Eye location method and system based on gray projection and multi-angle combined HOG features

Abstract

The application belongs to the technical field of computers and discloses a human eye positioning method and system based on gray scale projection and multi-angle combined HOG features. Firstly, in the image collected by a near-infrared camera, a local adaptive binarization and positioning image centroid combined mode is used to obtain a face region in the image. Then, a gray scale integral projection and least square fitting combined method is used to preliminarily determine all candidate points that may be human eyes. Then, through a gray scale co-occurrence matrix and the maximum connected domain area, most non-human eye candidate points are removed from the preliminarily obtained candidate points, and a gradient image is used to adjust the positions of the candidate points. Finally, through a multi-angle combined HOG feature extraction method, a feature vector of a candidate region centered on the candidate points is extracted, and the feature vector is input into multiple trained classifiers for classification and discrimination. The application has high accuracy and low operation time in the task of positioning human eyes from a face image.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a method and system for human eye localization based on grayscale projection and multi-angle combined HOG features. Background Technology

[0002] As one of the most important visual organs in the human face, the human eye can reflect an individual's physiological state, gaze direction, and behavioral intentions. Its features have significant application value in fields such as fatigue detection, gaze tracking, iris recognition, and human-computer interaction. Therefore, accurate eye localization is the prerequisite and foundation for subsequent feature extraction and analysis. In the field of computer vision, face detection and eye localization are usually presented as hierarchical tasks, that is, the face region is detected first, and then key parts such as the eyes are further precisely localized ([Wikipedia][1]). However, due to the similarity of facial structures and their susceptibility to changes in posture, occlusion, and lighting, eye localization remains a challenging research problem.

[0003] In existing technologies, traditional human eye localization methods mainly rely on the analysis of image grayscale information. Typical methods include grayscale integral projection function method, variance projection function method, and mixed projection function method. These methods typically determine the location of the human eye region by projecting grayscale onto the image horizontally or vertically and utilizing the extreme value features of the projection curve. Because these methods are computationally simple and have low complexity, they can achieve relatively fast coarse human eye localization under ideal conditions of uniform illumination and high image quality. However, in real-world applications, images often suffer from uneven illumination, shadow interference, and blurry images, leading to distortion of the grayscale distribution. This causes the extreme value features of the projection function to become unstable, significantly reducing localization accuracy. Furthermore, methods based on simple grayscale features have poor robustness to noise and complex backgrounds, easily resulting in false positives or false negatives.

[0004] With the development of computer vision and artificial intelligence technologies, human eye localization methods based on machine learning and deep learning have gradually become a research hotspot. These methods typically construct classifiers or deep neural network models and train them on a large number of human and non-human eye samples, thereby achieving high-precision localization results in complex environments. For example, automatically extracting multi-level features through convolutional neural networks can effectively improve the adaptability to changes in lighting, pose, and occlusion. However, these methods usually rely on large-scale data training and involve complex model calculations, resulting in high computational costs and significant hardware resource requirements, thus limiting their application in embedded devices or real-time scenarios.

[0005] In summary, existing technologies still have significant shortcomings in practical applications: on the one hand, while deep learning or machine learning-based eye localization methods have high detection accuracy, their computational complexity is high, making it difficult to meet the requirements of real-time or low-power applications; on the other hand, although traditional grayscale projection-based eye localization methods are computationally efficient, they have poor robustness in complex lighting and noise environments, making it difficult to guarantee stable localization results. Therefore, how to improve the robustness and accuracy of eye localization algorithms in complex environments while maintaining low computational complexity has become an urgent technical problem to be solved in this field. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a method and system for human eye localization based on grayscale projection and multi-angle combined HOG features, which can be applied to scenarios where it is necessary to locate the human eye in images acquired by a near-infrared camera.

[0007] This invention includes the following steps:

[0008] S1: Considering that the images captured by the near-infrared camera contain background parts that interfere with subsequent localization, it is necessary to extract the face region from the image to minimize the proportion of the background in the image. Since the grayscale information of the background part in the images captured by the near-infrared camera is relatively uniform, a method combining local adaptive binarization and localization of the image centroid can effectively obtain the face region in the image;

[0009] S2: A method combining grayscale integral projection and least squares fitting is used to roughly screen all candidate points in a face image that could possibly be human eyes. This method can select all candidate points for human eyes on a face, but may include some non-human eye candidate points, such as eyebrows and noses;

[0010] S3: To remove non-human eye candidate points and position the human eye candidate points closer to the center of the human eye, the candidate points obtained in S2 need to be filtered and adjusted. This includes three steps: filtering candidate points based on the gray-level co-occurrence matrix, filtering candidate points based on the maximum connected component, and adjusting the position of the human eye candidate points;

[0011] S4: Extract candidate regions centered on the candidate points selected in S3. Use the multi-angle combined HOG feature extraction method proposed in this invention to extract features from the candidate regions, and then pass these features through multiple trained classifiers (i.e., multiple trained logistic regression models) to obtain the final discrimination result.

[0012] Furthermore, step S1 specifically includes:

[0013] Since the background region in the image captured by the near-infrared camera has relatively uniform grayscale values, the scaled image is first converted into a binarized image through local adaptive thresholding to reduce the interference of the background region on face localization. Then, the grayscale centroid (i.e., the first moment of the image) of the binarized image is used as the center of the face region, and the height and width of the face region are determined based on the second moment of the binarized image.

[0014] Furthermore, step S2 specifically includes:

[0015] In the face image obtained in S1, candidate points are searched pixel by pixel. If the gray value of the current pixel is less than a certain threshold, then the candidate point will be centered on the current pixel. The image block is cropped out. Then, in order to reduce the interference of uneven lighting on the left and right sides of the image block, the gray value of the image block is averaged. That is, with the center line of the image block as the axis, the gray value of each pixel in the image block is the average of the gray value of the current pixel and the gray value of the symmetrical position of the current pixel.

[0016] Calculate the horizontal and vertical gray-level integral projection sequences for the image patch. For both gray-level integral projections, perform the following calculation steps: Since the gray-level distribution in the eye region has a characteristic of "large at both ends and small in the middle," its projection sequence should show a trend of first decreasing and then increasing. Calculate the minimum value, the maximum value of the left and right halves of the gray-level integral projection sequence, and their corresponding positions in the sequence. If the minimum value and its position in the sequence do not meet certain conditions, the pixel is considered not a candidate point. Perform least-squares fitting on the sequence between the positions corresponding to the maximum value in the left half and the positions corresponding to the maximum value in the right half, and calculate the fitting error. If the error is greater than a certain threshold, the pixel is considered not a candidate point.

[0017] The above steps yield multiple candidate points. Considering the possibility of multiple candidate points being located close to each other, a simple clustering method is used to merge nearby candidate points into the same cluster, and then the candidate points in each cluster are merged into a single candidate point by taking the positional average.

[0018] Furthermore, step S3 specifically includes:

[0019] First, calculate the candidate region (assuming this candidate region is centered on the candidate point). The histogram of an image patch is used to determine the dynamic range of image gray levels. After obtaining the effective dynamic range of image gray levels, the gray-level co-occurrence matrix of the image patch is calculated using uniform quantization. Finally, the information entropy of this gray-level co-occurrence matrix is ​​used as an evaluation metric. If this metric is less than a certain threshold, the candidate point is discarded to remove falsely detected candidate points such as skin and background. This method can also be used as part of image quality assessment to remove blurred candidate points.

[0020] Then, the maximum connected region area of ​​the image was used as a filtering criterion to remove falsely detected eyebrow and nose regions. The dynamic range of gray levels in the candidate regions was statistically analyzed to determine the minimum gray level. and maximum grayscale value Then, the grayscale range is divided into four equal segments, each segment having a length of [length missing]. Select all gray values ​​less than in the candidate region. The point is taken as the starting point of the search. For all search starting points (let the gray value of the starting point be...),... The area of ​​its connected components is calculated recursively, where the adjacency condition is that the gray level is less than 1. It is also connected to the current position 8; among the connected component areas corresponding to all search starting points, the largest connected component area is selected. If this area is less than a set threshold, the candidate region is considered not to be the human eye.

[0021] Finally, the positions of the candidate points need to be adjusted to be closer to the center of the human eye. The candidate region is calculated (assuming this candidate region is centered on the candidate point). The Sobel gradient image of the image patch is calculated. In the upper half of the gradient image, the position corresponding to the minimum value of its horizontal gray-level integral projection is calculated, and the area of ​​the gradient image above this position is masked to reduce the interference of the eyebrows. The centroid of the image after masking the eyebrows is calculated, and the position of the candidate point is updated according to the deviation between the centroid and the image center. The above steps are iterated until the position change of the candidate point is less than a certain value or the maximum number of iterations is reached.

[0022] Furthermore, step S4 specifically includes:

[0023] Extracting from the candidate points selected by S3 as the center The candidate region. Establish a [something] within this candidate region. A sliding window, centered on the candidate points selected by S3. The image patch serves as the starting window (denoted as SW), and along a certain angle... The window moves towards the edge of the candidate region with a fixed step size. During the movement, the window traverses a total of 5 regions. The image patches are processed. The HOG features of the image patches traversed by the window during its movement are extracted (excluding the initial window SW), and these features are concatenated into a one-dimensional HOG feature vector, denoted as the angle. The corresponding HOG feature vector.

[0024] Using the same method, extract another angle. The corresponding HOG feature vector. Centered on the candidate region. The HOG feature vector corresponding to the image patch is the center of symmetry. The corresponding HOG feature vector and The corresponding HOG feature vectors are concatenated into a one-dimensional long HOG feature vector, and then the L2 norm normalization operation is performed on this feature vector to obtain the angle. With angle The corresponding combined HOG feature vector.

[0025] The final discrimination result is obtained by passing multiple combined HOG feature vectors through multiple pre-trained classifiers.

[0026] The purpose of this invention is to provide a human eye localization method based on grayscale projection and multi-angle combined HOG features, including cropping the face region, determining candidate points, candidate point screening and adjustment, and human eye discrimination based on multi-angle combined HOG features;

[0027] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, the computer program being executed by the processor causing the processor to perform the steps of the human eye localization method based on grayscale projection and multi-angle combined HOG features.

[0028] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the human eye localization method based on grayscale projection and multi-angle combined HOG features.

[0029] Another objective of this invention is to provide an information data processing terminal for implementing the aforementioned human eye positioning system based on grayscale projection and multi-angle combined HOG features.

[0030] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0031] First, this invention proposes a human eye localization method based on grayscale projection and multi-angle combined HOG features, used for detecting and locating human eyes in images. Combining the advantages of traditional human eye localization algorithms and machine learning algorithms, it can provide an effective and fast human eye localization method for systems or devices that require the use of human eye features.

[0032] Second, the human eye localization method proposed in this invention, based on grayscale projection and multi-angle combined HOG features, aims to address the shortcomings of traditional human eye localization algorithms, such as weak robustness, and the time-consuming nature of machine learning and deep learning-based human eye localization algorithms. The technological advancements of the human eye localization method based on grayscale projection and multi-angle combined HOG features are mainly reflected in the following aspects:

[0033] Fast execution time and strong robustness: Traditional human eye localization algorithms are simple in steps, but their accuracy in practical applications is not ideal. Conventional machine learning and deep learning methods, while achieving high accuracy, usually suffer from high computational costs. The method proposed in this invention combines the advantages of traditional algorithms and machine learning algorithms, avoiding the problem of excessively long program execution time caused by extracting features pixel by pixel from the original image for discrimination and classification, and achieving high accuracy in practical applications.

[0034] Third, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0035] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0036] This invention effectively reduces the dependence of high-precision human eye positioning technology on high-computing-power platforms, reduces algorithm deployment costs, and improves the response speed and recognition accuracy of human eye positioning-related products in practical applications. Simultaneously, this technical solution possesses strong versatility and portability, allowing for flexible adaptation to different hardware platforms and application scenarios. It can be applied to the field of iris recognition and has high application value.

[0037] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0038] Compared to existing technologies where traditional methods suffer from insufficient accuracy, and machine learning and deep learning methods are heavily reliant on computing power and storage resources, making it difficult to simultaneously achieve both accuracy and efficiency, this invention provides a technical implementation for human eye positioning that balances real-time performance, accuracy, and deployability. It offers significant advantages in the lightweight and practical application of high-performance human eye positioning functions. Therefore, this invention, to a certain extent, overcomes the shortcomings of existing human eye positioning technologies in simultaneously achieving both performance and efficiency, and has positive implications for the development of related technologies. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a structural diagram of the human eye localization system based on grayscale projection and multi-angle combined HOG features provided in an embodiment of the present invention;

[0041] Figure 2 This is a flowchart of the human eye localization method based on grayscale projection and multi-angle combined HOG features provided in the embodiments of the present invention;

[0042] Figure 3 This is an example diagram of the multi-angle combined HOG feature extraction method provided in the embodiments of the present invention;

[0043] Figure 4 These are the four Sobel gradient operators used in the multi-angle combined HOG feature extraction method provided in this embodiment of the invention;

[0044] Figure 5 This is relevant evidence of the technical effects obtained by the embodiments of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0046] To address the problems existing in the prior art, the present invention provides a human eye localization method and system based on grayscale projection and multi-angle combined HOG features. The present invention will be described in detail below with reference to the accompanying drawings.

[0047] The following are specific examples demonstrating how to apply a human eye localization method based on grayscale projection and multi-angle combined HOG features.

[0048] Example: Iris Recognition System

[0049] In a practical iris recognition system, the iris recognition system must first locate and obtain the image of the human eye from the image captured by the near-infrared camera.

[0050] 1. Extracting the face region: The face region in the image is obtained by using local adaptive binarization and calculating the first moment of the image;

[0051] 2. Determine candidate points: Use gray-scale integral projection and least squares fitting to initially determine candidate points for the human eye;

[0052] 3. Preliminary screening of candidate points: Candidate points for the human eye are screened using the gray-level co-occurrence matrix and the maximum connected component of the image;

[0053] 4. Candidate point position adjustment: The position of the candidate points is iteratively adjusted based on the centroid of the image gradient;

[0054] 5. Human eye discrimination based on multi-angle combined HOG features: The HOG features of the candidate region are extracted using a multi-angle combined HOG feature extraction algorithm. The features are then further filtered out by multiple trained logistic regression models to identify human eye candidate points.

[0055] This embodiment demonstrates the applicability of a human eye localization method based on grayscale projection and multi-angle combined HOG features in iris recognition applications. This method can effectively improve the accuracy and security of iris recognition. By combining the advantages of traditional algorithms and machine learning algorithms, it effectively reduces computation time while improving accuracy.

[0056] See Figure 1 The human eye localization system based on grayscale projection and multi-angle combined HOG features of this invention is structurally designed as follows: First, in the image acquired by the near-infrared camera, the face region in the image is obtained by using local adaptive binarization and calculating the first moment of the image; then, the face region is initially determined by using a combination of grayscale integral projection and least squares fitting; next, most non-eye candidate points are removed from the initially obtained candidate points using the grayscale co-occurrence matrix and the maximum connected region area, and the position of the candidate points is adjusted using gradient images; finally, the multi-angle combined HOG feature extraction method is used to extract the features of the image patch centered on the candidate points, and these features are input into multiple pre-trained classifiers for classification and discrimination.

[0057] The human eye localization method based on grayscale projection and multi-angle combined HOG features involves multiple steps. The following is a detailed description of each step of the method:

[0058] Step S1: Cropping the face region

[0059] By employing local adaptive binarization and calculating the first moment of the image, the face region in the image is obtained, effectively removing redundant background parts from the original image.

[0060] Step S2: Determine candidate points

[0061] A method combining grayscale integral projection and least squares fitting was used to initially screen candidate points for human eyes. This method can select all candidate points for human eyes, but it also selects some non-eye candidate points.

[0062] Step S3: Filter candidate points

[0063] By utilizing the gray-level co-occurrence matrix and the maximum connected region area of ​​the image, most non-eye candidate points, including nose, eyebrows, and skin, can be effectively removed from all candidate points obtained in S2.

[0064] Step S4: Adjusting the position of candidate points

[0065] By utilizing the Sobel gradient image and horizontal grayscale integral projection, the interference of eyebrows on the candidate region of the human eye is masked. Then, the positions of the candidate points are updated by iteratively finding the centroid of the gradient image.

[0066] Step S5: Human eye discrimination based on multi-angle combined HOG features

[0067] HOG features of candidate regions are extracted using a multi-angle combined HOG feature extraction method. These features are then passed through multiple trained logistic regression models to identify candidate eye points. Finally, the eye image is cropped based on these candidate eye points for subsequent processing steps.

[0068] like Figure 2 As shown, the human eye localization method based on grayscale projection and multi-angle combined HOG features provided in this embodiment of the invention includes the following steps:

[0069] S1. Considering the presence of background areas in the images captured by the near-infrared camera that can interfere with subsequent localization, it is necessary to extract the face region from the image to minimize the proportion of the background. Since the grayscale information of the background in the near-infrared camera images is relatively uniform, a method combining local adaptive binarization and image centroid localization can effectively extract the face region from the image.

[0070] Step S1 specifically involves performing local adaptive binarization on the image to obtain a binarized image. The center of the face region... Select in binary image The gray centroid, i.e.

[0071]

[0072] in, and They represent The height and width. Then the height of the face region. Hekuan It can be determined based on the second moment of the binarized image, i.e.

[0073]

[0074] S2. A method combining grayscale integral projection and least squares fitting is used to roughly screen all candidate points in a face image that could possibly be human eyes. This method can select all candidate points for human eyes on a face, but may include some non-human eye candidate points, such as eyebrows and noses.

[0075] Step S2 specifically involves traversing the face image obtained in S1 pixel by pixel. If the grayscale value of the current pixel is less than a certain threshold, then a crop centered on the current pixel is extracted. Image blocks Then, an averaging operation is performed on the image, that is, for... In the middle position The pixel value is the sum of the pixel value at the current position and the pixel value at the location located therein. The average of the sum of pixel values ​​is calculated as follows:

[0076]

[0077] in, The result is rounded down. Then, the horizontal and vertical grayscale integral projection sequences of the image patch are calculated:

[0078]

[0079] in, and These represent the current image blocks. Row indexes and column indexes.

[0080] In the calculated grayscale integral projection sequence, the minimum value, the maximum values ​​of the left and right halves, and their corresponding positions in the sequence are determined. If the minimum value and its position in the sequence do not meet certain conditions, the pixel is considered not a candidate point. A least-squares fit is performed between the positions corresponding to the maximum values ​​of the left and right halves of the sequence, and the fitting error is calculated. If the error exceeds a certain threshold, the pixel is considered not a candidate point.

[0081] Considering the possibility of multiple candidate points being located close to each other, a simple clustering method is used to merge nearby candidate points into the same cluster, and the candidate points in each cluster are merged into one candidate point by taking the position average method.

[0082] S3. To remove non-human eye candidate points and position the human eye candidate points closer to the center of the human eye, the candidate points obtained in S2 need to be filtered and adjusted. This includes three steps: filtering candidate points based on the gray-level co-occurrence matrix, filtering candidate points based on the maximum connected component, and adjusting the position of the human eye candidate points.

[0083] Step S3 specifically involves first calculating the candidate region (assuming this candidate region is centered on the candidate point). The histogram of an image patch is used to determine the dynamic range of image gray levels. After obtaining the effective dynamic range of image gray levels, the gray-level co-occurrence matrix of the image patch is calculated using uniform quantization. Finally, the information entropy of this gray-level co-occurrence matrix is ​​used as an evaluation metric. If this metric is less than a certain threshold, the candidate point is discarded to remove falsely detected candidate points such as skin and background. This method can also be used as part of image quality assessment to remove blurred candidate points.

[0084] The maximum connected component area of ​​the image is used to further filter and remove falsely detected eyebrow and nose regions. The dynamic range of gray levels in the candidate regions is statistically analyzed to determine the minimum gray level. and maximum grayscale value Then, the grayscale range is divided into four equal segments, each segment having a length of [length missing]. Select all gray values ​​less than in the candidate region. The point is taken as the starting point of the search. For all search starting points (let the gray value of the starting point be...),... The area of ​​its connected components is calculated recursively, where the adjacency condition is that the gray level is less than 1. It is also connected to the current position 8; among the connected component areas corresponding to all search starting points, the largest connected component area is selected. If this area is less than a set threshold, the candidate region is considered not to be the human eye.

[0085] Finally, the positions of the candidate points need to be adjusted to be closer to the center of the human eye. The candidate region (centered on the candidate point) is then calculated. Sobel gradient image of image patch ,Right now

[0086]

[0087] in Indicates the image location index; and These are the Sobel gradients of the image in the horizontal and vertical directions, respectively. Find the gradient image. The location corresponding to the minimum value of the horizontal gray-level integral projection in the upper half of the image. The portion above this location is masked in the image to reduce eyebrow interference; based on the gradient image after masking the eyebrows... The position of the candidate point is updated by the deviation between the centroid and the image center, and the position of the candidate point is updated iteratively.

[0088] S4. Using the candidate points selected in S3 as the center, extract the candidate region. Extract the features of the candidate region using the multi-angle combined HOG feature extraction method proposed in this invention, and then pass these features through multiple trained classifiers (i.e., multiple trained logistic regression models) to obtain the final discrimination result:

[0089] Step S4 specifically involves truncating the candidate points selected in S3 as the center. Candidate regions. Establish a... A sliding window, centered on the candidate points selected by S3. The image patch serves as the starting window (denoted as SW), and along a certain angle... The window moves towards the edge of the candidate region with a fixed step size. During the movement, the window traverses a total of 5 regions. Image blocks.

[0090] Extract the HOG features of the four image patches traversed by the sliding window (this step does not include the initial window SW). Specifically, the steps involve using... Figure 4 The four Sobel gradient operators shown (including Sobel gradient operators at 0°, 45°, 90°, and 135°) are used to calculate the gradient images of the image patch in the four directions (denoted as ). ). Amplitude of the gradient image and angle It can be determined by the following formula:

[0091]

[0092]

[0093] in, Indicates the position index; This indicates that the maximum value is taken; then, the Sobel gradient image is traversed pixel by pixel, and the HOG feature vector is calculated according to the following formula. :

[0094]

[0095] Where % represents the modulo operation; express The length is usually set to 9.

[0096] After the above steps, the sliding window can be obtained along the angle. The HOG feature vectors of the four image patches traversed during the movement. These four feature vectors are concatenated into a one-dimensional feature vector of length 36, denoted as the angle. The corresponding HOG feature vector.

[0097] Extract the angle following the same steps. The corresponding HOG feature vector and the HOG feature vector corresponding to the starting window SW. The corresponding HOG feature vector and The corresponding HOG feature vectors are connected into a one-dimensional feature vector of length 81, with the HOG feature vector corresponding to the initial window SW as the center of symmetry. This vector is then normalized using the L2 norm to obtain the angle. With angle The corresponding combined HOG feature vector.

[0098] Finally, the combined HOG feature vectors are passed through multiple pre-trained classifiers to obtain the final discrimination result.

[0099] An application embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of a human eye localization method based on grayscale projection and multi-angle combined HOG features.

[0100] An application embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform steps of a human eye localization method based on grayscale projection and multi-angle combined HOG features.

[0101] An application embodiment of the present invention provides an information data processing terminal, which is used to implement a human eye positioning system based on grayscale projection and multi-angle combined HOG features.

[0102] This invention can be implemented in various ways, including hardware, software, and a combination of both. Hardware implementation may employ dedicated logic circuits, while software implementation may be implemented by a microprocessor or dedicated hardware through code stored on a storage medium. Those skilled in the art will recognize that the technology can be implemented by computer-executed instructions or processor control code, which may be stored on various forms of media (such as disks, CD / DVD-ROMs), programmable memories (such as ROM firmware), or provided via data carriers (such as optical or electronic signals). Implementation of this invention may involve the use of hardware circuits, such as very large-scale integrated circuits, gate arrays, logic chips, transistors, etc., or programmable hardware devices, such as field-programmable gate arrays, programmable logic devices, etc., or may employ software executed by different types of processors, or a combination of both, such as through firmware implementation.

[0103] The dataset used to train the logistic regression classifier contains 22,100 eye images (including left and right eyes) and 32,925 non-eye images (including nose, mouth, eyebrows, and other non-eye images, in a ratio of approximately 9:9:1), totaling 55,025 images.

[0104] The HOG feature extraction process first involves setting a feature at the center of the image. A sliding window is used to extract the HOG feature vectors of the image patches contained within the sliding window, denoted as C. The window is then moved at different angles from the image center, and the HOG feature vectors of the image patches traversed by the window are extracted with a fixed step size (a total of 4 HOG feature vectors are extracted) until the window encounters the image boundary. The 4 extracted HOG feature vectors are then concatenated into a one-dimensional vector. Next, the HOG feature vectors extracted from any two different directions are symmetrically concatenated around C to obtain a long feature vector (this long feature vector should contain 9 HOG feature vectors). This vector is then normalized to the L2 norm and used as the input to the logistic regression model. Figure 3 This is a schematic diagram of the feature extraction algorithm proposed in this invention.

[0105] During training, a logistic regression model was used, with the test set comprising 10% of the total dataset, and a maximum convergence count of 500. Following the standards and processing procedures described above, nine pairs of feature vectors were selected to train nine logistic regression models (classifiers) at angles of 0-180°, 90-270°, 0-90°, 15-195°, 165-345°, 30-330°, 150-210°, 30-150°, and 210-330°. The prediction results were then aggregated and evaluated. During testing, the nine classifiers were used jointly on the test set. The accuracy rate for eye-based discrimination reached 98.86%, the accuracy rate for non-eye-based discrimination reached 100%, and the overall accuracy rate was 99.51%. The average processing time per image was approximately 0.024 seconds.

[0106] In summary, compared with existing related methods, the method proposed in this invention has considerable accuracy and low time complexity, and is more suitable for practical applications that require human eye positioning.

[0107] I. Specific application areas or related products of this invention.

[0108] An embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the human eye localization method based on grayscale projection and multi-angle combined HOG features.

[0109] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the human eye localization method based on grayscale projection and multi-angle combined HOG features.

[0110] An embodiment of the present invention provides an information data processing terminal, which is used to implement the aforementioned human eye positioning system based on grayscale projection and multi-angle combined HOG features.

[0111] II. Evidence related to the technical effects obtained by the embodiments of the present invention

[0112] The eye localization algorithm proposed in this invention has been deployed in an iris recognition module. This module communicates with the host computer software via a serial port protocol, allowing users to preview images captured by the module in real time on the host computer software interface. In the preview image, a prompt box that follows the real-time movement of the eyes can be observed in the face area, indicating that the module can accurately and in real-time locate the eye area based on the proposed eye localization algorithm. For specific results, please refer to [link / reference]. Figure 5 .

[0113] Implementation Method 1: A near-infrared camera is used to acquire a frontal facial image of a person. The original image is first subjected to local adaptive binarization. Then, the center position of the face is determined based on the gray-level centroid of the binary image, and the size of the cropped area is determined using the second-order moment to obtain the face region. This method demonstrates that facial cropping based on the imaging characteristic of a uniform background gray level provides stable input for subsequent candidate point search. Implementation Method 2: Low-gray-level pixels are searched pixel-by-pixel within the cropped face region. Local image blocks are cropped at positions that meet the gray-level conditions. Horizontal and vertical gray-level integral projection sequences are calculated separately, and candidate points that satisfy the distribution pattern of large pixels on both sides and small pixels in the middle are selected. Positions that do not conform to the trend of eye structure changes are then eliminated using least-squares fitting error. This method shows that gray-level projection and fitting discrimination are not used independently but jointly define the gray-level distribution pattern of candidate points. Implementation Method 3: Adjacent candidate points are aggregated, and the average position within the same cluster is taken as the new candidate center, avoiding multiple duplicate candidate points for the same eye. This method demonstrates that the present invention not only focuses on detecting the eyes but also further controls the concentration of candidate point distribution, improving subsequent screening efficiency. Implementation Method Four: Perform grayscale dynamic range statistics on the candidate regions, construct a grayscale co-occurrence matrix, and calculate information entropy to remove textureless and blurry regions. This method demonstrates that the present invention utilizes texture statistical features to exclude skin and flat backgrounds, ensuring that candidate points retain eye regions with stable local texture structures. Implementation Method Five: In the candidate regions after texture filtering, recursively search connected components starting from low grayscale points, calculate the area of ​​the largest connected component, and remove non-target regions such as eyebrows and nose shadows. This method demonstrates the seamless integration of the grayscale co-occurrence matrix and the area of ​​the largest connected component; the former addresses false detections due to lack of texture, while the latter addresses false detections due to structural morphology. The combination of the two significantly improves candidate authenticity. Implementation Method Six: Extract candidate regions centered on the retained candidate points, establish sliding windows moving in different directions, extract multi-angle combined directional gradient histogram features, and input them into multiple logistic regression models for discrimination. Before this, use Sobel gradient images for upper interference masking, and iteratively correct the candidate point positions based on centroid offset to make the final discrimination center closer to the real pupil region. This approach demonstrates that candidate point correction and multi-angle feature classification are coupled by a mechanism. First, spatial alignment accuracy is improved through gradient centroid iteration. Then, the ability to express eye-oriented texture is enhanced through multi-angle combined features. Finally, robust recognition is achieved by leveraging multi-model outputs.

[0114] It should be noted that embodiments of the present invention can be implemented using hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented using hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or using software executed by various types of processors, or using a combination of the above-described hardware circuitry and software, such as firmware.

[0115] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for extracting and correcting candidate points for the human eye based on grayscale projection, characterized in that, Includes the following steps: A near-infrared face image is acquired, and the near-infrared face image is subjected to local adaptive binarization processing. The face region is determined based on the gray centroid and second moment of the binarization result. Candidate points are searched pixel by pixel within the face area. Local image blocks are constructed for pixels that meet the preset gray level conditions. The horizontal gray level integral projection sequence and the vertical gray level integral projection sequence of the local image blocks are calculated respectively. Initial candidate points are selected based on the extreme value distribution characteristics of the projection sequence and the least squares fitting error. Initial candidate points that are adjacent in location are aggregated to obtain cluster candidate points; For each cluster candidate point, the first round of screening is carried out based on the information entropy of the gray-level co-occurrence matrix of the candidate region, the second round of screening is carried out based on the area of ​​the maximum connected region of the candidate region, and the position is adjusted iteratively based on the centroid offset of the gradient image of the candidate region to obtain the corrected human eye candidate points.

2. The method for extracting and correcting candidate points of the human eye as described in claim 1, characterized in that, The process of determining the face region based on the gray-level centroid and second-order moment of the binarization result is as follows: The gray-level centroid of the binarized image is used as the center of the face region. The second-order moment representation result of the binarized image is used as the basis for determining the height and width parameters of the face region. The face region is then cropped according to the height and width parameters.

3. The method for extracting and correcting candidate points of the human eye as described in claim 1, characterized in that, The initial candidate points are selected based on the extreme value distribution characteristics of the projection sequence and the least squares fitting error, specifically as follows: Find the minimum value, the maximum value of the left half, and the maximum value of the right half of the horizontal gray-scale integral projection sequence and the vertical gray-scale integral projection sequence respectively, and their corresponding positions. Determine whether the position of the minimum value is located between the positions of the two maximum values, and determine whether the maximum value and the minimum value on both sides meet the preset difference condition. Between the position corresponding to the maximum value in the left half and the position corresponding to the maximum value in the right half, least squares fitting is performed on the projection sequence, and the current pixel is determined as an initial candidate point based on whether the fitting error is less than a preset threshold.

4. The method for extracting and correcting candidate points of the human eye as described in claim 1, characterized in that, The position adjustment based on the centroid offset of the candidate region gradient image is specifically as follows: The candidate region is extracted centered on the candidate point, and the Sobel gradient image of the candidate region is calculated. Calculate the minimum position of the horizontal integral projection in the upper half of the gradient image, and then mask the area above that position. Calculate the centroid of the masked gradient image, and update the candidate point positions based on the deviation between the centroid and the candidate region center; Repeat the masking, centroid calculation, and position update process until the change in the candidate point position is less than the preset value, or the maximum number of iterations is reached.

5. A human eye discrimination method based on multi-angle combined directional gradient histogram features, characterized in that, Includes the following steps: Obtain candidate points to be judged, and extract candidate regions centered on the candidate points; A sliding window is established within the candidate region, with the central image block where the candidate point is located as the starting window. The sliding window is controlled to move sequentially along at least two different directions, and the orientation gradient histogram features of multiple image blocks are collected during the movement in each direction. The directional gradient histogram features obtained from different directions are symmetrically combined to form multiple sets of combined feature vectors, and the L2 norm normalization is performed on each combined feature vector. The normalized combination feature vectors are input into multiple pre-trained logistic regression classifiers, and the candidate point is determined to be the center position of the human eye based on the output of the multiple logistic regression classifiers.

6. The human eye discrimination method as described in claim 5, characterized in that, The control sliding window moves sequentially along at least two different directions, and during the movement in each direction, it acquires the directional gradient histogram features of multiple image patches, specifically: Using the central image patch where the starting window is located as a reference, move along the first direction by a fixed step size and traverse 4 image patches. Extract the directional gradient histogram features of the traversed 4 image patches and concatenate them into a first direction feature vector. Move a fixed step size along the second direction corresponding to the first direction and traverse 4 image patches. Extract the orientation gradient histogram features of the traversed 4 image patches and concatenate them into a feature vector of the second direction. Using the directional gradient histogram features of the central image block as the center of symmetry, the first-direction feature vector and the second-direction feature vector are concatenated to form a combined feature vector.

7. The human eye discrimination method as described in claim 5, characterized in that, The process of determining whether a candidate point is at the center of the human eye based on the output results of multiple logistic regression classifiers is as follows: Set up logistic regression classifiers for different combinations of directions; The discriminant values ​​output by each logistic regression classifier are summarized, and the human eye category determination result of the candidate points is output based on the number of discriminant values, the size of the discriminant values, or the weighted discriminant result.

8. A human eye localization system based on grayscale projection and multi-angle combined directional gradient histogram features, characterized in that, include: The face region extraction module is used to perform local adaptive binarization processing on near-infrared face images and determine the face region based on the gray centroid and second moment of the binarization result. The candidate point coarse screening module is used to search for candidate points pixel by pixel within the face area and obtain initial candidate points based on the horizontal gray-scale integral projection sequence, the vertical gray-scale integral projection sequence, and the least squares fitting error. The candidate point correction module is used to filter and adjust the position of the initial candidate points based on the information entropy of the gray-level co-occurrence matrix, the area of ​​the maximum connected region, and the centroid shift of the gradient image, and output the corrected human eye candidate points. The feature extraction module is used to extract candidate regions centered on the corrected human eye candidate points and extract the combined directional gradient histogram features of the candidate regions along multiple directions. The classification and discrimination module is used to input the combined directional gradient histogram features into multiple pre-trained logistic regression classifiers and output the human eye localization result.

9. The human eye positioning system as described in claim 8, characterized in that, The candidate point correction module includes: The information entropy filtering unit is used to construct a gray-level co-occurrence matrix based on the effective dynamic range of gray levels in the candidate regions, and to filter out candidate points with insufficient texture based on the information entropy of the gray-level co-occurrence matrix. The connected component filtering unit is used to divide grayscale segments based on the minimum and maximum grayscale values ​​of candidate regions, calculate the area of ​​connected components using low grayscale pixels as the search starting point, and filter out non-human eye candidate points based on the maximum connected component area. The position adjustment unit is used to update the position of candidate points based on the centroid offset after the gradient image is masked.

10. The human eye positioning system as described in claim 8, characterized in that, The classification and discrimination module includes: The feature combination unit is used to concatenate the histogram features of the directional gradients obtained in different directions with the symmetry center of the central image block where the candidate point is located, forming multiple sets of combined feature vectors, and performing L2 normalization on each combined feature vector. The multi-classifier discrimination unit is used to input the normalized combination feature vectors into multiple logistic regression classifiers respectively, and to fuse the output results of multiple logistic regression classifiers to output the final human eye localization result.

Images