An artificial intelligence visual inspection method based on eigenface method

Abstract

The present invention discloses an artificial intelligence visual detection method based on a feature face method, which relates to the field of artificial intelligence technology, including synchronously collecting image data of visible light images, infrared images and depth images of the target object's face, and performing data preprocessing to form a multimodal facial feature standard data set; dynamically adjusting the weights in the feature vectors of each modality according to the current time, geographic location information and historical data statistics to form a comprehensive feature vector; using the comprehensive feature vector, identifying the identity of the target object through a machine learning classifier; when the recognition result reaches a preset confidence threshold, outputting the recognition conclusion, otherwise it will return to the feature extraction step. The present invention realizes adaptive optimization of the recognition strategy by integrating the feature face method of visible light, infrared and depth images, combined with a dynamic weight adjustment strategy, and effectively solves the problems of accuracy and robustness of face recognition in complex environments.

Description

Technical Field

[0001] The invention relates to the field of artificial intelligence technology, in particular to an artificial intelligence visual detection method based on a feature face method. Background Art

[0002] In the rapid evolution of current artificial intelligence visual detection technology, the eigenface method, as a classic face recognition strategy, effectively reduces the complexity of high-dimensional face image data through principal component analysis and realizes effective feature extraction. However, this traditional method mainly focuses on visible light image processing. When faced with complex and changeable real-world environments, such as extreme lighting conditions, occlusion or camouflage, its recognition accuracy and robustness are severely challenged. In addition, the complementary advantages of infrared and depth information are ignored, limiting the application potential of face recognition technology in complex scenarios. The deficiencies of existing technologies in feature fusion and adaptive adjustment are particularly prominent. They lack the ability to dynamically optimize the recognition model according to environmental changes and time sensitivity, resulting in large fluctuations in recognition performance at different times and locations, affecting the stability and practicality of the system. Summary of the invention

[0003] In view of the above existing problems, the present invention is proposed.

[0004] Therefore, the present invention provides an artificial intelligence visual detection method based on the feature face method to solve the problems of accuracy and robustness of face recognition in complex environments.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] In the first aspect, an embodiment of the present invention provides an artificial intelligence visual detection method based on a feature face method, which includes synchronously collecting image data of visible light images, infrared images and depth images of the target object's face, and performing data preprocessing to form a multimodal facial feature standard data set; extracting each modal feature vector of visible light image features, infrared thermal map features and depth features in the multimodal facial feature standard data set through a feature face analysis algorithm, heat map region segmentation, and 3D geometric feature description; dynamically adjusting the weights in the feature vectors of each modality according to the current time, geographic location information and historical data statistics to form a comprehensive feature vector; using the comprehensive feature vector to identify the identity of the target object through a machine learning classifier; when the recognition result reaches a preset confidence threshold, outputting the recognition conclusion, otherwise it will return to the feature extraction step.

[0007] As a preferred solution of the artificial intelligence visual detection method based on the eigenface method described in the present invention, the synchronous acquisition process includes utilizing adaptive exposure control and dynamic thermal balance adjustment to optimize image quality and reduce interference from environmental factors.

[0008] As a preferred solution of the artificial intelligence visual inspection method based on the eigenface method described in the present invention, the data preprocessing includes image correction, alignment, denoising and normalization processing.

[0009] As a preferred solution of the artificial intelligence visual detection method based on the eigenface method of the present invention, wherein: the eigenface analysis algorithm, heat map region segmentation, and 3D geometric feature description are used to extract the visible light image feature vector in the multimodal facial feature standard data set, and the specific steps are:

[0010] The multimodal facial feature standard dataset P is expressed as:

[0011] P={P vis ,P ir ,P dep}

[0012] Among them, P vis is the visible light image, P ir is the infrared image, P dep is the depth image;

[0013] Based on the multimodal facial feature standard data set, the PCA method is used to extract the eigenface basis matrix V, which is expressed as:

[0014]

[0015] Among them, v i is the i-th eigenvector, λ i is the non-zero eigenvalue corresponding to the i-th eigenvector, and K is the variable parameter;

[0016] The visible light image P is transformed into vis Projected into the feature space, generating the visible light image feature vector F vis The expression is:

[0017] F vis =V T ·P vis

[0018] Among them, V T is the transpose of the eigenface basis matrix V.

[0019] As a preferred solution of the artificial intelligence visual detection method based on the eigenface method of the present invention, the infrared thermal image features and depth feature vectors in the multimodal facial feature standard data set are extracted, and the specific steps are as follows:

[0020] Binarization segmentation of infrared image P using Otsu method irGet the hot zone map and cold zone map, calculate the average temperature and temperature variance in the hot zone map, and generate the infrared heat map feature vector F tm The expression is:

[0021] F tm =[T avg ,σ 2 ]

[0022] Among them, T avg is the average temperature of all measurement points in the infrared thermal image, σ 2 is the temperature variance of infrared thermal image;

[0023] Extract depth map P dep The facial contour and key points in the image are used to calculate local and global geometric features and obtain the deep feature vector F. gm The expression is:

[0024]

[0025] Among them, VLAD is the local aggregation vector function, D loc is around the key point k i The local depth descriptor of , n is the total number of key points;

[0026] The visible light image feature vector F vis , infrared heat map feature vector F tm and the deep feature vector F gm Perform vector concatenation to form a multimodal feature vector set F mi The expression is: F mi =[F vis ,F tm ,F gm ].

[0027] As a preferred solution of the artificial intelligence visual detection method based on the eigenface method of the present invention, the weights in the feature vectors of each modality are dynamically adjusted according to the current time, geographical location information and historical data statistics to form a comprehensive feature vector. The specific steps are as follows:

[0028] Based on the collected current time t, geographic location information g and historical data statistics H v , dynamically adjust the weights of visible light, infrared and depth images respectively;

[0029] Dynamic weight adjustment expression w for visible light images v is:

[0030]

[0031] Among them, α is the constant coefficient, t is the time in hours, and β is the baseline weight;

[0032] Dynamic weight adjustment expression w for infrared images i r is:

[0033]

[0034] Among them, γ k is the basic weight of environment category k, f t (t) is a linear function that changes smoothly over time, 1 g＝k is the indicator function, e is the environmental factor code, k is the environmental category, and N is the total number of environmental categories;

[0035] Depth image dynamic weight adjustment expression w g m is:

[0036]

[0037] Among them, δ 0 ,δ 1 ,δ 2 To adjust the model parameters of the weights, is the average value of historical data statistics;

[0038] Based on the dynamic weights of the visible light image, infrared image and depth image, a comprehensive feature vector F is formed. ce The expression is: F ce =w v is·F vis +w i r.F tm +w g m·F gm .

[0039] As a preferred solution of the artificial intelligence visual detection method based on the eigenface method of the present invention, wherein: the identity of the target object is identified by using the comprehensive feature vector through a machine learning classifier, and the specific steps are:

[0040] The comprehensive feature vector F ce The comprehensive feature vector F is divided into 70% training set, 15% validation set and 15% test set. ce The training set is normalized by L2 norm and the expression is:

[0041]

[0042] Among them, F norm is the training set after L2 normalization;

[0043] Use cross-validation to select the best set of hyperparameters and train the gradient boosting tree model based on the L2-normalized training set;

[0044] The generalization ability of the gradient boosting tree classification model was evaluated using the accuracy, recall, F1 score, and area under the AUC-ROC curve in the test set;

[0045] Based on the performance of the test set and combined with business requirements, set an initial recognition confidence threshold;

[0046] Through the online learning mechanism, the performance of the gradient boosting tree classification model in practical applications is continuously monitored, user feedback and recognition results are collected, and the confidence threshold is dynamically adjusted to the optimal value according to the actual false recognition rate and missed recognition rate.

[0047] As a preferred solution of the artificial intelligence visual detection method based on the feature face method of the present invention, when the recognition result reaches a preset confidence threshold, the recognition conclusion is output, otherwise it returns to the feature extraction step, and the specific steps are:

[0048] When the probability score value output by the trained gradient boosting tree model is greater than or equal to the currently set confidence threshold, the target object is deemed to have been successfully identified and the corresponding identification conclusion is output; otherwise, the identification is considered insufficient and the feature re-extraction process is triggered, i.e., returning to the image preprocessing step for re-identification.

[0049] In a second aspect, an embodiment of the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the artificial intelligence visual detection method based on the feature face method as described in the first aspect of the present invention is implemented.

[0050] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the artificial intelligence visual detection method based on the feature face method as described in the first aspect of the present invention.

[0051] The beneficial effects of the present invention are as follows: the present invention not only broadens the dimension of feature extraction by synchronously collecting and deeply fusing visible light, infrared and depth image data, but also further enhances the system's ability to resist environmental interference, and adopts advanced feature face analysis, heat map region segmentation and 3D geometric feature description technology to achieve accurate extraction of multimodal features. Secondly, the weights of each modal feature are dynamically adjusted according to the current time, geographic location information and historical data analysis results, which improves the environmental adaptability and recognition accuracy of the recognition system, so that stable high-performance recognition effects can be maintained under various lighting conditions, geographical environments and individual differences. Therefore, the present invention not only solves the defects of the prior art in multimodal fusion and dynamic environmental adaptability, but also ensures the accuracy and robustness of face recognition through a highly intelligent feature weight adjustment mechanism, providing strong support for the advancement of artificial intelligence visual detection technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0053] Figure 1 This is a flow chart of the artificial intelligence visual detection method based on the eigenface method in Example 1.

[0054] Figure 2 This is a flow chart of the machine learning classifier in Example 1 identifying the identity of the target object. DETAILED DESCRIPTION

[0055] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0056] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0057] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0058] Example 1, reference Figure 1, which is the first embodiment of the present invention, and provides an artificial intelligence visual detection method based on the feature face method, comprising the following steps:

[0059] S1. Synchronously collect image data of visible light images, infrared images and depth images of the target object's face, and perform data preprocessing to form a multimodal facial feature standard data set.

[0060] Furthermore, the synchronous acquisition process includes the use of adaptive exposure control and dynamic thermal balance adjustment, which monitors the light intensity of the scene in real time through the built-in photosensors, and dynamically adjusts the exposure parameters according to the preset algorithm. For example, in low-light environments, the system may extend the shutter time or increase the aperture, or increase the ISO sensitivity to increase the amount of light entering; while in strong light, the opposite measures may be taken to prevent overexposure. For multi-camera systems, all devices need to be adjusted in coordination to ensure the consistency and comparability of all images, and to optimize image quality in a step-by-step manner;

[0061] Preferably, the infrared camera needs to be thermally calibrated before shooting. Dynamic thermal balance adjustment further monitors and compensates for the impact of ambient temperature and device temperature changes on the image in real time. It automatically adjusts the black body temperature reference point of the image by analyzing the background radiation temperature, so that the color mapping of the thermal image remains stable. Even if the ambient temperature changes, it can ensure the accuracy of the target temperature measurement and reduce the interference of environmental factors;

[0062] It should be noted that adaptive exposure control is a technology that automatically adjusts camera exposure parameters such as shutter speed, aperture size, ISO sensitivity, etc., in order to obtain images with the best exposure effect under different lighting conditions. This technology is particularly important in multi-camera synchronous acquisition systems because it can ensure that each frame of the image has appropriate brightness and contrast even in an environment where the light intensity changes rapidly, avoiding overexposure or underexposure;

[0063] It should also be noted that adaptive exposure control and dynamic thermal balance adjustment are indispensable technologies for synchronously acquiring multimodal image data. They work together to ensure that high-quality image data can be obtained under various complex environmental conditions, providing a solid foundation for subsequent image processing and feature extraction.

[0064] S2. Data preprocessing includes image correction, alignment, denoising and normalization.

[0065] Furthermore, the acquired image data is subjected to temporal layer correction. In fMRI functional magnetic resonance imaging, temporal layer correction usually involves correcting the possible differences between images acquired at different time points in the scanning sequence to ensure the consistency and continuity of the time series data. Although this step is mainly for fMRI data, similar corrections may also be applied to dynamic image sequences, such as video frames, to ensure the coherence between frames.

[0066] In the processing of the collected infrared images, radiation correction is required. Radiation correction is used to correct the image brightness inconsistency caused by factors such as uneven sensor response or atmospheric absorption, to ensure accurate measurement of the image radiation. For image distortion caused by lens distortion, change in viewing angle or uneven lighting conditions, geometric correction is required. Geometric correction restores the true geometric shape of the image through mathematical transformation.

[0067] Furthermore, aligning the acquired image data is the process of accurately matching the image data of different modalities, such as visible light, infrared, and depth, to the same spatial coordinate system. This step is crucial for multimodal fusion, ensuring that the image data obtained from different sensors can be superimposed on each other and reflect the same anatomical or physical structure. For example, structural images such as MRI and functional images such as fMRI need to be accurately aligned to ensure the accuracy of the analysis;

[0068] When denoising the acquired image data, filters such as Gaussian filters and median filters should be used to reduce random noise or spots in the image while trying to maintain image details. For time series data such as fMRI sequences, specific time series analysis methods may be needed to remove noise that varies over time, such as ARIMA models or sliding window averages.

[0069] The image is converted to the frequency domain through Fourier transform, and then the high-frequency noise components are removed from the spectrum and then inversely transformed back to the spatial domain;

[0070] Furthermore, the collected image data is normalized after image correction, alignment, and denoising, and the grayscale value range of the image is scaled to a predetermined interval such as 0 to 1 or -1 to 1, so that different images are comparable and convenient for subsequent processing and comparison. The feature vectors are standardized, such as z-score standardization, to ensure that different features have the same scale and avoid some features dominating the model learning process due to their large value range. The data is scaled to fall into a specific small interval, such as [-1,1] or [0,1]. This is necessary for many machine learning algorithms because they are sensitive to the scale of the input data;

[0071] It should be noted that data preprocessing is the basis for ensuring image data quality, improving analysis accuracy, and promoting multimodal data fusion. Through these meticulous processing steps, errors in the analysis process can be effectively reduced and the performance of the final recognition or classification model can be improved.

[0072] S3. Through the eigenface analysis algorithm, heat map region segmentation, and 3D geometric feature description, the feature vectors of each modality of visible light image features, infrared heat map features, and depth features in the multimodal facial feature standard dataset are extracted.

[0073] Furthermore, through the eigenface analysis algorithm, heat map region segmentation, and 3D geometric feature description, the feature vectors of each modality of visible light image features, infrared heat map features, and depth features in the multimodal facial feature standard data set are extracted. The specific steps are as follows:

[0074] The expression of the multimodal facial feature standard dataset P is:

[0075] P={P vis ,P ir ,P dep}

[0076] Among them, P vis is the visible light image, P ir is the infrared image, P dep is the depth image;

[0077] Based on the multimodal facial feature standard data set, the PCA method is used to extract the eigenface basis matrix V, which is expressed as:

[0078]

[0079] Among them, v i is the i-th eigenvector, λ i is the non-zero eigenvalue corresponding to the i-th eigenvector, and K is the variable parameter;

[0080] It should be noted that the relationship between the eigenface basis matrix V and the multimodal facial feature standard data set P is established through principal component analysis (PCA). PCA is a statistical method used for data dimensionality reduction and feature extraction. It looks for the principal components of the data, that is, the direction of maximum variance in the data. In this method, PCA is used to extract key features from the multimodal facial feature standard data set P to facilitate subsequent facial recognition and analysis. The eigenface basis matrix V is the result of PCA analysis. It consists of multiple eigenvectors, which are the eigenvectors of the covariance matrix of the original data set P. They are sorted by the square root of their respective eigenvalues ​​after standardization. The eigenvalue reflects the proportion of the eigenvector in the data explanation, that is, the amount of change in the data along that direction. Specifically, each column vector v in the eigenface basis matrix V iis a feature vector of the data set P, which represents a change pattern of the data. These feature vectors are orthogonal, meaning that they are independent of each other and have no overlapping information. The feature vector v i With eigenvalue λ i The larger the eigenvalue, the more the eigenvector can explain the variance of the data set, that is, the more data changes it captures; the variable parameter K determines how many eigenvectors to retain. Usually, we will select the first K eigenvectors with higher eigenvalues ​​because they can capture the main changes in the data set. In this way, we can use fewer eigenvectors to approximate the original data, thereby achieving data dimensionality reduction.

[0081] In summary, the relationship between the eigenface basis matrix V and the multimodal facial feature standard dataset P is: V is obtained by performing PCA analysis on P, and it contains the principal components of the dataset P, that is, the most important eigenvectors. These eigenvectors are used in the subsequent image feature extraction and recognition process. By projecting the image data into the feature space defined by V, the key features of the face can be effectively extracted for identity recognition.

[0082] The visible light image P is transformed into vis Projected into the feature space, generating the visible light image feature vector F vis The expression is:

[0083] F vis =V T ·P vis

[0084] Among them, V T is the transpose of the eigenface basis matrix V;

[0085] Preferably, the eigenface analysis algorithm is a dimensionality reduction technique commonly used in face recognition. First, based on the visible light image P in the multimodal facial feature standard dataset vis , the principal component analysis PCA method is applied. PCA calculates the covariance matrix of the data set and decomposes it to obtain a set of orthogonal eigenvectors v1, V2, ..., vk, which represent the main change directions of the original data. The importance of each eigenvector is measured by its corresponding eigenvalue λ1, λ2, ..., λk. Usually only the vectors corresponding to the larger eigenvalues ​​are retained to reduce the dimension. The eigenface basis matrix V is composed of these eigenvectors divided by the square root of their corresponding eigenvalues ​​in a standardized manner;

[0086] Furthermore, the Otsu method is used to perform binary segmentation of the infrared image P ir Get the hot zone map and cold zone map, calculate the average temperature and temperature variance in the hot zone map, and generate the infrared heat map feature vector F tm The expression is:

[0087] F tm =[T avg ,σ 2 ]

[0088] Among them, T avg is the average temperature of all measurement points in the infrared thermal image, σ 2 is the temperature variance of infrared thermal image;

[0089] Preferably, the Otsu method is an automatic threshold selection method that maximizes the inter-class variance to determine the optimal binarization threshold, and then calculates the average temperature T in the heat map. avg and temperature variance σ 2 , these two statistics constitute the infrared thermal map feature vector F tm , they can reflect the temperature distribution characteristics of the face.

[0090] It should be noted that the Otsu method is used to perform binary segmentation of the infrared image P ir Get the hot zone map, cold zone map and infrared heat map feature vector F tm , the specific steps are as follows:

[0091] First, from the infrared image P ir Calculate the gray level histogram;

[0092] Based on the histogram, the cumulative distribution function (CDF) is calculated, which will help us understand the sum of pixels below each gray level. Its expression is:

[0093]

[0094] Among them, j = 0 represents the starting value of the summation, starting from gray level 0, and i is the end value of the summation, that is, gray level i, which means that we accumulate all gray levels less than or equal to i, and the histogram H j represents the number of pixels of gray level j;

[0095] The Otsu algorithm is used to determine the optimal threshold T, which will minimize the internal variance between the background and foreground while maximizing the inter-class variance between them. The expression of the average gray value is:

[0096]

[0097] Where N is the total number of pixels in the image, and L is the maximum value of the grayscale level (e.g. 256);

[0098] For each possible threshold T, calculate the background w 0 and the foreground 1 The weights of the 0and μ 1 , the specific expression is as follows:

[0099]

[0100] w 1 (T) = 1-w 0 (T);

[0101]

[0102] Based on the above w 0 (T), w 1 (T), μ 0 (T) and μ 1 (T) value, calculate the between-class variance The expression is:

[0103]

[0104] Ultimately, the choice makes The maximum T is used as the optimal threshold to segment the infrared image Pir. All pixels with grayscale values ​​less than or equal to T are marked as cold areas (for example, the value is 0), and all pixels with grayscale values ​​greater than T are marked as hot areas (for example, the value is 1). If Pir(x, y) ≤ T, then B(x, y) = 0 (cold area), otherwise B(x, y) = 1 (hot area)

[0105] Separate the hot and cold maps from the binary image B. The hot map will contain only pixels with a value of 1 in B, while the cold map will contain pixels with a value of 0.

[0106] In the heat map, calculate the average temperature T avg And the temperature variance σ2, the expression is:

[0107]

[0108] Among them, x represents the column coordinate in the image, and y represents the row coordinate in the image;

[0109] Finally, we get the hot zone map, cold zone map and infrared heat map feature vector F tm , which contains the average temperature and temperature variance of the hot zone, which will be used as the input features for subsequent analysis. This process fully utilizes the advantages of the Otsu method, that is, automatically finding the optimal segmentation threshold, thereby effectively segmenting the infrared image into hot and cold areas, providing valuable information for subsequent feature extraction and recognition.

[0110] Furthermore, extract the depth map P depThe facial contours and key points in the image are usually significant signs of facial structure, such as the tip of the nose and the corners of the eyes. Secondly, the VLAD method is used to calculate local and global geometric features. VLAD calculates the difference between each descriptor and its cluster center and accumulates these differences to obtain the deep feature vector F. gm The expression is:

[0111]

[0112] Among them, VLAD is the local aggregation vector function, D loc is around the key point k i Local depth descriptors, which can contain information such as depth change and curvature, n is the total number of key points;

[0113] The visible light image feature vector F vis , infrared heat map feature vector F tm and the deep feature vector F gm Perform vector concatenation to form a multimodal feature vector set F mi The expression is: F mi =[F vis ,F tm ,F gm ];

[0114] It should be noted that this set integrates the structural features of visible light images, the temperature distribution features of infrared images, and the spatial geometric features of depth images, providing a rich information basis for subsequent identity recognition.

[0115] S4. According to the current time, geographic location information and historical data statistics, dynamically adjust the weights in each modal feature vector to form a comprehensive feature vector.

[0116] Furthermore, based on the collected current time t, geographic location information g and historical data statistics H v , dynamically adjust the weights of visible light, infrared and depth images respectively;

[0117] Dynamic weight adjustment expression w for visible light images v is:

[0118]

[0119] Among them, α is the constant coefficient, t is the time in hours, and β is the baseline weight;

[0120] It should be noted that the weight here changes periodically with time t in hours, and a sine function is used to simulate the impact of lighting conditions at different time periods of the day on the quality of visible light images. The constant coefficient α controls the amplitude of the weight change, while β is the baseline weight for the entire day, ensuring that there is a basic level of importance at all time points. For example, at night or during periods of low light, the weight may be lower because the recognition information of visible light images may not be as rich as other modalities at this time;

[0121] Furthermore, the infrared image dynamic weight adjustment expression w i r is:

[0122]

[0123] Among them, γ k is the basic weight of environment category k, reflecting the general importance of infrared image information under this environment type, e is the environmental factor code, and f t t is a linear function that changes smoothly over time. The weight adjustment here depends on the environment category k and the linear function that changes smoothly over time f t t, 1 g＝k is an indicator function. It takes the value of 1 when the current environmental factor e matches the category k, otherwise it takes the value of 0. It is used to determine whether the current environment is suitable for a specific weight adjustment. k is the environmental category, and N is the total number of environmental categories.

[0124] Depth image dynamic weight adjustment expression w g m is:

[0125]

[0126] Among them, δ 0 ,δ 1, δ 2 To adjust the model parameters of the weights, it is used to adjust the importance of historical data impact and geographical location. is the average value of historical data statistics;

[0127] Based on the dynamic weights of the visible light image, infrared image and depth image, a comprehensive feature vector F is formed. ce The expression is: F ce =w v is·F vis +w i r.F tm +w g m·F gm .

[0128] It should also be noted that the comprehensive feature vector F ce, integrates the information of all modalities, and optimizes the adaptability to the current environment and conditions by dynamically adjusting the weights, thereby improving the accuracy and robustness of identity recognition.

[0129] S5. Use the comprehensive feature vector to identify the identity of the target object through a machine learning classifier. When the recognition result reaches the preset confidence threshold, the recognition conclusion is output, otherwise it will return to the feature extraction step.

[0130] Furthermore, in order to verify the effectiveness and generalization ability of the model, the comprehensive feature vector F ce The model is divided into 70% training set, 15% validation set for adjusting hyperparameters, and 15% test set for final testing of model performance, and the comprehensive feature vector F is ce The training set is normalized by L2 norm and the expression is:

[0131]

[0132] Among them, F norm is the training set after L2 normalization;

[0133] Preferably, before training the model, the comprehensive feature vector F in the training set is ce Perform L2 norm normalization, that is, divide each eigenvector by its Euclidean length, which is the vector's own modulus. The purpose of this is to ensure that all eigenvectors have the same scale and avoid some eigenvalues ​​being too large to dominate the model learning, thereby improving the stability and convergence speed of the model;

[0134] Use cross-validation to select the best set of hyperparameters and train the gradient boosting tree model based on the L2-normalized training set;

[0135] Preferably, cross-validation is used to train and validate the model by repeatedly dividing the training set into different subsets, so as to evaluate the performance of the model on different subsets and select the best hyperparameter combination, such as the number of trees, learning rate, etc., to achieve the best generalization ability;

[0136] Furthermore, the generalization ability of the gradient boosted tree classification model was evaluated using the accuracy, recall, F1 score, and area under the AUC-ROC curve in the test set;

[0137] It should be noted that the performance of the GBT model is evaluated using the test set, focusing on several key indicators: Accuracy, Recall, F1 Score, and Area Under the Curve. These indicators help understand the frequency of the model identifying the correct target, the ability to identify all positive examples, the balance between precision and recall, and the model's ability to distinguish between positive and negative samples.

[0138] Based on the performance of the test set and combined with business needs, an initial recognition confidence threshold is set. This threshold is the standard for judging whether the recognition result is acceptable. If it is higher than this threshold, the recognition is considered successful, and if it is lower than this threshold, the recognition is considered failed.

[0139] Through the online learning mechanism, the performance of the gradient boosting tree classification model in actual applications is continuously monitored, user feedback and recognition results are collected, and the confidence threshold is dynamically adjusted according to the actual false recognition rate and missed recognition rate to find an optimal balance between the false recognition rate and missed recognition rate, so that the overall recognition performance is optimal. This means that the model will continuously optimize itself according to new recognition results and user feedback. If the false recognition rate or missed recognition rate is found to be too high in actual applications, the system will dynamically adjust the recognition confidence threshold according to these actual performances to achieve the best recognition effect and user experience;

[0140] When the probability score value output by the trained gradient boosting tree model is greater than or equal to the currently set confidence threshold, the identity of the target object is deemed to be successfully recognized and the corresponding recognition conclusion is output; otherwise, the recognition is considered insufficient and the feature re-extraction process is triggered, that is, returning to the image preprocessing step for re-recognition to obtain a more accurate recognition result. This mechanism ensures the system's self-correction capability and continuous optimization.

[0141] This embodiment also provides a computer device, which is suitable for the case of an artificial intelligence visual detection method based on a feature face method, comprising: a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the artificial intelligence visual detection method based on a feature face method as proposed in the above embodiment.

[0142] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0143] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the artificial intelligence visual detection method based on the feature face method proposed in the above embodiment is implemented; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0144] In summary, the present invention has created a highly integrated and adaptable artificial intelligence visual detection solution through: synchronous multimodal data acquisition, eigenface analysis, heat map and deep feature description and dynamic weight adjustment algorithm, which significantly enhances the accuracy and robustness of face recognition, and has important value and far-reaching impact on promoting the practical application of artificial intelligence visual technology.

[0145] Example 2, referring to Table 1, is the second example of the present invention. To further verify the advancement of the present invention, experimental simulation data of an artificial intelligence visual detection method based on a feature face method is provided.

[0146] Four volunteers of different ages, genders and skin colors were selected as test subjects to ensure the broad representativeness of the sample. An intelligent visual detection system equipped with a high-precision camera module (including visible light, infrared and depth sensors) was used. The system has adaptive exposure control and dynamic thermal balance adjustment functions to ensure image quality and reduce environmental interference. Image collection was carried out in different time periods (morning, noon and evening), different weather conditions (sunny, cloudy and rainy) and different locations (indoors, outdoors and semi-sheltered) to ensure that the data covers all-weather and multiple scenarios.

[0147] First, the system automatically and synchronously collects the visible light image P of the target object’s face. vis , infrared image P ir and the depth image P dep Subsequently, these images are processed with image correction, alignment, denoising and normalization to eliminate the effects of uneven illumination, geometric distortion and noise, forming a high-quality multimodal facial feature standard dataset P.

[0148] Secondly, the PCA method is used to extract the eigenface basis matrix V from the data set, and further generate the visible light image eigenvector F vis 、Infrared heat map feature vector F tm and the deep feature vector F gm The analysis of infrared thermal images also considers the uniformity of temperature distribution, while the depth map deeply analyzes the geometric features of facial contours. v , dynamically adjust the weight of each modal feature vector. For example, a lower weight is given to infrared features during the day and a higher weight is given at night; the weight of depth features is increased in indoor environments to cope with insufficient light.

[0149] Then, according to the adjusted weights, the modal eigenvectors are combined into a comprehensive eigenvector F ce , providing input to the machine learning model.

[0150] Finally, the comprehensive feature vector is trained, verified and tested using the gradient boosting tree model, the model is optimized based on indicators such as accuracy and recall, and the initial recognition confidence threshold is set. The model is continuously optimized through the online learning mechanism to ensure recognition performance.

[0151] The details are shown in Table 1:

[0152] Table 1 Experimental record table

[0153]

[0154]

[0155] By comparing the experimental data and analyzing the data in the above table, it can be clearly seen that the effectiveness and superiority of the method of the present invention are generally higher than that of the traditional single-modal recognition system. For example, in indoor morning conditions, although the quality of visible light images is slightly inferior for V001, due to the effective complementarity of depth features and infrared features, the comprehensive accuracy reaches 0.98, which is about 10% higher than that of the single-modal system. In the afternoon outdoor environment, V002, due to direct sunlight, which may cause the quality of visible light images to deteriorate, the system automatically reduces the weight of visible light features and increases the weight of infrared features, maintaining a high accuracy of 0.96, which proves the effectiveness of the dynamic weight adjustment strategy. In the semi-shaded environment at night, although the light conditions are not good, the high precision of the depth features and the uniformity index of the infrared features work together to ensure the accurate recognition of the system, showing the strong adaptability of the method to complex environments.

[0156] Comparing V001 and V004, although both are in the morning, the different geographical locations significantly affect the recognition effect. V001 performs better indoors, which shows that the system can dynamically optimize according to the specific environment and improve the robustness of recognition.

[0157] The present invention uses an artificial intelligence visual detection system based on the feature face method, and comprehensively utilizes the synchronous acquisition and advanced preprocessing of visible light, infrared and depth images to achieve effective extraction of multimodal features. The dynamic weight adjustment mechanism optimizes the feature vector weights according to time, geographic location and historical data, significantly improving the recognition accuracy and robustness in a variety of environments. Experimental data proves that this method improves the comprehensiveness and accuracy of feature expression through PCA, heat map analysis, 3D geometric description and other technologies, especially in complex light and environmental changes. The model ensures a high degree of recognition accuracy and good generalization ability through sophisticated training and online learning strategies, showing significant innovation and advantages over existing technologies, and enhancing its practical value and future application potential in the field of identity authentication.

[0158] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An artificial intelligence visual detection method based on eigenface method, characterized by: include, Synchronously collect image data of visible light images, infrared images and depth images of the target object's face, and perform data preprocessing to form a multimodal facial feature standard data set; Extracting feature vectors of each modality of visible light image features, infrared thermal image features and depth features in the multimodal facial feature standard data set through eigenface analysis algorithm, thermal image region segmentation and 3D geometric feature description; According to the current time, geographic location information and historical data statistics, dynamically adjust the weights in the feature vectors of each modality to form a comprehensive feature vector; Using the comprehensive feature vector, identifying the identity of the target object through a machine learning classifier; When the recognition result reaches the preset confidence threshold, the recognition conclusion is output, otherwise it returns to the feature extraction step; The visible light image feature vector in the multimodal facial feature standard data set is extracted by using a feature face analysis algorithm, heat map region segmentation, and 3D geometric feature description. The specific steps are as follows: The multimodal facial feature standard dataset P is expressed as: P={P vis ,P ir ,P dep } Among them, P vis is the visible light image, P ir is the infrared image, P dep is the depth image; Based on the multimodal facial feature standard data set, the PCA method is used to extract the eigenface basis matrix V, which is expressed as: Among them, v i is the i-th eigenvector, λ i is the non-zero eigenvalue corresponding to the i-th eigenvector, and K is the variable parameter; The visible light image P is transformed into vis Projected into the feature space, generating the visible light image feature vector F vis The expression is: F vis =V T ·P vis Among them, V T is the transpose of the eigenface basis matrix V; Extracting infrared thermal image features and depth feature vectors from the multimodal facial feature standard dataset, specifically, the steps are: Binarization segmentation of infrared image P using Otsu method ir Get the hot zone map and cold zone map, calculate the average temperature and temperature variance in the hot zone map, and generate the infrared heat map feature vector F tm The expression is: F tm =[T avg ,s 2 ] Among them, T avg is the average temperature of all measurement points in the infrared thermal image, σ 2 is the temperature variance of infrared thermal image; Extract depth map P dep The facial contour and key points in the image are used to calculate local and global geometric features and obtain the deep feature vector F. gm The expression is: Among them, VLAD is the local aggregation vector function, D loc is around the key point k i The local depth descriptor of , n is the total number of key points; The visible light image feature vector F vis , infrared heat map feature vector F tm and the deep feature vector F gm Perform vector concatenation to form a multimodal feature vector set F mi The expression is: F mi =[F vis ,F tm ,F gm ].

2. The artificial intelligence visual inspection method based on the eigenface method as claimed in claim 1, characterized in that: The synchronous acquisition process includes utilizing adaptive exposure control and dynamic thermal balance adjustment to optimize image quality and reduce interference from environmental factors.

3. The artificial intelligence visual inspection method based on the eigenface method as claimed in claim 2, characterized in that: The data preprocessing includes image correction, alignment, denoising and normalization.

4. The artificial intelligence visual inspection method based on the eigenface method as claimed in claim 3, characterized in that: According to the current time, geographic location information and historical data statistics, the weights in the modal feature vectors are dynamically adjusted to form a comprehensive feature vector. The specific steps are as follows: Based on the collected current time t, geographic location information g and historical data statistics H v , dynamically adjust the weights of visible light, infrared and depth images respectively; Dynamic weight adjustment expression w for visible light images v is: Among them, α is the constant coefficient, t is the time in hours, and β is the baseline weight; Dynamic weight adjustment expression w for infrared images i r is: Among them, γ k is the basic weight of environment category k, f t (t) is a linear function that changes smoothly over time, 1 g＝k is the indicator function, e is the environmental factor code, k is the environmental category, and N is the total number of environmental categories; Depth image dynamic weight adjustment expression w g m is: Among them, δ0,δ 1, δ2 is the model parameter for adjusting weights, is the average value of historical data statistics; Based on the dynamic weights of the visible light image, infrared image and depth image, a comprehensive feature vector F is formed. ce The expression is: F ce =w v is·F vis +w i r.F tm +w g m·F gm .

5. The artificial intelligence visual inspection method based on the eigenface method as claimed in claim 4, characterized in that: The comprehensive feature vector is used to identify the identity of the target object through a machine learning classifier, and the specific steps are: The comprehensive feature vector F ce The comprehensive feature vector F is divided into 70% training set, 15% validation set and 15% test set. ce The training set is normalized by L2 norm and the expression is: Among them, F norm is the training set after L2 normalization; Use cross-validation to select the best set of hyperparameters and train the gradient boosting tree model based on the L2-normalized training set; The generalization ability of the gradient boosting tree classification model was evaluated using the accuracy, recall, F1 score, and area under the AUC-ROC curve in the test set; Based on the performance of the test set and combined with business requirements, set an initial recognition confidence threshold; Through the online learning mechanism, the performance of the gradient boosting tree classification model in practical applications is continuously monitored, user feedback and recognition results are collected, and the confidence threshold is dynamically adjusted to the optimal value according to the actual false recognition rate and missed recognition rate.

6. The artificial intelligence visual inspection method based on the eigenface method as claimed in claim 5, characterized in that: When the recognition result reaches the preset confidence threshold, the recognition conclusion is output, otherwise it returns to the feature extraction step, and the specific steps are: When the probability score value output by the trained gradient boosting tree model is greater than or equal to the currently set confidence threshold, the target object is deemed to have been successfully identified and the corresponding identification conclusion is output; otherwise, the identification is considered insufficient and the feature re-extraction process is triggered, i.e., returning to the image preprocessing step for re-identification.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the artificial intelligence visual detection method based on the eigenface method described in any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the artificial intelligence visual detection method based on the eigenface method described in any one of claims 1 to 6 are implemented.

Images