Image recognition model optimization method based on artificial intelligence

Abstract

The present invention discloses an image recognition model optimization method based on artificial intelligence, and specifically relates to the field of image recognition technology. The whole automatic labeling process is divided into monitoring sub-intervals, labeling boundary error information, labeling error response information, and environmental interference information of the automatic labeling process in the monitoring sub-intervals are obtained, a labeling hidden danger identification model is constructed, a labeling hidden danger identification index is generated, the labeling abnormality degree existing in the automatic labeling process is evaluated, and the automatic labeling process is identified and classified, and the labeling hidden dangers are accurately identified. The identification of intelligent labeling hidden dangers effectively improves the accuracy of image labeling and reduces the occurrence of erroneous labeling. The labeled products in the monitoring sub-interval corresponding to the generated labeling hidden danger signal are transferred to manual labeling, and the labeling difference information after manual intervention is obtained. The preset labeling hidden danger identification index threshold is dynamically adjusted in combination with the difference time fluctuation information.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and more specifically, to an image recognition model optimization method based on artificial intelligence. Background Art

[0002] Image recognition technology is an important direction in the field of computer vision, which aims to automatically identify and classify objects, scenes or features in images through computer algorithms. With the rapid development of deep learning, image recognition technology based on artificial intelligence has made significant progress and is widely used in security monitoring, medical image analysis, autonomous driving, smart retail and other fields.

[0003] Before building an image recognition model, it is usually necessary to use image annotation technology to annotate the product images. The annotated product images can provide the label information required for training for the image recognition model, so that the image recognition model can learn specific features and patterns. The image annotation dataset can also be used to evaluate the performance of the model. By comparing the difference between the model prediction results and the true labels, the image recognition model can be optimized. Image annotation technology mainly includes manual annotation and automatic annotation. Manual annotation is done manually, with high accuracy but low efficiency. Automatic annotation relies on deep learning algorithms, and directly annotates new images through training models. Although it is efficient, its accuracy may be affected by the quality of training data. In addition, there is still a lack of intelligent identification of annotation risks in the annotation process in the field of automatic annotation. Especially in smart retail, since product packaging may have different degrees of deformation, if there are annotation risks in the automatic annotation process, the deformation of the product packaging may not be effectively identified, and the deformed product may be incorrectly labeled as other categories. If the process with annotation risks is not intervened in time, the accuracy of automatic annotation may continue to decline. With the accumulation of incorrect annotations, the overall quality of training data declines, resulting in distortion of features and patterns learned by the image recognition model, thereby affecting the accuracy of the image recognition model. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides an image recognition model optimization method based on artificial intelligence to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The image recognition model optimization method based on artificial intelligence includes the following steps:

[0007] Step S1, dividing the entire automatic labeling process into monitoring sub-intervals, and obtaining labeling boundary error information, labeling error response information, and environmental interference information of the automatic labeling process in the monitoring sub-intervals;

[0008] Step S2, constructing a labeling hidden danger identification model according to the labeling boundary error information, the labeling error response information, and the environmental interference information, generating a labeling hidden danger identification index, and evaluating the degree of labeling anomalies in the automatic labeling process;

[0009] Step S3, comparing the marked hidden danger identification index with a preset marked hidden danger identification index threshold, and classifying the automatic marking process;

[0010] Step S4, determining whether to perform manual intervention according to the identification and classification results, obtaining the labeling difference information after the manual intervention, and dynamically adjusting the preset labeling hidden danger identification index threshold in combination with the difference time fluctuation information.

[0011] In a preferred embodiment, the marking boundary error information includes a marking boundary error coefficient, the marking error response information includes a marking error response coefficient, and the environmental interference information includes an anti-interference pressure coefficient.

[0012] In a preferred embodiment, the logic for obtaining the annotation boundary error coefficient is as follows:

[0013] Obtain the annotation boundaries of similar products after automatic annotation, place the obtained annotation boundaries of similar products in the same coordinate system, and calculate the annotation boundary coordinate difference BJ between similar products according to the Euclidean distance calculation method. The expression is as follows Where (x1 i ,y1 i ) and (x2 i , y2 i ) represent the coordinates of two similar products at the same point i;

[0014] Compare the annotation boundary coordinate difference with the preset annotation boundary difference threshold, count the total number of times N when the annotation boundary coordinate difference is greater than the annotation boundary difference threshold, and calculate the annotation boundary error coefficient bzb, which is expressed as follows: In the formula, BJ n It represents the labeled boundary coordinate difference when the labeled boundary coordinate difference is greater than the labeled boundary difference threshold for the nth time. BJZ represents the preset labeled boundary difference threshold. n={1,2,...,N}, where N is a positive integer.

[0015] In a preferred embodiment, the acquisition logic of the label error response coefficient is as follows:

[0016] When a labeling error is detected, the detection time T is recorded start and the time to complete the re-labeling T end ;

[0017] Calculate the response time T response =T end -T start, calculate the average response time Indicates the number of times the mth labeling error is detected, m = {1, 2, ..., M}, M represents the total number of times the labeling error is detected;

[0018] Compare the response time with the expected response time. When the response time is less than the expected response time, mark this re-labeling as successful re-labeling, and count the number of successful re-labelings ES;

[0019] Calculate the success rate success, the expression is as follows Where M represents the total number of times labeling errors are detected;

[0020] Calculate the label error response coefficient bzc, the expression is as follows Where a 1 、a 2 Represents the preset proportionality coefficients of the success rate and the inverse of the average response time respectively.

[0021] In a preferred embodiment, the logic for obtaining the anti-interference pressure coefficient is as follows:

[0022] An automatic annotation experiment is conducted in an interference-free environment to measure the accuracy ZQ of image annotation within the time period TS. The expression is as follows Where Z1 represents the number of correct annotations within the time period TS, and Z2 represents the total number of annotations within the time period TS;

[0023] Use the light intensity sensor to collect the actual light intensity during the automatic labeling process in real time and calculate the anti-interference pressure coefficient Where TD represents the time period when the actual light intensity increases, and ZQW represents the accuracy rate during the time period when the actual light intensity increases.

[0024] In a preferred embodiment, a labeling hidden danger identification model is constructed according to the labeling boundary error coefficient, the labeling error response coefficient, and the anti-interference pressure coefficient to generate a labeling hidden danger identification index yhz, according to the following formula yhz=e (α*bzb-β*bzc-γ*kgr) , where α, β, and γ represent the preset proportional coefficients of the annotation boundary error coefficient, the annotation error response coefficient, and the anti-interference pressure coefficient, respectively, and α, β, and γ are all greater than 0.

[0025] In a preferred embodiment, in step S3, the annotated hidden danger identification index is compared with a preset annotated hidden danger identification index threshold, and the automatic annotation process is classified as follows:

[0026] If the marked hidden danger identification index is greater than the marked hidden danger identification index threshold, a marked hidden danger signal is generated;

[0027] If the labeled hidden danger identification index is less than or equal to the labeled hidden danger identification index threshold, a labeled stable signal is generated.

[0028] In a preferred embodiment, the annotated products in the monitoring sub-interval corresponding to the generated annotated hidden danger signal are transferred to manual labeling to obtain the annotated difference information after manual intervention, wherein the annotated difference information includes the overlap coefficient of the annotated area;

[0029] The logic for obtaining the overlap coefficient of the annotation area is as follows:

[0030] For the same annotated product, obtain the annotated area MJ1 under automatic annotating and the annotated area MJ2 under manual annotating, and calculate the area overlap MJC, which is expressed as follows: MJC = |MJ1-MJ2|;

[0031] Set an initial maximum area overlap, traverse all the annotated products in the monitoring sub-interval, calculate the area overlap of each annotated product in turn for each annotated product in the monitoring sub-interval, and compare it with the maximum area overlap of the current record. If the area overlap of the current annotated product is greater than the maximum area overlap of the current record, update the maximum area overlap, that is, use the area overlap of the current annotated product as the maximum area overlap. At the same time, obtain the annotation completion time of the current annotated product in the automatic annotation process, add it to the annotation product difference time set, and mark the annotation product difference time set as TC = {tc j}, tc j Indicates the labeling completion time of the corresponding labeling product in the automatic labeling process when the maximum value of the j-th region coincidence degree is updated, j = {1, 2, ..., J}, J is a positive integer;

[0032] Calculate the overlap coefficient qcd of the marked area. The expression is as follows Where gxc represents the update times of the maximum value of regional overlap, and zcs represents the number of all annotated products in the monitoring sub-interval;

[0033] The difference time fluctuation information includes the difference time fluctuation coefficient;

[0034] The logic for obtaining the difference time fluctuation coefficient is as follows:

[0035] Calculate the average of the labeling completion time in the labeling product difference time set The expression is as follows

[0036] Calculate the standard deviation sd of the labeling completion time in the labeling product difference time set. The expression is as follows

[0037]

[0038] Calculate the difference time fluctuation coefficient cys, the expression is as follows

[0039] In a preferred embodiment, the overlap coefficient of the marked area and the difference time fluctuation coefficient are normalized, and the normalized overlap coefficient of the marked area and the difference time fluctuation coefficient are used as adjustment factors to dynamically adjust the preset threshold of the marked hidden danger identification index. The specific adjustment function is as follows: x =zby*e -(qcd+cys) , where zby x It represents the adjusted threshold of the hidden danger identification index, and zby represents the threshold of the hidden danger identification index.

[0040] Technical effects and advantages of the present invention:

[0041] 1. The present invention divides the entire automatic labeling process into monitoring sub-intervals, obtains the labeling boundary error information, labeling error response information, and environmental interference information of the automatic labeling process in the monitoring sub-intervals, constructs a labeling hidden danger identification model, generates a labeling hidden danger identification index, evaluates the degree of labeling abnormality in the automatic labeling process, compares the labeling hidden danger identification index with a preset labeling hidden danger identification index threshold, identifies and classifies the automatic labeling process, accurately identifies labeling hidden dangers, and intelligently identifies labeling hidden dangers, which effectively improves the accuracy of image labeling and reduces the occurrence of erroneous labeling. The labeled products in the monitoring sub-interval corresponding to the generated labeling hidden danger signal are transferred to manual labeling, and the labeling difference information after manual intervention is obtained. The preset labeling hidden danger identification index threshold is dynamically adjusted in combination with the difference time fluctuation information to adapt to real-time scene changes, effectively solve the labeling hidden danger problem in the automatic labeling process, and provide a strong guarantee for the performance optimization of the image recognition model. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to facilitate understanding by those skilled in the art, the present invention is further described below in conjunction with the accompanying drawings;

[0043] Figure 1 Schematic diagram of the structure of the method of the embodiment of the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] Example: Figure 1The present invention provides an image recognition model optimization method based on artificial intelligence, which includes the following steps:

[0046] Step S1, dividing the entire automatic labeling process into monitoring sub-intervals, and obtaining labeling boundary error information, labeling error response information, and environmental interference information of the automatic labeling process in the monitoring sub-intervals;

[0047] Step S2, constructing a labeling hidden danger identification model according to the labeling boundary error information, the labeling error response information, and the environmental interference information, generating a labeling hidden danger identification index, and evaluating the degree of labeling anomalies in the automatic labeling process;

[0048] Step S3, comparing the marked hidden danger identification index with a preset marked hidden danger identification index threshold, and classifying the automatic marking process;

[0049] Step S4, determining whether to perform manual intervention according to the identification and classification results, obtaining the labeling difference information after the manual intervention, and dynamically adjusting the preset labeling hidden danger identification index threshold in combination with the difference time fluctuation information;

[0050] Step S1, dividing the entire automatic labeling process into monitoring sub-intervals, and obtaining labeling boundary error information, labeling error response information, and environmental interference information of the automatic labeling process in the monitoring sub-intervals;

[0051] The marking boundary error information includes a marking boundary error coefficient, the marking error response information includes a marking error response coefficient, and the environmental interference information includes an anti-interference pressure coefficient;

[0052] The labeling boundary error coefficient is used to measure the error degree between the labeling boundaries when the automatic labeling technology labels the same type of products. When labeling the same product, the labeling results are very different, which means that there may be a large labeling risk in the automatic labeling process, and the operating performance of the automatic labeling strategy is gradually declining. By calculating the labeling boundary error coefficient, it can help identify the labeling risks that may exist in the automatic labeling process, and provide an important basis for the optimization and improvement of automatic labeling technology. By continuously monitoring and analyzing the labeling boundary error coefficient, the accuracy and consistency of automatic labeling can be effectively improved, ensuring the efficiency and reliability of the image recognition model in practical applications.

[0053] The logic for obtaining the annotation boundary error coefficient is as follows:

[0054] Obtain the annotation boundaries of similar products after automatic annotation, place the obtained annotation boundaries of similar products in the same coordinate system, and calculate the annotation boundary coordinate difference BJ between similar products according to the Euclidean distance calculation method. The expression is as follows Where (x1 i ,y1i ) and (x2 i , y2 i ) represent the coordinates of two similar products at the same point i;

[0055] It should be noted that for size standardization and coordinate transformation of different annotated images, commonly used methods include the nearest neighbor interpolation algorithm, Lanczos interpolation algorithm for scaling the annotated image, Min-Max normalization, Z-score normalization for normalizing the annotated image, and affine transformation algorithm for coordinate transformation of the annotated image to ensure that all annotation boundaries are compared in the same reference frame;

[0056] Compare the annotation boundary coordinate difference with the preset annotation boundary difference threshold, count the total number of times N when the annotation boundary coordinate difference is greater than the annotation boundary difference threshold, and calculate the annotation boundary error coefficient bzb, which is expressed as follows: In the formula, BJ n represents the labeled boundary coordinate difference when the labeled boundary coordinate difference is greater than the labeled boundary difference threshold for the nth time, BJZ represents the preset labeled boundary difference threshold, n={1,2,...,N}, N is a positive integer;

[0057] The labeling error response coefficient is used to measure the degree of response of whether the product can be re-labeled within the expected time period when a product labeling error occurs during the automatic labeling process. The labeling error response coefficient can be used to help evaluate the efficiency and accuracy of the automatic labeling technology, ensure that labeling errors are corrected in a timely manner, thereby improving the overall labeling quality and reducing the impact of labeling errors on the subsequent image recognition model training and application. The larger the labeling error response coefficient, the more efficient and timely the system can be when labeling errors occur, thereby successfully correcting the errors within the expected time. This means that the automatic labeling technology has a strong response capability, can quickly identify and correct labeling problems, and effectively improve the accuracy of labeling and the overall labeling quality. Conversely, the smaller the labeling error response coefficient, the weaker the system's response capability, the labeling errors cannot be corrected within a reasonable time, and the greater the probability that there may be labeling risks in the automatic labeling process, which may have a negative impact on the subsequent training of the image recognition model. This may lead to the accumulation of erroneous labeling, which in turn affects the image recognition model's learning of specific features and reduces recognition accuracy;

[0058] The logic for obtaining the label error response coefficient is as follows:

[0059] When a labeling error is detected, the detection time T is recorded start and the time to complete the re-labeling T end ;

[0060] Calculate the response time T response =Tend -T start , calculate the average response time Indicates the number of times the mth labeling error is detected, m = {1, 2, ..., M}, M represents the total number of times the labeling error is detected;

[0061] Compare the response time with the expected response time. When the response time is less than the expected response time, mark this re-labeling as successful re-labeling, and count the number of successful re-labelings ES;

[0062] Calculate the success rate success, the expression is as follows Where M represents the total number of times labeling errors are detected;

[0063] Calculate the label error response coefficient bzc, the expression is as follows Where a 1 、a 2 The preset proportionality coefficients representing the success rate and the inverse of the average response time respectively;

[0064] It should be noted that before calculating the labeling error response coefficient based on the success rate and the average response time, the success rate and the average response time need to be normalized. Common normalization methods include Min-Max normalization and Z-Score normalization. 1 、a 2 It is set according to the actual situation. For example, the expert empowerment method is adopted, that is, experts in related fields are invited to determine the preset proportion coefficients of various indicators through professional opinion surveys and comprehensive evaluations. I will not elaborate on this here;

[0065] The anti-interference pressure coefficient is used to measure the adaptability and anti-interference ability caused by environmental changes in the automatic labeling process. Specifically, the anti-interference pressure coefficient reflects the degree to which the automatic labeling technology can maintain labeling accuracy and stability when facing external interference factors. When the anti-interference pressure coefficient is large, it means that the automatic labeling technology has a strong adaptability to environmental changes. Even under unfavorable external conditions, it can still complete the labeling task more accurately, and the labeling quality and performance are less affected. At this time, the automatic labeling process can cope with the interference of light, background or other factors well to ensure the robustness of the labeling. When the anti-interference pressure coefficient is small, it means that it is more sensitive to external interference. Environmental changes may significantly affect the labeling effect, resulting in an increased incidence of labeling errors. At this time, the reliability and consistency of automatic labeling in actual applications are poor, and there may be a certain degree of labeling risks;

[0066] The logic for obtaining the anti-interference pressure coefficient is as follows:

[0067] An automatic annotation experiment is conducted in an interference-free environment to measure the accuracy ZQ of image annotation within the time period TS. The expression is as follows Where Z1 represents the number of correct annotations within the time period TS, and Z2 represents the total number of annotations within the time period TS;

[0068] It should be noted that the environmental interference in the present invention includes but is not limited to light intensity;

[0069] Use the light intensity sensor to collect the actual light intensity during the automatic labeling process in real time and calculate the anti-interference pressure coefficient Where TD represents the time period when the actual light intensity increases, and ZQW represents the accuracy rate during the time period when the actual light intensity increases;

[0070] Step S2, constructing a labeling hidden danger identification model according to the labeling boundary error information, the labeling error response information, and the environmental interference information, generating a labeling hidden danger identification index, and evaluating the degree of labeling anomalies in the automatic labeling process;

[0071] According to the annotation boundary error coefficient, annotation error response coefficient, and anti-interference pressure coefficient, a annotation hidden danger identification model is constructed to generate the annotation hidden danger identification index yhz, based on the following formula yhz = e (α*bzb-β*bzc-γ*kgr) , where α, β, and γ represent the preset proportional coefficients of the annotation boundary error coefficient, the annotation error response coefficient, and the anti-interference pressure coefficient, respectively, and α, β, and γ are all greater than 0;

[0072] It should be noted that before constructing the annotation hidden danger identification model, the annotation boundary error coefficient, the annotation error response coefficient, and the anti-interference pressure coefficient need to be normalized. Common normalization methods include Min-Max normalization and Z-Score standardization. α, β, and γ are set according to actual conditions. For example, the expert empowerment method is adopted, that is, experts in related fields are invited to determine the preset proportional coefficients of various indicators through professional opinion surveys and comprehensive evaluations.

[0073] From the above calculation expression, it can be seen that the larger the labeling boundary error coefficient, the smaller the labeling error response coefficient, and the smaller the anti-interference pressure coefficient, the larger the labeling hidden danger identification index is, indicating that the probability of labeling hidden dangers in the automatic labeling process is greater. Conversely, the smaller the labeling boundary error coefficient, the larger the labeling error response coefficient, and the larger the anti-interference pressure coefficient, the smaller the labeling hidden danger identification index is, indicating that the probability of labeling hidden dangers in the automatic labeling process is smaller.

[0074] Step S3, comparing the marked hidden danger identification index with the preset marked hidden danger identification index threshold, and classifying the automatic marking process, as follows:

[0075] If the labeling hidden danger identification index is greater than the labeling hidden danger identification index threshold, it indicates that there is a labeling hidden danger in the automatic labeling process within the monitoring sub-interval, and a labeling hidden danger signal is generated;

[0076] If the labeling hidden danger identification index is less than or equal to the labeling hidden danger identification index threshold, it indicates that there is no obvious labeling hidden danger in the automatic labeling process within the monitoring sub-interval, and a labeling stability signal is generated;

[0077] Step S4, determining whether to perform manual intervention according to the identification and classification results, obtaining the labeling difference information after the manual intervention, and dynamically adjusting the preset labeling hidden danger identification index threshold in combination with the difference time fluctuation information;

[0078] Transferring the annotated products in the monitoring sub-interval corresponding to the generated annotated hidden danger signal to manual labeling, and obtaining the annotated difference information after manual intervention, wherein the annotated difference information includes the overlap coefficient of the annotated area;

[0079] The overlap coefficient of the annotation area is used to measure the difference between the automatic annotation and manual annotation of the same annotation product, and specifically indicates how much overlap there is between the two annotation areas. The calculation of the overlap coefficient of the annotation area can effectively evaluate the accuracy of automatic annotation, and help determine whether there are signs of continued expansion of the annotation risks in the automatic annotation process, which helps to more accurately adjust the threshold of the annotation risk identification index;

[0080] The logic for obtaining the overlap coefficient of the annotation area is as follows:

[0081] For the same annotated product, obtain the annotated area MJ1 under automatic annotating and the annotated area MJ2 under manual annotating, and calculate the area overlap MJC, which is expressed as follows: MJC = |MJ1-MJ2|;

[0082] Set an initial maximum area overlap, traverse all the annotated products in the monitoring sub-interval, calculate the area overlap of each annotated product in turn for each annotated product in the monitoring sub-interval, and compare it with the maximum area overlap of the current record. If the area overlap of the current annotated product is greater than the maximum area overlap of the current record, update the maximum area overlap, that is, use the area overlap of the current annotated product as the maximum area overlap. At the same time, obtain the annotation completion time of the current annotated product in the automatic annotation process, add it to the annotation product difference time set, and mark the annotation product difference time set as TC = {tc j}, tc j Indicates the labeling completion time of the corresponding labeling product in the automatic labeling process when the maximum value of the j-th region coincidence degree is updated, j = {1, 2, ..., J}, J is a positive integer;

[0083] It should be noted that the initial maximum value of the region overlap is a default minimum value, such as 0;

[0084] Calculate the overlap coefficient qcd of the marked area. The expression is as follows Where gxc represents the update times of the maximum value of regional overlap, and zcs represents the number of all annotated products in the monitoring sub-interval;

[0085] The difference time fluctuation information includes the difference time fluctuation coefficient;

[0086] The difference time fluctuation coefficient is used to measure the degree of fluctuation of the labeling completion time of different labeling products in the labeling product difference time set during the automatic labeling process, reflecting the fluctuation of the labeling products with labeling risks in the time dimension. By calculating the difference time fluctuation coefficient, it is helpful to adjust the threshold of the labeling risk identification index more accurately.

[0087] The logic for obtaining the difference time fluctuation coefficient is as follows:

[0088] Calculate the average of the labeling completion time in the labeling product difference time set The expression is as follows

[0089] Calculate the standard deviation sd of the labeling completion time in the labeling product difference time set. The expression is as follows

[0090]

[0091] Calculate the difference time fluctuation coefficient cys, the expression is as follows

[0092] The overlap coefficient of the marked area and the difference time fluctuation coefficient are normalized, and the normalized overlap coefficient of the marked area and the difference time fluctuation coefficient are used as adjustment factors to dynamically adjust the preset threshold of the marked hidden danger identification index. The specific adjustment function is as follows: x =zby*e -(qcd+cys) , where zby x represents the adjusted threshold value of the hidden danger identification index, and zby represents the threshold value of the hidden danger identification index;

[0093] The present invention divides the whole process of automatic labeling into monitoring sub-intervals, obtains labeling boundary error information, labeling error response information, and environmental interference information of the automatic labeling process in the monitoring sub-intervals, constructs a labeling hidden danger identification model, generates a labeling hidden danger identification index, evaluates the degree of labeling anomalies in the automatic labeling process, compares the labeling hidden danger identification index with a preset labeling hidden danger identification index threshold, identifies and classifies the automatic labeling process, accurately identifies labeling hidden dangers, and intelligently identifies labeling hidden dangers to effectively improve the accuracy of image labeling and reduce the occurrence of erroneous labeling. The labeled products in the monitoring sub-interval corresponding to the generated labeling hidden danger signal are transferred to manual labeling, and the labeling difference information after manual intervention is obtained. The preset labeling hidden danger identification index threshold is dynamically adjusted in combination with the difference time fluctuation information to adapt to real-time scene changes, effectively solve the labeling hidden danger problem in the automatic labeling process, and provide a strong guarantee for the performance optimization of the image recognition model;

[0094] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0095] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0096] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An image recognition model optimization method based on artificial intelligence, characterized in that: The steps include: Step S1, dividing the entire automatic labeling process into monitoring sub-intervals, and obtaining labeling boundary error information, labeling error response information, and environmental interference information of the automatic labeling process in the monitoring sub-intervals; Step S2, constructing a labeling hidden danger identification model according to the labeling boundary error information, the labeling error response information, and the environmental interference information, generating a labeling hidden danger identification index, and evaluating the degree of labeling anomalies in the automatic labeling process; Step S3, comparing the marked hidden danger identification index with a preset marked hidden danger identification index threshold, and classifying the automatic marking process; Step S4, determining whether to perform manual intervention according to the identification and classification results, obtaining the labeling difference information after the manual intervention, and dynamically adjusting the preset labeling hidden danger identification index threshold in combination with the difference time fluctuation information; The marking boundary error information includes a marking boundary error coefficient, the marking error response information includes a marking error response coefficient, and the environmental interference information includes an anti-interference pressure coefficient; The logic for obtaining the annotation boundary error coefficient is as follows: Obtain the annotation boundaries of similar products after automatic annotation, place the obtained annotation boundaries of similar products in the same coordinate system, and calculate the annotation boundary coordinate difference BJ between similar products according to the Euclidean distance calculation method. The expression is as follows Where (x1 i , y1 i ) and (x2 i , y2 i ) represent the coordinates of two similar products at the same point i; Compare the annotation boundary coordinate difference with the preset annotation boundary difference threshold, count the total number of times N when the annotation boundary coordinate difference is greater than the annotation boundary difference threshold, and calculate the annotation boundary error coefficient bzb, which is expressed as follows: In the formula, BJ n It represents the labeled boundary coordinate difference when the labeled boundary coordinate difference is greater than the labeled boundary difference threshold for the nth time. BJZ represents the preset labeled boundary difference threshold. n={1,2,...,N}, where N is a positive integer.

2. The method for optimizing an image recognition model based on artificial intelligence according to claim 1, characterized in that: The logic for obtaining the label error response coefficient is as follows: When a labeling error is detected, the detection time T is recorded start and the time to complete the re-labeling T end ; Calculate the response time T response =T end -T start , calculate the average response time M represents the total number of labeling errors detected; Compare the response time with the expected response time. When the response time is less than the expected response time, mark this re-labeling as successful re-labeling, and count the number of successful re-labelings ES; Calculate the success rate success, the expression is as follows Where M represents the total number of times labeling errors are detected; Calculate the label error response coefficient bzc, the expression is as follows Where a1 and a2 represent the preset proportional coefficients of the success rate and the inverse of the average response time, respectively.

3. The method for optimizing an image recognition model based on artificial intelligence according to claim 1, characterized in that: The logic for obtaining the anti-interference pressure coefficient is as follows: An automatic annotation experiment is conducted in an interference-free environment to measure the accuracy ZQ of image annotation within the time period TS. The expression is as follows Where Z1 represents the number of correct annotations within the time period TS, and Z2 represents the total number of annotations within the time period TS; Use the light intensity sensor to collect the actual light intensity during the automatic labeling process in real time and calculate the anti-interference pressure coefficient Where TD represents the time period when the actual light intensity increases, and ZQW represents the accuracy rate during the time period when the actual light intensity increases.

4. The method for optimizing an image recognition model based on artificial intelligence according to claim 1, characterized in that: According to the annotation boundary error coefficient, annotation error response coefficient, and anti-interference pressure coefficient, a annotation hidden danger identification model is constructed to generate the annotation hidden danger identification index yhz, based on the following formula yhz = e (α*bzb-β*bzc-γ*kgr) , where α, β, and γ represent the preset proportional coefficients of the annotation boundary error coefficient, the annotation error response coefficient, and the anti-interference pressure coefficient, respectively, and α, β, and γ are all greater than 0.

5. The method for optimizing an image recognition model based on artificial intelligence according to claim 4, characterized in that: In step S3, the annotation hidden danger identification index is compared with the preset annotation hidden danger identification index threshold, and the automatic annotation process is classified as follows: If the marked hidden danger identification index is greater than the marked hidden danger identification index threshold, a marked hidden danger signal is generated; If the labeled hidden danger identification index is less than or equal to the labeled hidden danger identification index threshold, a labeled stable signal is generated.

6. The method for optimizing an image recognition model based on artificial intelligence according to claim 5, characterized in that: Transferring the annotated products in the monitoring sub-interval corresponding to the generated annotated hidden danger signal to manual labeling, and obtaining the annotated difference information after manual intervention, wherein the annotated difference information includes the overlap coefficient of the annotated area; The logic for obtaining the overlap coefficient of the annotation area is as follows: For the same annotated product, obtain the annotated area MJ1 under automatic annotating and the annotated area MJ2 under manual annotating, and calculate the area overlap MJC, which is expressed as follows: MJC = |MJ1-MJ2|; Set an initial maximum area overlap, traverse all the annotated products in the monitoring sub-interval, calculate the area overlap of each annotated product in turn for each annotated product in the monitoring sub-interval, and compare it with the maximum area overlap of the current record. If the area overlap of the current annotated product is greater than the maximum area overlap of the current record, update the maximum area overlap, that is, use the area overlap of the current annotated product as the maximum area overlap. At the same time, obtain the annotation completion time of the current annotated product in the automatic annotation process, add it to the annotation product difference time set, and mark the annotation product difference time set as TC = {tc j }, tc j Indicates the labeling completion time of the corresponding labeling product in the automatic labeling process when the maximum value of the j-th region coincidence degree is updated, j = {1, 2, ..., J}, J is a positive integer; Calculate the overlap coefficient qcd of the marked area. The expression is as follows Where gxc represents the update times of the maximum value of regional overlap, and zcs represents the number of all annotated products in the monitoring sub-interval; The difference time fluctuation information includes the difference time fluctuation coefficient; The logic for obtaining the difference time fluctuation coefficient is as follows: Calculate the average of the labeling completion time in the labeling product difference time set The expression is as follows Calculate the standard deviation sd of the labeling completion time in the labeling product difference time set. The expression is as follows Calculate the difference time fluctuation coefficient cys, the expression is as follows 7. The method for optimizing an image recognition model based on artificial intelligence according to claim 6, characterized in that: The overlap coefficient of the marked area and the difference time fluctuation coefficient are normalized, and the normalized overlap coefficient of the marked area and the difference time fluctuation coefficient are used as adjustment factors to dynamically adjust the preset threshold of the marked hidden danger identification index. The specific adjustment function is as follows: x =zby*e -(qcd+cys) , where zby x It represents the adjusted threshold of the hidden danger identification index, and zby represents the threshold of the hidden danger identification index.

Images