as to comprehensively quantitatively evaluate the performance of the agent department. When the agent department running bad running signal is generated, the running of the agent department can be automatically suspended, and the relevant maintenance personnel is notified, which helps to maximize the availability and performance of the system.

[0045] 3. The ratio (comprehensive bad ratio) of the occupancy time and the time length of the generated agent department running acceptable signal in the comprehensive performance monitoring interval is monitored, the trend of poor performance of the agent department is identified even if no explicit running bad signal is generated, potential problems are found in advance, the reliability and stability of the performance of the agent department are improved, and potential failures and performance fluctuations are reduced. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A schematic diagram of a distributed face recognition method of the present application is shown.

[0047] Figure 2 A structural schematic diagram of a distributed face recognition system of the present application is shown. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0049] Embodiment 1

[0050] Figure 1 A distributed face recognition method of the present application is given, which includes the following steps:

[0051] Step S1: Collecting agent detection failure information, the agent detection failure information includes an identification failure bad evaluation value, calculating a face recognition failure rate and an identification failure close occurrence ratio, weighting and summing the face recognition failure rate and the identification failure close occurrence ratio to calculate the identification failure bad evaluation value.

[0052] By comparing the identification failure bad evaluation value with the identification failure judgment threshold, an agent department available bad signal or an agent department available normal signal is generated.

[0053] Step S2: Collecting feature extraction performance stability information, the feature extraction performance stability information includes a feature extraction comprehensive value and a face feature extraction difference evaluation value.

[0054] Step S3: When a normal signal is generated for the agent part, the poor evaluation value of the recognition error, the comprehensive value of the feature extraction, and the difference evaluation value of the facial feature extraction are normalized to calculate the comprehensive evaluation coefficient of the agent performance stability; the comprehensive evaluation coefficient of the agent performance stability is compared with the first threshold value and the second threshold value of the agent performance judgment to generate an excellent agent part operation signal, an acceptable agent part operation signal, or a poor agent part operation signal.

[0055] Step S4: When the agent part poor operation signal is not generated, the comprehensive poor ratio is calculated, and the agent comprehensive performance poor signal or the agent comprehensive performance normal signal is generated by comparing the comprehensive poor ratio with the comprehensive poor judgment threshold.

[0056] In step S1, the specific steps of the agent unit in the face recognition process include face detection and collecting agent detection error information. The agent detection error information reflects the situation where the agent unit mistakenly identifies non-face targets as faces during face detection; for example, the appearance features of some non-face targets may be similar to faces, such as objects with face images printed on them, cartoon characters, etc., which may be mistakenly identified as faces. When the performance of the agent unit is poor, misrecognition may occur in complex scenes, and some non-face areas may be mistakenly marked as faces. This may be caused by factors such as lighting conditions, occlusion, and angle.

[0057] If the agency mistakenly identifies non-face objects as faces during face detection, if it is serious, it will have an adverse impact on the agency and the face recognition equipment:

[0058] This leads to a decrease in the performance of the proxy unit. In face detection, incorrectly identifying non-face targets as faces will process a large amount of unnecessary data, including incorrectly identified non-face targets; this will waste the computing resources of the proxy unit and reduce the efficiency of the proxy unit.

[0059] If the agent frequently misidentifies non-face objects as faces, this may lead to a decrease in the reliability of the face recognition device because the false recognition may lead to misidentification or rejection of legitimate faces.

[0060] Frequent errors may increase the time it takes for the agency to process tasks. The agency not only needs to process correct face recognition tasks, but also needs to process additional erroneous recognition tasks, causing congestion when a large number of people pass through the face recognition passage.

[0061] The agent detection error information includes the recognition error bad evaluation value. The specific acquisition logic of the recognition error bad evaluation value is:

[0062] An identification failure monitoring interval is obtained, the time length corresponding to the identification failure monitoring interval is set by a person skilled in the art according to actual monitoring requirements of the agent unit for identification of a face, the time length corresponding to the identification failure monitoring interval is a fixed value, but the range of the identification failure monitoring interval changes with real-time time, that is, one critical point of the identification failure monitoring interval is always real-time time; for example, the real-time time is 13:34, and the time length of the identification failure monitoring interval is 5 min, so the identification failure monitoring interval monitors the time interval between 13:29 and 13:34.

[0063] The number of all targets identified as a face by the agent unit in the identification failure monitoring interval is obtained, and the number of non-face targets incorrectly identified as a face by the agent unit in the identification failure monitoring interval is obtained.

[0064] The ratio of the number of non-face targets incorrectly identified as a face by the agent unit in the identification failure monitoring interval to the number of all targets identified as a face by the agent unit in the identification failure monitoring interval is calculated. The ratio of the number of non-face targets incorrectly identified as a face by the agent unit in the identification failure monitoring interval to the number of all targets identified as a face by the agent unit in the identification failure monitoring interval is marked as a face identification failure rate.

[0065] The number of all targets identified as a face by the agent unit is obtained: the data collected by the agent unit is traversed to check the identification result field of each identification event. For each identification event, it is checked whether the identification result is a face. If the identification result is a face, this event is counted as a target identified as a face by the agent unit, and the number of all targets identified as a face by the agent unit in the identification failure monitoring interval is counted.

[0066] The number of non-face targets incorrectly identified as a face by the agent unit is obtained: the data collected by the agent unit is traversed to check the identification result field of each identification event.

[0067] For each identification event, it is checked whether the identification result is a face. If the target is not actually a face, if this occurs, this event is counted as a case where the agent unit incorrectly identifies a non-face target as a face, and the number of non-face targets incorrectly identified as a face by the agent unit in the identification failure monitoring interval is counted.

[0068] The judgment logic for checking whether the identification result is a face but the target is not actually a face is as follows:

[0069] Based on subsequent face tracking and feature extraction, if they fail, it means that the target is not actually a face, which is not described here again.

[0070] Obtain the time points corresponding to the human faces when the agent unit mistakenly identifies the non-face targets within the recognition error monitoring interval, mark the time points corresponding to the human faces when the agent unit mistakenly identifies the non-face targets within the recognition error monitoring interval as error recognition points, and obtain the time interval between every two adjacent error recognition points within the recognition error monitoring interval.

[0071] The smaller the time interval between two adjacent erroneous recognition points within the recognition error monitoring interval, the higher the frequency at which the agent mistakenly recognizes non-face targets as faces.

[0072] If there are m erroneous recognition points within the recognition error monitoring interval, then when m is greater than 0, the number of time intervals between adjacent erroneous recognition points within the recognition error monitoring interval is m-1, and when m is equal to 0, the number of time intervals between adjacent erroneous recognition points within the recognition error monitoring interval is 0, and m is a natural number.

[0073] Set an error interval threshold. When the time interval between adjacent erroneous recognition points in the recognition error monitoring interval is less than the error interval threshold, it means that the time interval between each two adjacent erroneous recognition points is smaller. Count the number of times the time interval between adjacent erroneous recognition points in the recognition error monitoring interval is less than the error interval threshold.

[0074] The more times the time interval between adjacent erroneous recognition points in the recognition error monitoring interval is less than the error interval threshold, the more cases in which the time interval between each two adjacent erroneous recognition points in the recognition error monitoring interval is smaller, and the greater the adverse impact on the normal face recognition of the agent unit, which indicates that the probability of the agent failure in the recognition error monitoring interval is greater.

[0075] Calculate the recognition error close occurrence ratio, which is the ratio of the number of times the time interval between adjacent erroneous recognition points within the recognition error monitoring interval is less than the error interval threshold to the time length corresponding to the recognition error monitoring interval; the larger the recognition error close occurrence ratio, the more frequent face misrecognition occurs in the agent unit within the recognition error monitoring interval, indicating that the performance and stability of the agent unit have been adversely affected.

[0076] The error interval threshold is set by professionals in this field according to actual requirements for the interval of recognition errors of the agent, for example, the error interval threshold is set to 3s.

[0077] The face recognition error rate and the recognition error occurrence ratio are de-unitized, and the face recognition error rate and recognition error occurrence ratio after de-unitization are weighted and summed to calculate the recognition error bad evaluation value, which is expressed as: sszp = μ1*rssv+μ2*ssjb, wherein sszp, rssv, and ssjb are the recognition error bad evaluation value, the face recognition error rate, and the recognition error occurrence ratio, respectively, and μ1 and μ2 are the weights of the face recognition error rate and the recognition error occurrence ratio, respectively. μ1 and μ2 are both greater than 0 and μ1 and μ2 are fixed values. The size of μ1 and μ2 is set by professional and technical personnel in this field according to the actual size of the face recognition error rate and the recognition error occurrence ratio and other actual conditions, and will not be repeated here.

[0078] The greater the recognition error evaluation value, the worse the performance and stability of the agent unit for face recognition within the recognition error monitoring interval, which means that the worse the face misrecognition situation of the agent unit within the recognition error monitoring interval, the greater the probability of the agent failure.

[0079] Set a recognition error judgment threshold, and evaluate the performance and stability of the agent's face detection by comparing the recognition error evaluation value with the recognition error judgment threshold, so as to discover the availability of the agent's face detection function in advance.

[0080] When the recognition error bad evaluation value is greater than the recognition error judgment threshold, the agent part is generated. When the recognition error bad evaluation value is less than or equal to the recognition error judgment threshold, the agent part is generated.

[0081] When a poor proxy unit availability signal is generated, it indicates that the proxy unit's performance may have problems or be unstable in face detection. The proxy unit frequently makes errors or performs poorly when recognizing faces, which may have a significant adverse impact on subsequent face tracking and feature extraction. Based on the generated poor proxy unit availability signal, measures are taken to improve the proxy unit's performance to ensure normal face detection function. Specific measures include but are not limited to continuously improving and optimizing the face detection algorithm according to specific application scenarios and data and adjusting the algorithm's parameters.

[0082] When a normal signal is generated for the agent unit, it indicates that the performance of the agent unit is normal or stable in terms of face detection, and the performance of the agent unit is within the allowable operating range of the agent unit, but it does not mean that there is no possibility of problems.

[0083] The recognition error worst evaluation value is calculated based on the weighted sum of the face recognition error rate and the recognition error occurrence ratio. It is used to quantitatively monitor the performance and stability of the face detection function of the agent department, detect problems in advance and take measures to improve the reliability of the system, thereby reducing the risk of agent department failure.

[0084] In step S2, the agent's facial recognition process also includes feature extraction, a key step in facial recognition. By extracting key features from a facial image, such as facial contours, eyes, nose, and mouth, a unique facial feature vector can be created for identifying and verifying an individual. If the performance of feature extraction is adversely affected, the efficiency of subsequent server operations will be adversely affected.

[0085] The feature extraction performance stability information is collected, and the feature extraction performance stability information reflects the efficiency of facial feature extraction during the feature extraction process.

[0086] Feature extraction performance stability information includes the feature extraction comprehensive value and the face feature extraction difference evaluation value. The specific logic for obtaining the feature extraction comprehensive value is as follows:

[0087] A feature extraction monitoring set is obtained, which includes n feature extraction monitoring points. The number of feature extraction monitoring points in the feature extraction monitoring set is set by professional and technical personnel in this field based on the actual monitoring requirements of the feature extraction performance of the agent.

[0088] The number of feature extraction monitoring points in the feature extraction monitoring set is fixed, and the feature extraction monitoring set always collects the n feature extraction monitoring points closest to the real time.

[0089] The feature extraction monitoring set always collects the n feature extraction monitoring points closest to the real time in order to analyze the efficiency of feature extraction in the recent period of time and to better evaluate the performance of feature extraction.

[0090] Each feature extraction monitoring point corresponds to a facial feature extraction time, and the facial feature extraction time corresponding to each feature extraction monitoring point in the feature extraction monitoring set is obtained.

[0091] The facial feature extraction time refers to the time it takes for the agent to extract the features of a single facial data. The specific method for obtaining the time it takes to extract the features of a single facial data is as follows:

[0092] Before feature extraction begins, a timestamp is inserted to record the start time. The facial feature extraction process, which involves extracting features from a single face, typically involves calculating a facial feature vector or descriptor. After feature extraction completes, a timestamp is inserted to record the end time. The start and end timestamps are used to calculate the time taken for feature extraction. The difference is calculated by subtracting the start timestamp from the end timestamp, representing the time taken for feature extraction of a single face.

[0093] The feature extraction comprehensive value is the ratio of the sum of the face feature extraction time corresponding to all the feature extraction monitoring points in the feature extraction monitoring set to the number of the feature extraction monitoring points.

[0094] The greater the feature extraction comprehensive value, the more time-consuming the face feature extraction in the feature extraction monitoring set, which means that the feature extraction operation is relatively slow and the performance of the agent part is poor.

[0095] The specific acquisition logic of the face feature extraction difference evaluation value is: the face feature extraction time corresponding to the feature extraction monitoring points in the feature extraction monitoring set is discretely analyzed, and the face feature extraction difference evaluation value is calculated, and its expression is: Wherein, n and q are the number of feature extraction monitoring points in the feature extraction monitoring set and the number of feature extraction monitoring points in the feature extraction monitoring set, q = 1, 2, 3, 4,..., n, n and q are positive integers greater than or equal to 1; rtcp, rtts q , ttzz are the face feature extraction difference evaluation value, the face feature extraction time corresponding to the qth feature extraction monitoring point in the feature extraction monitoring set, and the feature extraction comprehensive value.

[0096] The greater the face feature extraction difference evaluation value, the more dispersed the face feature extraction time in the feature extraction monitoring set; the greater the face feature extraction difference evaluation value, the greater the difference and fluctuation between the face feature extraction time of different feature extraction monitoring points, which reflects the instability of the feature extraction performance of the agent part.

[0097] It is worth noting that the feature extraction comprehensive value is the average value of the face feature extraction time corresponding to the feature extraction monitoring points in the feature extraction monitoring set.

[0098] In step S3, the agent detection failure information and the feature extraction performance stability information are comprehensively analyzed to determine the performance of the single agent part in operation, so as to early warn the state of the agent part.

[0099] The recognition failure bad evaluation value under the condition of generating agent part available bad signal is excluded, because measures have been taken under the condition of generating agent part available bad signal, so the recognition failure bad evaluation value under the condition of generating agent part available normal signal is analyzed.

[0100] When a normal signal is generated for the agent unit, the recognition error poor evaluation value, the feature extraction comprehensive value, and the facial feature extraction difference evaluation value are normalized, and the agent performance stability comprehensive evaluation coefficient is calculated based on the normalized recognition error poor evaluation value, the feature extraction comprehensive value, and the facial feature extraction difference evaluation value. For example, the present invention can use the following formula to calculate the agent performance stability comprehensive evaluation coefficient, which is expressed as follows: Among them, dxwp is the comprehensive evaluation coefficient of agent performance stability, α1, α2, and α3 are the preset proportional coefficients of the poor evaluation value of recognition error, the comprehensive value of feature extraction, and the difference evaluation value of face feature extraction, respectively, and α1, α2, and α3 are all greater than 0.

[0101] The greater the comprehensive evaluation coefficient of agent performance stability, the worse the operating performance of the agent unit, which may easily lead to the inability to perform face recognition in a timely manner, affecting the efficiency of face recognition, and thus causing blockage in the face recognition channel.

[0102] A first agent performance judgment threshold and a second agent performance judgment threshold are set, and the first agent performance judgment threshold is smaller than the second agent performance judgment threshold.

[0103] Compare the agent performance stability comprehensive evaluation coefficient with the first and second thresholds for agent performance judgment:

[0104] When the comprehensive evaluation coefficient of agent performance stability is less than the first threshold of agent performance judgment, an excellent agent operation signal is generated; when the comprehensive evaluation coefficient of agent performance stability is greater than or equal to the first threshold of agent performance judgment, and the comprehensive evaluation coefficient of agent performance stability is less than or equal to the second threshold of agent performance judgment, an acceptable agent operation signal is generated; when the comprehensive evaluation coefficient of agent performance stability is greater than the second threshold of agent performance judgment, a poor agent operation signal is generated.

[0105] When an excellent signal of the agent unit operation is generated, the agent unit is operating normally and no action is required.

[0106] When an acceptable signal for the operation of the proxy unit is generated, it indicates that the proxy unit may have a small amount of poor performance, but it does not affect the current performance of the proxy unit.

[0107] When a signal indicating poor performance is generated for an agent, indicating poor performance, the agent is suspended and, based on the generated signal, a professional technician is assigned to inspect and repair the agent. The agent's unfinished tasks (feature extraction) are then transferred to another agent that is generating a signal indicating excellent performance. (If no agent is generating an excellent performance signal, the agent is transferred to another agent generating an acceptable performance signal. If no agent is generating an excellent performance signal, the agent is not assigned to another agent.)

[0108] Among them, the first threshold value for judging agent performance and the second threshold value for judging agent performance are set by professional and technical personnel in this field based on the size of the comprehensive evaluation coefficient of agent performance stability and other actual conditions such as the actual performance requirements of the agent department, and will not be repeated here.

[0109] Different types of signals are generated based on the range of the agent performance stability comprehensive evaluation coefficient, enabling appropriate action to be taken. The agent performance stability comprehensive evaluation coefficient is calculated by comprehensively analyzing the poor recognition error evaluation value, the feature extraction comprehensive value, and the facial feature extraction difference evaluation value, providing a comprehensive and quantitative assessment of agent performance. Furthermore, when a poor agent performance signal is generated, the agent operation can be automatically suspended and maintenance personnel notified. This helps maximize system availability and performance through task switching and agent maintenance.

[0110] In step S4, when the agent does not generate a poor agent operation signal, although the agent operates within an acceptable range of normal operation, if more acceptable agent operation signals are generated within a period of time, the adverse impact on the probability of failure of the agent is also relatively large.

[0111] Therefore, the comprehensive performance monitoring interval is obtained. The time length corresponding to the comprehensive performance monitoring interval is set by professional and technical personnel in this field based on the actual monitoring needs of the comprehensive requirements for the performance of the agent department within a period of time. The time length corresponding to the comprehensive performance monitoring interval is a fixed value, but the range of the comprehensive performance monitoring interval changes with the real-time time, that is, a critical point of the comprehensive performance monitoring interval is always the real-time time; for example, the real-time time is 13:34, and the time length of the comprehensive performance monitoring interval is 25 minutes, then the comprehensive performance monitoring interval monitors the time interval between 13:09 and 13:34.

[0112] When no agent unit operation bad operation signal is generated within the comprehensive performance monitoring interval, the occupied time for generating the agent unit operation acceptable signal within the comprehensive performance monitoring interval is obtained, and the ratio of the occupied time for generating the agent unit operation acceptable signal within the comprehensive performance monitoring interval to the time length corresponding to the comprehensive performance monitoring interval is marked as the comprehensive bad ratio.

[0113] Set a comprehensive bad judgment threshold. The comprehensive bad judgment threshold is set based on the size of the comprehensive bad ratio and other actual conditions such as the performance requirements of the agency within a period of time.

[0114] Compare the comprehensive bad ratio and the comprehensive bad judgment threshold:

[0115] When the overall poor performance ratio exceeds the overall poor performance threshold, a poor overall agent performance signal is generated. This indicates that the proportion of acceptable performance signals generated by the agent within the comprehensive performance monitoring interval is too high, significantly impacting the agent's current and future performance. The agent's operation is suspended. Based on the generated poor overall agent performance signal, professional technicians are assigned to inspect the agent and repair it if any problems are found. The agent's unfinished tasks (feature extraction) are then transferred to another agent that is generating excellent performance signals. (If no agent is generating excellent performance signals, this operation is not performed.)

[0116] When the comprehensive bad ratio is less than or equal to the comprehensive bad judgment threshold, a normal agent comprehensive performance signal is generated and no action is required.

[0117] By monitoring the ratio of the time it takes to generate an acceptable signal for the agent's operation to the length of time (comprehensive bad ratio) within the comprehensive performance monitoring interval, we can identify trends in poor agent performance even when no clear bad operation signal is generated. This helps to discover potential problems in advance, reduce the risk of future failures of the agent, improve the reliability and stability of the agent's performance, and reduce potential failures and performance fluctuations.

[0118] Example 2

[0119] The difference between Example 2 of the present invention and Example 1 is that this example introduces a distributed face recognition system.

[0120] Figure 2 A structural schematic diagram of a distributed face recognition system of the present invention is given. A distributed face recognition system includes a data processing module and an information acquisition module connected to the data processing module, a detection error judgment module, an agent operation evaluation module and a comprehensive status judgment module.

[0121] The information collection module collects agent detection error information, sends the agent detection error information to the data processing module, and calculates the recognition error bad evaluation value.

[0122] The detection error judgment module generates an agent part usable bad signal or an agent part usable normal signal by comparing the recognition error bad evaluation value with the recognition error judgment threshold.

[0123] The information collection module collects the feature extraction performance stability information, sends the feature extraction performance stability information to the data processing module, and calculates the feature extraction comprehensive value and the face feature extraction difference evaluation value.

[0124] When the detection error judgment module generates a normal signal for the agent part, the data processing module normalizes the recognition error poor evaluation value, the feature extraction comprehensive value and the face feature extraction difference evaluation value to calculate the agent performance stability comprehensive evaluation coefficient.

[0125] The agent operation evaluation module compares the agent performance stability comprehensive evaluation coefficient with the agent performance judgment first threshold and the agent performance judgment second threshold to generate an excellent agent operation signal, an acceptable agent operation signal or a poor agent operation signal.

[0126] When the agent operation evaluation module does not generate a poor operation signal for the agent part, the comprehensive bad ratio is calculated by the data processing module, and the comprehensive status judgment module generates a poor agent comprehensive performance signal or a normal agent comprehensive performance signal by comparing the comprehensive bad ratio and the comprehensive bad judgment threshold.

[0127] Example 3

[0128] A face recognition terminal is used to implement a distributed face recognition method and a distributed face recognition system. That is, the face recognition terminal is a face recognition device. The face recognition terminal consists of multiple agent parts and a server. The agent parts can simultaneously perform face detection, face tracking and feature extraction on people in multiple videos, and the server performs face recognition operations on pedestrians in the video.

[0129] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0130] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center through a wired (for example, infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0131] Those of ordinary skill in the art can realize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those of ordinary skill in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0132] Those of ordinary skill in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, device, and module can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0133] In several embodiments provided in the present application, it should be understood that the disclosed system, device, and method can be implemented in other ways. For example, the above-described device embodiments are only schematic, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, for example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed ones can be indirect coupling or communication connection through some interfaces, devices, or modules, which can be electrical, mechanical, or other forms.

[0134] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0135] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0136] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0137] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0138] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A distributed face recognition method, characterized in that: The method monitors the running status of the proxy part in the distributed face recognition process to ensure the stability of face detection, feature extraction and subsequent server face recognition operations, including the following steps: Step S1: Collecting proxy detection error information, including recognition error badness evaluation values, calculating face recognition error rate and recognition error close occurrence ratio, and weighted summing the face recognition error rate and recognition error close occurrence ratio to calculate the recognition error badness evaluation value; By comparing the recognition error bad evaluation value with the recognition error judgment threshold, an agent part available bad signal or an agent part available normal signal is generated; Step S2: collecting feature extraction performance stability information, the feature extraction performance stability information including the feature extraction comprehensive value and the face feature extraction difference evaluation value; Step S3: When a normal signal is generated for the agent unit, the recognition error poor evaluation value, the feature extraction comprehensive value, and the facial feature extraction difference evaluation value are normalized to calculate the agent performance stability comprehensive evaluation coefficient; the agent performance stability comprehensive evaluation coefficient is compared with the first agent performance judgment threshold and the second agent performance judgment threshold to generate an agent unit excellent operation signal, an agent unit acceptable operation signal, or an agent unit poor operation signal; Step S4: When the agent unit is not generating a poor operation signal, the comprehensive bad ratio is calculated, and by comparing the comprehensive bad ratio with the comprehensive bad judgment threshold, a poor agent comprehensive performance signal or a normal agent comprehensive performance signal is generated. The normal comprehensive performance signal is used to predict the reliability of the subsequent face recognition process supported by the agent unit.

2. A distributed face recognition method according to claim 1, characterized in that: In step S1, the specific logic for obtaining the poor evaluation value of the recognition error is: Obtaining a recognition error monitoring interval; obtaining the number of all targets recognized as faces by the agent within the recognition error monitoring interval, and obtaining the number of non-face targets mistakenly recognized as faces by the agent within the recognition error monitoring interval; Calculate the ratio of the number of non-face targets erroneously identified as faces by the agent within the recognition error monitoring interval to the number of all targets identified as faces by the agent within the recognition error monitoring interval; mark the ratio of the number of non-face targets erroneously identified as faces by the agent within the recognition error monitoring interval to the number of all targets identified as faces by the agent within the recognition error monitoring interval as the face recognition error rate; Obtaining time points corresponding to when the agent unit mistakenly identifies a non-face target as a face within the recognition error monitoring interval, marking the time points corresponding to when the agent unit mistakenly identifies a non-face target as a face within the recognition error monitoring interval as mistaken recognition points, and obtaining the time interval between every two adjacent mistaken recognition points within the recognition error monitoring interval; Count the number of times that the time interval between adjacent erroneous recognition points in the recognition error monitoring interval is less than the error interval threshold; The recognition error close occurrence ratio is calculated, and the recognition error close occurrence ratio is the ratio of the number of times the time interval between adjacent erroneous recognition points in the recognition error monitoring interval is less than the error interval threshold to the time length corresponding to the recognition error monitoring interval.

3. A distributed face recognition method according to claim 2, characterized in that: The face recognition error rate and the recognition error occurrence ratio are subjected to unit-removal processing, and the weighted sum of the face recognition error rate and the recognition error occurrence ratio after unit-removal processing is performed to calculate the recognition error evaluation value, which is expressed as follows: ,in, They are respectively the recognition error bad evaluation value, face recognition error rate and recognition error close occurrence ratio, are the weights of face recognition error rate and recognition error occurrence ratio, respectively. are greater than 0 and is a fixed value; When the recognition error bad evaluation value is greater than the recognition error judgment threshold, the agent part is generated. When the recognition error bad evaluation value is less than or equal to the recognition error judgment threshold, the agent part is generated.

4. The distributed face recognition method according to claim 1, wherein: In step S2, the specific logic for obtaining the feature extraction comprehensive value is: Obtain a feature extraction monitoring set, where the feature extraction monitoring set includes n feature extraction monitoring points; Obtain the facial feature extraction time corresponding to each feature extraction monitoring point in the feature extraction monitoring set; Calculate the feature extraction comprehensive value, which is the ratio of the sum of the facial feature extraction times corresponding to all feature extraction monitoring points in the feature extraction monitoring set to the number of feature extraction monitoring points.

5. A distributed face recognition method according to claim 4, characterized in that: The specific logic for obtaining the face feature extraction difference evaluation value is: discrete analysis is performed on the face feature extraction time corresponding to the feature extraction monitoring points in the feature extraction monitoring set, and the face feature extraction difference evaluation value is calculated. The expression is: ,in, are the number of feature extraction monitoring points in the feature extraction monitoring set and the serial number of the feature extraction monitoring points in the feature extraction monitoring set, , are all positive integers greater than or equal to 1; They are respectively the face feature extraction difference evaluation value, the feature extraction monitoring set The facial feature extraction time and feature extraction comprehensive value corresponding to each feature extraction monitoring point.

6. The distributed face recognition method according to claim 1, characterized in that: In step S3, the recognition error bad evaluation value, the feature extraction comprehensive value, and the face feature extraction difference evaluation value are normalized, and the agent performance stability comprehensive evaluation coefficient is calculated based on the normalized recognition error bad evaluation value, the feature extraction comprehensive value, and the face feature extraction difference evaluation value; Compare the agent performance stability comprehensive evaluation coefficient with the first and second thresholds for agent performance judgment: When the comprehensive evaluation coefficient of agent performance stability is less than the first threshold for agent performance judgment, an excellent agent operation signal is generated; When the comprehensive evaluation coefficient of agent performance stability is greater than or equal to the first threshold value of agent performance judgment, and the comprehensive evaluation coefficient of agent performance stability is less than or equal to the second threshold value of agent performance judgment, an acceptable signal for the agent unit operation is generated; when the comprehensive evaluation coefficient of agent performance stability is greater than the second threshold value of agent performance judgment, a poor operation signal for the agent unit operation is generated; The first threshold for determining the performance of the agent is smaller than the second threshold for determining the performance of the agent.

7. A distributed face recognition method according to claim 6, characterized in that: In step S4, a comprehensive performance monitoring interval is obtained; if no agent unit operation poor operation signal is generated within the comprehensive performance monitoring interval, the occupied time for generating the agent unit operation acceptable signal within the comprehensive performance monitoring interval is obtained, and the ratio of the occupied time for generating the agent unit operation acceptable signal within the comprehensive performance monitoring interval to the time length corresponding to the comprehensive performance monitoring interval is marked as the comprehensive poor ratio; Compare the comprehensive bad ratio and the comprehensive bad judgment threshold: When the comprehensive bad ratio is greater than the comprehensive bad judgment threshold, a poor agent comprehensive performance signal is generated; When the comprehensive bad ratio is less than or equal to the comprehensive bad judgment threshold, a normal agent comprehensive performance signal is generated and no action is required.

8. A distributed face recognition system, configured to implement the distributed face recognition method according to any one of claims 1 to 7, wherein the method ensures face recognition stability by monitoring the operating status of the agent unit, and wherein: It includes a data processing module and an information collection module in communication with the data processing module, and is used to collect information related to face detection and feature extraction of the agent unit; The detection error judgment module is used to assess the risk of face detection misidentification; the agent operation evaluation module is used to evaluate the stability of feature extraction; and the comprehensive state judgment module is used to predict the reliability of the agent's support for face recognition; The information collection module collects agent detection error information, sends the agent detection error information to the data processing module, and calculates the recognition error bad evaluation value; The detection error judgment module generates an agent part available bad signal or an agent part available normal signal by comparing the recognition error bad evaluation value with the recognition error judgment threshold; The information collection module collects the feature extraction performance stability information, sends the feature extraction performance stability information to the data processing module, and calculates the feature extraction comprehensive value and the face feature extraction difference evaluation value; When the detection error judgment module generates a normal signal for the agent part, the data processing module normalizes the recognition error poor evaluation value, the feature extraction comprehensive value, and the face feature extraction difference evaluation value to calculate the agent performance stability comprehensive evaluation coefficient; The agent operation evaluation module compares the agent performance stability comprehensive evaluation coefficient with the agent performance judgment first threshold and the agent performance judgment second threshold to generate an agent department operation excellent signal, an agent department operation acceptable signal, or an agent department operation poor signal; When the agent operation evaluation module does not generate a poor operation signal for the agent part, the comprehensive bad ratio is calculated by the data processing module, and the comprehensive status judgment module generates a poor agent comprehensive performance signal or a normal agent comprehensive performance signal by comparing the comprehensive bad ratio and the comprehensive bad judgment threshold.

9. Face recognition terminal, characterized by: A distributed face recognition method for implementing any one of claims 1-7, wherein the method ensures the stability of face recognition by monitoring the operating status of the agent unit; the face recognition terminal is composed of multiple agent units and a server, and the agent unit can simultaneously perform face detection, face tracking and feature extraction on people in multiple videos. The operating status of the agent unit is monitored by the method, and the server performs face recognition operations on pedestrians in the video.

Images