A method for intelligent facial recognition of people in smart security cameras

Abstract

This application relates to the field of image recognition technology, specifically to an intelligent facial recognition method for smart security cameras. This method analyzes the impact of noise in low-light environments on facial images acquired by the security camera. During image filtering, based on the differences in directionality and continuity between noise and facial texture, an adaptive strategy for adjusting the filtering intensity is proposed. This effectively distinguishes between different distribution characteristics of facial texture features and noise, thereby optimizing the noise filtering effect and reducing errors in facial key point recognition and localization. In this way, the accuracy and stability of facial recognition based on smart security cameras are further improved.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, specifically to a method for intelligent recognition of people's appearance using smart security cameras. Background Technology

[0002] With the acceleration of urbanization, urban population density is increasing, and mobility is gradually strengthening, making security issues increasingly important. In terms of access control, traditional security systems are widely used in office buildings, residential areas, and important institutions. However, these systems rely on shift management of security personnel and reactive responses, facing structural problems such as response delays, low recognition accuracy, and insufficient data integration capabilities. With technological advancements, intelligent facial recognition technology based on smart security cameras, with its significant advantages of non-contact operation, high precision, and high efficiency, has become a core technology of modern security systems and has demonstrated enormous potential in improving public safety.

[0003] However, in low-light or nighttime environments, severe noise interference can obscure facial skin texture, eyebrows, and other detailed features. Non-local means denoising algorithms can preserve texture edge details while removing noise, and are therefore often used to address this type of noise problem. Current techniques typically use fixed filter coefficients to apply non-local means denoising algorithms, but they fail to consider that noise interference can severely impact facial feature extraction and matching in low-light conditions. Specifically, noise can lead to inaccurate localization of facial feature points, resulting in matching and recognition anomalies. Summary of the Invention

[0004] In view of the above, it is necessary to provide a method for intelligent facial recognition of people in smart security cameras to solve the above problems.

[0005] One embodiment of this application provides a method for intelligent facial recognition of people using a smart security camera, the method comprising: Images are captured using security cameras, and facial region detection is performed to obtain images containing facial information; The image containing facial information is cropped to obtain the Region of Interest (ROI) image; image patches are set for each pixel in the ROI image, and a filter bank is used to obtain high-dimensional feature vectors; the pixels in the ROI image are filtered based on the gradient magnitude to obtain the pixels to be analyzed; the similarity and energy distribution characteristics of the high-dimensional feature vectors of each pixel to be analyzed and the pixels in its local range are analyzed to determine the comprehensive feature value of each pixel to be analyzed. Based on the comprehensive feature values, all pixels to be analyzed in the ROI region image are classified. In the image block corresponding to each pixel, the local feature value of each pixel is determined according to the difference in the number distribution of various types of pixels. The similarity features of the gray-level distribution of pixels between rows and columns are analyzed to obtain the distribution feature value of each pixel. Based on the local feature values ​​and distribution feature values, the optimal filtering coefficient of each pixel is determined, and the ROI region image is filtered and enhanced. Appearance matching is performed on the filtered and enhanced ROI region image to obtain the person's appearance recognition result.

[0006] Preferably, the pixels to be analyzed are specifically pixels whose gradient magnitude is greater than or equal to the optimal threshold; the optimal threshold is obtained by thresholding the gradient magnitude of all pixels in the ROI region image.

[0007] Preferably, determining the comprehensive feature value of each pixel to be analyzed specifically involves: Obtain the average similarity between the high-dimensional feature vectors of each pixel to be analyzed and all pixels in its eight neighborhoods; Principal component analysis is performed on the high-dimensional feature vectors corresponding to each pixel to be analyzed and all pixels in its eight neighborhoods, and the ratio of the largest eigenvalue to the sum of all eigenvalues ​​is calculated. The ratio obtained for each pixel to be analyzed is positively fused with the mean similarity value to obtain the comprehensive feature value of each pixel to be analyzed.

[0008] Preferably, the classification of all pixels to be analyzed in the ROI region image specifically involves: Threshold segmentation is performed on the comprehensive feature values ​​of all pixels to be analyzed in the ROI region image. Pixels to be analyzed with comprehensive feature values ​​greater than or equal to the segmentation threshold are designated as undetermined texture pixels, while pixels to be analyzed with comprehensive feature values ​​less than the segmentation threshold are designated as undetermined noise pixels.

[0009] Preferably, the local feature value of each pixel is obtained by the proportion of the difference between the total number of undetermined texture pixels and the total number of undetermined noise pixels in the total number of pixels to be analyzed in the ROI region image.

[0010] Preferably, the process of obtaining the distribution feature value of each pixel is as follows: Based on an image block containing any pixel, determine the gray-level feature value of each pixel according to the gray-level distribution within the local area of ​​each pixel in the image block; The gray-level feature values ​​of all pixels in each row of the image block are used to form the feature vector of each row, and the Euclidean distance between the feature vector of each row and the feature vectors of the other rows is calculated. The gray-level feature values ​​of all pixels in each column of the image block are used to form the feature vector of each column, and the Euclidean distance between the feature vector of each column and the feature vectors of the other columns is calculated. The mean of all Euclidean distances obtained from the image patch is calculated and used as the distribution feature value of any given pixel.

[0011] Preferably, the grayscale feature value of each pixel is specifically the mean of the absolute values ​​of the grayscale differences between each pixel and its eight neighboring pixels.

[0012] Preferably, the formula for determining the optimal filtering coefficient for each pixel is as follows: in, This represents the optimal filtering coefficient corresponding to pixel a; This is the normalization function; This represents the distribution characteristic value corresponding to pixel a; This represents the local feature value corresponding to pixel a; These are the preset base filter coefficients.

[0013] Preferably, the specific process of performing appearance matching on the filtered and enhanced ROI region image to obtain the person's appearance recognition result is as follows: Pose correction is performed on the filtered and enhanced ROI region, and the coordinates of a preset number of facial key points are detected and obtained. The facial key point coordinates are mapped, and the FaceNet model is used to obtain the correlation between the facial image corresponding to the ROI region and all facial data in the facial database. Based on the numerical value of the degree of correlation, the results of personnel appearance recognition are obtained.

[0014] Preferably, obtaining the person's appearance recognition result specifically includes: If the facial database contains data with a correlation greater than a preset similarity threshold with the corresponding facial image, the identity information corresponding to the image with the highest correlation is selected as the recognition result; otherwise, the recognition fails.

[0015] This application has at least the following beneficial effects: This application uses security cameras to capture images and employs a facial region detection method to accurately locate and extract image regions containing facial information. This effectively eliminates irrelevant background information, ensuring that subsequent processing focuses solely on the facial region, thus improving computational efficiency and recognition accuracy. By cropping the ROI (Region of Interest) image containing facial information, the processing of the region of interest can be concentrated, reducing interference from other backgrounds, thereby optimizing processing speed and enhancing focus on facial details, improving the accuracy of subsequent image analysis. Dividing the image into blocks and extracting high-dimensional feature vectors refines the information of each pixel, ensuring that every detail of the face (such as skin texture, eyebrows, and corners of the eyes) is accurately captured. High-dimensional feature vectors can more comprehensively express the features of each pixel, improving the ability to recognize facial information. Filtering pixels by gradient magnitude focuses attention on areas with significant changes in facial features, eliminating irrelevant smooth areas or noise. This helps improve the recognition accuracy of key feature points (such as corners of the eyes, bridge of the nose, and corners of the mouth) while reducing interference from irrelevant points. By analyzing the similarity and energy distribution between the pixel to be analyzed and surrounding pixels, facial texture information can be extracted more accurately. This method helps improve facial feature recognition capabilities, especially under complex lighting conditions, effectively identifying and preserving detailed information. By classifying pixels based on comprehensive feature values, it effectively distinguishes features of different facial regions, further optimizing the image segmentation and recognition process. Similarity analysis of pixel row and column grayscale distributions reveals local texture information in facial regions, further improving the robustness of facial recognition in low light or noisy environments. Based on the local and distributional features of each pixel, optimal filtering coefficients are selected for filtering and enhancement. This not only removes noise but also enhances facial details, making facial features clearer, especially in low-light environments, effectively improving facial image quality and increasing recognition accuracy. Appearance matching of the processed images accurately identifies and distinguishes different individuals. Combined with the preceding image preprocessing and feature extraction, appearance recognition becomes more stable and efficient, enabling real-time monitoring and identity recognition in dynamic or complex environments. Attached Figure Description

[0016] Figure 1 A flowchart of a method for intelligent facial recognition of people using a smart security camera, provided in this application; Figure 2 A flowchart for obtaining the optimal filter coefficients provided in this application. Detailed Implementation

[0017] In the description of the embodiments in this application, the words "exemplary," "or," and "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplary," "or," and "for example" is intended to present the relevant concepts in a specific manner.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this application's specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0019] It should also be noted that the terms "first" and "second" in this application and its accompanying drawings are used to distinguish similar objects, rather than to describe a specific order or sequence. The methods disclosed in the embodiments of this application or the methods shown in the flowcharts include one or more steps for implementing the method. Without departing from the scope of protection of this application, the execution order of multiple steps can be interchanged, and some steps can also be deleted.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0021] This application proposes an intelligent facial recognition method for smart security cameras, applied in the field of image recognition technology. (See attached document.) Figure 1 The method includes the following steps: S1: Use security cameras to capture images and perform facial region detection to obtain images containing facial information.

[0022] In the process of intelligent facial recognition based on smart security cameras, images of the corresponding scene are first acquired using the smart security cameras and transmitted to the backend server. Further, a publicly available facial recognition dataset is used to train a facial region detection model using the YOLOv10 algorithm. The YOLOv10 algorithm is a well-known technology and will not be described in detail here. The publicly available facial recognition dataset in this application was obtained through publicly available internet channels.

[0023] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0024] On the backend server, a face region detection model trained using the YOLOv10 algorithm is used to detect images acquired in real time by the smart security camera. The detection confidence level is set to 0.66 to obtain images containing facial information.

[0025] S2: Cropping the image containing facial information yields the ROI region image; setting image patches for each pixel in the ROI region image and using a filter bank to obtain high-dimensional feature vectors; filtering the pixels in the ROI region image based on gradient magnitude to obtain the pixels to be analyzed; analyzing the similarity and energy distribution characteristics of the high-dimensional feature vectors of each pixel to be analyzed and its local pixels to determine the comprehensive feature value of each pixel to be analyzed.

[0026] After acquiring images containing facial information, key facial regions are filtered. Since conventional facial recognition typically focuses on the closest individuals, images containing facial information undergo further filtering. The facial region with the largest area is selected as the Region of Interest (ROI), and then cropped to obtain the image of that ROI. Furthermore, images acquired at night or in low-light conditions are often subject to significant noise interference, which may obscure details such as facial skin texture and eyebrows. During facial feature extraction and matching, noise interference can lead to inaccurate localization of feature points (such as the corners of the eyes and mouth), thus affecting the accuracy of facial recognition and matching. Therefore, preprocessing of the ROI image is necessary to reduce noise interference.

[0027] Considering that traditional denoising algorithms blur edges and textures while smoothing noise, destroying facial details such as facial contours, wrinkles, and moles, and affecting facial recognition performance, while non-local means denoising algorithms can preserve texture and edge details while denoising, this application uses a non-local means denoising algorithm to filter and denoise the image. The specific steps are as follows: In the ROI region image, a search window and image patch size are set for each pixel. When performing feature matching on the facial region, since there are many non-local similar structures on the face (such as symmetrical facial areas), using a larger search window has a higher probability of finding similar patches, thus improving the denoising quality. Therefore, in this embodiment, a window with a side length of 21 is set as the search window for each pixel in the ROI region image. When the number of pixels in the image cannot satisfy a search window of sufficient size, the region containing pixels within the corresponding search window area is selected as its search window. For example, when there is only one pixel to the left of a pixel, and the horizontal axis width of the window cannot meet the requirement of 21 pixels, the missing pixel on the left does not need to be supplemented, and the region containing pixels within the corresponding search window size is directly obtained as the search window for that pixel. At the same time, a window with a side length of 7 is set as the image patch for each pixel in the ROI region image. A larger image patch can better capture local facial features (such as eyes and lips), improve the stability of image patch matching, and combat the problem of low signal-to-noise ratio in low light. When the number of pixels in an image is insufficient to form a sufficiently large image block, this application selects the region containing pixels within the corresponding image block region as its image block.

[0028] Using the ROI region image as input, a Gabor filter bank is used to calculate the high-dimensional feature vector for each pixel in the image, obtaining the high-dimensional feature vector corresponding to each pixel. Obtaining high-dimensional feature vectors based on the Gabor filter bank is well-known and will not be elaborated further. In this embodiment, the scale parameter of the Gabor filter bank is set to a standard value of 5, and the specific side lengths of the scale windows are as follows: The angle parameters are set to 8, and the specific angles are as follows: .

[0029] After images are acquired through smart security cameras, vectors from areas such as skin texture and eyebrows typically exhibit significant and continuous responses in a few adjacent dimensions, forming a clear dominant direction. These features show high similarity and correlation across multiple scales, while the vectors of noise points show smaller and more chaotic responses in all dimensions, lacking a clear dominant direction. The responses are discontinuous and jump between different directions, resulting in low overall correlation.

[0030] Based on the above analysis, the gradient magnitude of all pixels in the image is first obtained. The Otsu thresholding method is used to calculate the maximum inter-class variance to obtain the optimal threshold. Since texture, edge and noise points in the facial region usually have large gradient magnitudes, while pixels of smooth skin do not have large gradient magnitudes, all pixels with gradient magnitudes greater than or equal to the optimal threshold are taken as pixels to be analyzed; otherwise, the pixels are taken as smooth pixels and will not be analyzed in the future.

[0031] The analysis process involves examining the pixels in the image to obtain the average similarity between each pixel and the high-dimensional feature vectors corresponding to all pixels in its eight-neighborhood. In this embodiment, the similarity of the high-dimensional feature vectors is calculated using cosine similarity. A higher average similarity indicates that the feature vectors of the pixel are more aligned with the surrounding pixels, resulting in better spatial continuity. Furthermore, the high-dimensional feature vectors corresponding to each pixel and all pixels in its eight-neighborhood are used as input, and Principal Component Analysis (PCA) is employed to calculate the eigenvalues ​​of the covariance matrix. The output is the eigenvalues ​​corresponding to each pixel and its eight-neighborhood. The PCA algorithm is a well-known technique and will not be elaborated upon further. Considering that after PCA dimensionality reduction, when the feature distribution is relatively concentrated, the largest eigenvalue usually contains the majority of the energy, based on this, the ratio of the largest eigenvalue to the sum of all eigenvalues ​​is calculated among the eigenvalues ​​corresponding to each pixel to be analyzed and its eight neighboring pixels. This ratio is then positively fused with the average similarity value to obtain the comprehensive feature value for each pixel to be analyzed. In this embodiment, the positive fusion of multiple variables is performed using a multiplication method. A larger comprehensive feature value indicates a more concentrated local feature distribution at the pixel, a smoother spatial distribution, and a closer alignment of the feature vector directions between the pixel and its surrounding points, consistent with the distribution characteristics of skin texture pixels.

[0032] S3: Based on the comprehensive feature values, classify all pixels to be analyzed in the ROI region image. In the image block corresponding to each pixel, determine the local feature value of each pixel according to the difference in the number distribution of various types of pixels. Analyze the similarity features of the gray-level distribution of pixels between rows and columns to obtain the distribution feature value of each pixel. Based on the local feature values ​​and distribution feature values, determine the optimal filtering coefficient for each pixel and perform filtering and enhancement processing on the ROI region image.

[0033] Furthermore, the comprehensive feature values ​​of all pixels to be analyzed are used as input, and the Otsu thresholding method is used to obtain the segmentation threshold. The Otsu thresholding method is a well-known technique and will not be elaborated further. Pixels to be analyzed with comprehensive feature values ​​greater than or equal to the segmentation threshold are designated as texture pixels, and pixels with comprehensive feature values ​​less than the segmentation threshold are designated as noise pixels. In the image block corresponding to each pixel, the proportion of the difference between the total number of texture pixels and the total number of noise pixels to be analyzed is calculated; a very small positive number 0.1 is added to the denominator to ensure that the value is not zero. Min-Max normalization is used to normalize the proportion values ​​corresponding to all pixels in the ROI region image. The normalization range is [0.1, 1], ensuring that the value is not zero. The normalized value is used as the local feature value of each pixel. The larger the local feature value, the more dominant the texture distribution features are within the image block corresponding to the pixel.

[0034] Considering that facial skin textures (such as pore clusters and fine wrinkles) typically exhibit a large number of similar or repetitive local distribution features in local areas, resulting in a relatively uniform and orderly distribution of texture features; while noise points are usually randomly generated in low-light scenes, leading to a more chaotic and disordered distribution of noise in the image, typically exhibiting irregular distribution in local areas, therefore, based on the image block corresponding to each pixel, the mean of the absolute value of the gray-level difference between each pixel and its eight neighboring pixels in the image block is calculated, and the obtained mean is used as the gray-level feature value of the corresponding pixel. Furthermore, the grayscale feature values ​​of all pixels in each row of the image block are used to form the feature vector of each row, and the Euclidean distance between the feature vector of each row and the feature vectors of the other rows is calculated. Correspondingly, the grayscale feature values ​​of all pixels in each column of the image block are used to form the feature vector of each column, and the Euclidean distance between the feature vector of each column and the feature vectors of the other columns is calculated. The mean of all the Euclidean distances obtained for the image block is calculated as the distribution feature value of each pixel in the image block. The larger the obtained distribution feature value, the more disordered the feature distribution within the image block corresponding to the pixel is, and it cannot represent similar or repetitive features, which is more in line with the distribution characteristics of complex noise. The smaller the obtained distribution feature value, the more uniform the feature distribution within the image block corresponding to the pixel is, and the local distribution features of pixels in different rows and columns show more obvious similarity and repetition, which is in line with the distribution characteristics of facial texture.

[0035] Furthermore, considering the differences in facial texture and noise features within the image, and taking into account the significant noise characteristics, it is necessary to enhance noise filtering to prevent noise interference from affecting the accurate localization of feature points (such as the corners of the eyes and mouth). Conversely, for images with prominent facial skin texture features, it is necessary to reduce the filtering effect to retain more detailed texture features and prevent the texture from being excessively smoothed. Based on this, the optimal filtering coefficients for each pixel are obtained as follows: in, This represents the optimal filtering coefficient corresponding to pixel a; For the normalization function, the Min-Max normalization algorithm is used in this embodiment for normalization, and the normalization range is [0.1,1]. This represents the distribution characteristic value corresponding to pixel a; This represents the local feature value corresponding to pixel a; The preset base filter coefficients are set to a standard value of 10. The flowchart for obtaining the optimal filter coefficients is as follows: Figure 2 As shown.

[0036] It should be noted that the formula for calculating the optimal filter coefficients only applies to the pixels to be analyzed; the basic filter coefficients are still used for smoothing pixels.

[0037] Existing technologies typically use conventional values ​​as filtering coefficients for nonlocal means denoising algorithms. However, they fail to consider the severe noise interference introduced by low light conditions at night, obscuring delicate features such as facial skin texture and eyebrows. During facial feature acquisition and matching, noise interference prevents accurate localization of feature points (such as the corners of the eyes and mouth), leading to matching and recognition anomalies. Therefore, this application fully considers the differences in skin texture and interfering noise characteristics in facial images under complex noise conditions after ROI cropping in low light. For each pixel in the image, the optimal filtering coefficient is calculated based on its local and distribution feature values. Then, using this optimal filtering coefficient, nonlocal means denoising is performed within a set search window.

[0038] After filtering and denoising the ROI region image, in order to improve the image matching and recognition effect, the denoised ROI region image is used as input, and the contrast-limited adaptive histogram equalization algorithm is used for image enhancement processing. The contrast-limited adaptive histogram equalization algorithm is a well-known technology, and will not be described in detail here; thus, the preprocessing operation for the ROI region image is completed.

[0039] S4: Perform facial matching on the filtered and enhanced ROI region image to obtain the facial recognition results.

[0040] A facial database containing facial images of registered individuals and their corresponding identity information is pre-acquired. After acquiring the pre-processed Region of Interest (ROI) images, face alignment is performed to eliminate the influence of pose changes. An affine transformation is used to rotate the face to a standard pose (eyes horizontal, nose tip centered). Using the pose-corrected ROI image as input, facial feature point detection is performed based on a publicly available active shape model, identifying and labeling 68 facial keypoint coordinates. Affine transformation and active shape models are well-known techniques and will not be elaborated upon further.

[0041] Furthermore, the labeled ROI region image is scaled to a preset size as input, and the FaceNet model is used for spatial vector mapping to map the face region in the ROI region image into a 128-dimensional Euclidean space vector. In this embodiment, the preset size is set to... Pixels; the FaceNet model is a well-known technology, so details will not be elaborated further. The cosine similarity is calculated between the facial image mapped to a 128-dimensional Euclidean space vector and the corresponding 128-dimensional Euclidean space vector of each facial data in the facial database to obtain the correlation between the facial image corresponding to the ROI region and all facial data in the facial database.

[0042] A similarity threshold is set; in this embodiment, the similarity threshold is 0.7. When the correlation between the facial image corresponding to the ROI region and the facial data in the facial database is greater than the similarity threshold, the image with the highest correlation with the corresponding facial image in the facial database is selected for matching and information feedback. When there is no image in the facial database with a correlation greater than the similarity threshold with the acquired facial image, the matching fails and feedback is provided. Thus, intelligent recognition of personnel appearance based on smart security cameras is completed.

[0043] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description; sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0044] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for intelligent facial recognition of people using smart security cameras, characterized in that, The method includes the following steps: Images are captured using security cameras, and facial region detection is performed to obtain images containing facial information; The image containing facial information is cropped to obtain the Region of Interest (ROI) image; image patches are set for each pixel in the ROI image, and a filter bank is used to obtain high-dimensional feature vectors; the pixels in the ROI image are filtered based on the gradient magnitude to obtain the pixels to be analyzed; the similarity and energy distribution characteristics of the high-dimensional feature vectors of each pixel to be analyzed and the pixels in its local range are analyzed to determine the comprehensive feature value of each pixel to be analyzed. Based on the comprehensive feature values, all pixels to be analyzed in the ROI region image are classified. In the image block corresponding to each pixel, the local feature value of each pixel is determined according to the difference in the number distribution of various types of pixels. The similarity features of the gray-level distribution of pixels between rows and columns are analyzed to obtain the distribution feature value of each pixel. Based on the local feature values ​​and distribution feature values, the optimal filtering coefficient of each pixel is determined, and the ROI region image is filtered and enhanced. Appearance matching is performed on the filtered and enhanced ROI region image to obtain the person's appearance recognition result.

2. The intelligent facial recognition method for smart security cameras as described in claim 1, characterized in that, The pixels to be analyzed are specifically pixels whose gradient magnitude is greater than or equal to the optimal threshold; the optimal threshold is obtained by thresholding the gradient magnitude of all pixels in the ROI region image.

3. The intelligent facial recognition method for smart security cameras as described in claim 1, characterized in that, The determination of the comprehensive feature value of each pixel to be analyzed is specifically as follows: Obtain the average similarity between the high-dimensional feature vectors of each pixel to be analyzed and all pixels in its eight neighborhoods; Principal component analysis is performed on the high-dimensional feature vectors corresponding to each pixel to be analyzed and all pixels in its eight neighborhoods, and the ratio of the largest eigenvalue to the sum of all eigenvalues ​​is calculated. The ratio obtained for each pixel to be analyzed is positively fused with the mean similarity value to obtain the comprehensive feature value of each pixel to be analyzed.

4. The intelligent facial recognition method for smart security cameras as described in claim 1, characterized in that, The classification of all pixels to be analyzed in the ROI region image is specifically as follows: Threshold segmentation is performed on the comprehensive feature values ​​of all pixels to be analyzed in the ROI region image. Pixels to be analyzed with comprehensive feature values ​​greater than or equal to the segmentation threshold are designated as undetermined texture pixels, while pixels to be analyzed with comprehensive feature values ​​less than the segmentation threshold are designated as undetermined noise pixels.

5. The intelligent facial recognition method for smart security cameras as described in claim 4, characterized in that, The local feature value of each pixel is obtained by taking the proportion of the difference between the total number of undetermined texture pixels and the total number of undetermined noise pixels in the total number of pixels to be analyzed in the ROI region image.

6. The method for intelligent facial recognition of personnel in a smart security camera as described in claim 1, characterized in that, The process of obtaining the distribution feature value of each pixel is as follows: Based on an image block containing any pixel, determine the gray-level feature value of each pixel according to the gray-level distribution within the local area of ​​each pixel in the image block; The gray-level feature values ​​of all pixels in each row of the image block are used to form the feature vector of each row, and the Euclidean distance between the feature vector of each row and the feature vectors of the other rows is calculated. The gray-level feature values ​​of all pixels in each column of the image block are used to form the feature vector of each column, and the Euclidean distance between the feature vector of each column and the feature vectors of the other columns is calculated. The mean of all Euclidean distances obtained from the image patch is calculated and used as the distribution feature value of any given pixel.

7. The intelligent facial recognition method for smart security cameras as described in claim 6, characterized in that, The grayscale feature value of each pixel is specifically the mean of the absolute values ​​of the grayscale differences between each pixel and its eight neighboring pixels.

8. The intelligent facial recognition method for smart security cameras as described in claim 1, characterized in that, The specific formula for determining the optimal filtering coefficient for each pixel is as follows: in, This represents the optimal filtering coefficient corresponding to pixel a; This is the normalization function; This represents the distribution characteristic value corresponding to pixel a; This represents the local feature value corresponding to pixel a; These are the preset base filter coefficients.

9. The intelligent facial recognition method for smart security cameras as described in claim 1, characterized in that, The specific process of performing appearance matching on the filtered and enhanced ROI region image to obtain the person's appearance recognition result is as follows: Pose correction is performed on the filtered and enhanced ROI region, and the coordinates of a preset number of facial key points are detected and obtained. The facial key point coordinates are mapped, and the FaceNet model is used to obtain the correlation between the facial image corresponding to the ROI region and all facial data in the facial database. Based on the numerical value of the degree of correlation, the results of personnel appearance recognition are obtained.

10. The intelligent facial recognition method for smart security cameras as described in claim 9, characterized in that, The obtained personnel appearance recognition results are specifically as follows: If the facial database contains data with a correlation greater than a preset similarity threshold with the corresponding facial image, the identity information corresponding to the image with the highest correlation is selected as the recognition result; otherwise, the recognition fails.

Images