Iris recognition method based on graph neural network and anti-aliasing convolutional neural network
By combining the iris recognition method of graph neural network and anti-aliased convolutional neural network, the problem of complex sample collection and insufficient robustness of the iris recognition algorithm in drug users' detection is solved, and efficient and accurate iris recognition is achieved, suitable for drug control sites. and security checks.
Patent Information
- Application Number
- CN202411236582.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-04
- Publication Date
- 2025-01-07
AI Technical Summary
The existing drug users detection methods have problems such as complex sample collection process, inconvenience, high cost, and environmental impact, and the robustness and accuracy of the iris recognition algorithm in complex environments are insufficient.
Using an iris recognition method based on graph neural network and anti-aliased convolutional neural network, the graph structure data of iris images is processed through graph neural network, combined with anti-aliased convolutional neural network, reduce the serrated effect, extract highly accurate feature vectors, and pass K-nearest neighbor algorithm performs matching recognition.
It realizes efficient and accurate iris recognition in complex environments, reduces detection costs, and provides a contactless, fast and accurate biometric technology suitable for drug control sites and safety inspections.
Smart Images
Figure CN119274226A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of image processing and biometric identification technology, and in particular relates to an iris recognition method based on a graph neural network and an anti-aliasing convolutional neural network. Background Art
[0002] In the current field of drug addict detection, traditional biological detection methods are mainly relied on, including urine testing, saliva testing, blood testing, hair testing, etc. These methods have their own advantages and disadvantages, but all have limitations to varying degrees.
[0003] Although urine testing is widely used, its sample collection process requires a high degree of cooperation from the person being tested, and there is a risk of stealing and replacing samples. At the same time, the collection cycle is long and there are many variable factors in the intermediate process, which can easily affect the accuracy of the test results. Although saliva testing is more convenient to collect samples, it is easily interfered by various factors such as food, beverages, and individual physiological status, affecting the reliability of the test results. Although blood testing is more accurate, the testing process is complicated, the cost is high, and there are strict requirements on timeliness. Samples need to be collected in a short time and laboratory analysis is required. Hair testing has higher accuracy and a longer detection window, but it is not suitable for drug control sites, and it is costly and difficult to implement.
[0004] In view of the limitations of the above traditional detection methods, researchers are constantly exploring new detection technologies to improve the efficiency and accuracy of detection. As an emerging biometric recognition technology, iris recognition technology has attracted much attention due to its unique iris texture characteristics (uniqueness, stability, non-contact and anti-counterfeiting). However, the robustness and accuracy of current iris recognition algorithms in complex environments (such as lighting changes, eye occlusion, etc.) still need to be improved. Summary of the invention
[0005] In order to solve the above problems, the present invention provides an iris recognition method based on graph neural network and anti-aliasing convolutional neural network. By combining the advantages of graph neural network GNN in graph structure data processing and the powerful ability of CNN in image processing, accurate segmentation and feature extraction of iris images can be achieved, and complex relationships in graph structure data can be effectively learned and represented. The anti-aliasing CNN can reduce the aliasing effect while maintaining the clarity of image edges, thereby improving the accuracy of feature extraction.
[0006] The present invention aims to solve the various deficiencies in existing drug addict detection methods, provide a more efficient, accurate and low-cost iris recognition technology, and bring new breakthroughs in the field of drug addict detection.
[0007] The present invention specifically provides an iris recognition method based on a graph neural network and an anti-aliasing convolutional neural network, and the iris recognition method comprises the following steps:
[0008] S1, real-time acquisition of the target person's iris image, pre-processing of the acquired image, conversion to grayscale image, and image filtering;
[0009] S2. The present invention uses a graph neural network (GNN) to treat each pixel in the iris image as a node in the graph, and constructs an adaptive edge based on multifaceted knowledge to achieve effective feature aggregation and accurately segment the iris area.
[0010] S3, after segmentation, the iris image is normalized using Daugman's linear mapping method;
[0011] S4. Based on deep feature extraction of anti-aliasing convolutional neural network (CNN), feature vectors with texture information are extracted; the above feature vectors are encoded through an adaptive threshold coding method to generate binary iris codes with higher variability, and the K-nearest neighbor algorithm (KNN) is used for matching and recognition.
[0012] Furthermore, the image preprocessing in step S1, which converts the grayscale image, and the image filtering, are specifically as follows:
[0013] The basic formula for converting a color image into a grayscale image is as follows:
[0014] Gray=0.299×R+0.587×G+0.114×B
[0015] In the formula, R, G, and B are the brightness of the three color channels RGB in the color image; the image denoising process uses Gaussian filtering to eliminate noise, and its calculation formula includes:
[0016]
[0017] Where G(x, y) is the Gaussian kernel function and δ is the standard deviation of the Gaussian distribution.
[0018] Furthermore, in step S2, GNN is used to achieve accurate iris segmentation, and its principles include:
[0019] GNN regards individual pixels as nodes in a graph and builds adaptive edges based on multiple aspects of knowledge, including visual similarity, positional correlation, and semantic consistency for feature aggregation.
[0020] Specifically, we learn visual relations between nodes based on visual similarity and define these relations using a visual adjacency matrix (VAM), which measures visual similarity using cosine similarity in the embedding space:
[0021]
[0022] Among them, X i and X jRespectively represent V i and V j The eigenvector of Θ v is the linear projection matrix that maps the nodes into the embedding space. V is the VAM, V ij is the element in the i-th row and j-th column of V. The spatial position is another non-ignorable information for iris segmentation. Adjacent nodes are more likely to belong to the same eye component than distant nodes. A position adjacency matrix (PAM) gives the correlation of spatial information between nodes:
[0023]
[0024] Among them (h i , w i ) and (h i , w j ) is v i and v j The position vector. P is PAM, P ij Represents the element in the i-th row and j-th column of P. Similarly, semantic consistency maps nodes to semantic space and learns pseudo labels to define relationships based on the consistency of labels. The semantic adjacency matrix (SAM) follows semantic consistency to represent the relationship between nodes:
[0025]
[0026] in and Yes i and v j 's pseudo-labels.
[0027] VAM, PAM and SAM are integrated into two basic aggregation modules, namely encoder block and decoder block. Both blocks utilize propagation rules to aggregate nodes and update their features. The encoder block has rich visual and position features but less semantic information, and only utilizes VAM and PAM to maximize the exploration of visual similarity and position correlation. In the encoder block, the total adjacency matrix of the propagation process can be defined as:
[0028] A encoder =SoftMax(V+P)
[0029] Compared with the encoder block, the decoder block integrates all three adjacency matrices for feature aggregation. The total adjacency matrix of its propagation process is a combination of VAM, PAM and SAM, denoted as:
[0030] A decoder =SoftMax(V+P+S)
[0031] Furthermore, in step S3, the iris image is normalized using Daugman's linear mapping method after segmentation, and the calculation formula includes:
[0032] x=(rr i ) / (r0-r i )*W
[0033] y=θ / 2π*H
[0034] Where x and y are the normalized image coordinates, r is the distance from the origin to the iris, θ is the angle between the point and the center of the iris relative to the horizontal line, and W and H are the width and height of the normalized image.
[0035] Furthermore, in step S4, based on deep feature extraction of anti-aliasing convolutional neural network CNN, feature vectors with texture information are extracted. These feature vectors are encoded by adaptive threshold encoding method to generate binary iris codes with higher variability. K-nearest neighbor algorithm (KNN) is used for classification. The steps are as follows:
[0036] Specifically, the feature extraction process uses ResNet50 as the backbone network, in which each convolutional layer outputs a feature map tensor (FMT). These FMTs contain representations of the input image at different levels of abstraction:
[0037] F (l) =σ(K*F (l-1) +b)
[0038] Among them, K represents the convolution kernel, * represents the convolution operation, b is the bias term, σ is the activation function, and F (l) It is a FMT. In an iris recognition system, these FMTs are captured and stored for subsequent processing. Based on the properties of the iris image, a unique mask can be generated. This mask is used to remove irrelevant or noisy areas from the FMT, leaving only the parts that are useful for recognition.
[0039] The feature vector is encoded by an adaptive threshold coding method, which dynamically adjusts the threshold according to the local statistical characteristics of the FMT and converts each element in the FMT into a binary value:
[0040]
[0041] where f i is an element in FMT, and T is the threshold. The encoding process will focus on the area within the mask and use the K-nearest neighbor algorithm (KNN) to classify the results:
[0042]
[0043]
[0044] Among them, b test is the sample to be tested, b i is a standard sample, and the category is the test sample is classified as the category that appears most frequently among the K nearest neighbors.
[0045] Furthermore, the present invention also proposes an iris recognition system based on a graph neural network and an anti-aliasing convolutional neural network, comprising:
[0046] Iris acquisition module, used for real-time acquisition of iris images and image preprocessing;
[0047] The image segmentation module constructs adaptive edges based on various knowledge to perform feature aggregation and accurately segment the iris.
[0048] The iris classification module extracts the feature vector with texture information and uses the K-nearest neighbor algorithm (KNN) for classification through adaptive threshold coding.
[0049] Compared with the prior art, the beneficial effects are:
[0050] (1) By combining GNN and advanced anti-aliasing CNN technology in image processing, the present invention can more accurately identify subtle features in iris images and maintain high recognition accuracy even in complex environments (such as lighting changes, eye occlusion, etc.). GNN can capture the complex graph structure information in iris texture, while anti-aliasing CNN effectively reduces information loss and distortion in the image processing process, thereby significantly improving the accuracy and robustness of iris recognition;
[0051] (2) The detection method of the present invention is different from the detection methods such as urine and blood that require direct sample collection. Iris recognition technology is a non-contact biometric identification method. This means that during the detection process, the person being tested does not need to experience any discomfort or privacy infringement, which greatly improves the user experience. At the same time, non-contact detection also reduces the risk of cross-infection and is more in line with hygiene and safety requirements;
[0052] (3) Compared with traditional detection methods, such as blood testing, the present invention often requires complex sample collection and laboratory analysis processes, which are time-consuming and costly. However, the iris recognition technology based on GNN and anti-aliasing CNN can achieve rapid detection due to its efficient image processing capabilities and algorithm optimization. This not only shortens the detection cycle, but also reduces the detection cost, making the technology easier to promote and apply to various scenarios, including drug control sites, security inspections and other fields. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 A schematic diagram of a flow chart of an iris recognition method based on a graph neural network and an anti-aliasing convolutional neural network according to the present invention;
[0054] Figure 2 Schematic diagram of iris abnormality. DETAILED DESCRIPTION
[0055] The following is a detailed description of a specific implementation of an iris recognition method based on a graph neural network and an anti-aliasing convolutional neural network according to the present invention in conjunction with the accompanying drawings.
[0056] like Figure 1 As shown, the iris recognition method of the present invention comprises the following steps:
[0057] S1, real-time acquisition of the target person's iris image, pre-processing of the acquired image, conversion to grayscale image, and image filtering;
[0058] S2. The present invention uses a graph neural network (GNN) to treat each pixel in the iris image as a node in the graph, and constructs an adaptive edge based on multifaceted knowledge to achieve effective feature aggregation and accurately segment the iris area.
[0059] S3, after segmentation, the iris image is normalized using Daugman's linear mapping method;
[0060] S4. Based on deep feature extraction of anti-aliasing convolutional neural network (CNN), feature vectors with texture information are extracted; the above feature vectors are encoded through an adaptive threshold coding method to generate binary iris codes with higher variability, and the K-nearest neighbor algorithm (KNN) is used for matching and recognition.
[0061] The image preprocessing and grayscale image conversion and image filtering in step S1 are specifically as follows:
[0062] The basic formula for converting a color image into a grayscale image is as follows:
[0063] Gray=0.299×R+0.587×G+0.114×B
[0064] In the formula, R, G, and B are the brightness of the three color channels RGB in the color image; the image denoising process uses Gaussian filtering to eliminate noise, and its calculation formula includes:
[0065]
[0066] Where G(x, y) is the Gaussian kernel function and δ is the standard deviation of the Gaussian distribution.
[0067] In step S2, GNN is used to achieve accurate iris segmentation. The principles include:
[0068] GNN regards individual pixels as nodes in a graph and builds adaptive edges based on multiple aspects of knowledge, including visual similarity, positional correlation, and semantic consistency for feature aggregation.
[0069] Specifically, we learn visual relations between nodes based on visual similarity and define these relations using a visual adjacency matrix (VAM), which measures visual similarity using cosine similarity in the embedding space:
[0070]
[0071] Among them, X i and X j Respectively represent V i and V j The eigenvector of Θ v is the linear projection matrix that maps the nodes into the embedding space. V is the VAM, V ij is the element in the i-th row and j-th column of V. The spatial position is another non-ignorable information for iris segmentation. Adjacent nodes are more likely to belong to the same eye component than distant nodes. A position adjacency matrix (PAM) gives the correlation of spatial information between nodes:
[0072]
[0073] Among them (h i , w i ) and (h i , w j ) is v i and v j The position vector. P is PAM, P ij Represents the element in the i-th row and j-th column of P. Similarly, semantic consistency maps nodes to semantic space and learns pseudo labels to define relationships based on the consistency of labels. The semantic adjacency matrix (SAM) follows semantic consistency to represent the relationship between nodes:
[0074]
[0075] in and Yes i and v j 's pseudo-labels.
[0076] VAM, PAM and SAM are integrated into two basic aggregation modules, namely encoder block and decoder block. Both blocks utilize propagation rules to aggregate nodes and update their features. The encoder block has rich visual and position features but less semantic information, and only utilizes VAM and PAM to maximize the exploration of visual similarity and position correlation. In the encoder block, the total adjacency matrix of the propagation process can be defined as:
[0077] A encoder =SoftMax(V+P)
[0078] Compared with the encoder block, the decoder block integrates all three adjacency matrices for feature aggregation. The total adjacency matrix of its propagation process is a combination of VAM, PAM and SAM, denoted as:
[0079] A decoder =SoftMax(V+P+S)
[0080] In step S3, after segmentation, the iris image is normalized using Daugman's linear mapping method, and the calculation formula includes:
[0081] x=(rr i ) / (r0-r i )*W
[0082] y=θ / 2π*H
[0083] Where x and y are the normalized image coordinates, r is the distance from the origin to the iris, θ is the angle between the point and the center of the iris relative to the horizontal line, and W and H are the width and height of the normalized image.
[0084] In step S4, based on the deep feature extraction of anti-aliasing convolutional neural network CNN, feature vectors with texture information are extracted. These feature vectors are encoded by adaptive threshold encoding method to generate binary iris codes with higher variability. K-nearest neighbor algorithm (KNN) is used for classification. The steps are as follows:
[0085] Specifically, the feature extraction process uses ResNet50 as the backbone network, in which each convolutional layer outputs a feature map tensor (FMT). These FMTs contain representations of the input image at different levels of abstraction:
[0086] F (l) =σ(K*F (l-1) +b)
[0087] Among them, K represents the convolution kernel, * represents the convolution operation, b is the bias term, σ is the activation function, and F (l)It is a FMT. In an iris recognition system, these FMTs are captured and stored for subsequent processing. Based on the properties of the iris image, a unique mask can be generated. This mask is used to remove irrelevant or noisy areas from the FMT, leaving only the parts that are useful for recognition.
[0088] The feature vector is encoded by an adaptive threshold coding method, which dynamically adjusts the threshold according to the local statistical characteristics of the FMT and converts each element in the FMT into a binary value:
[0089]
[0090] where f i is an element in FMT, and T is the threshold. The encoding process will focus on the area within the mask and use the K-nearest neighbor algorithm (KNN) to classify the results:
[0091]
[0092]
[0093] Among them, b test is the sample to be tested, b i is a standard sample, and the category is the test sample is classified as the category that appears most frequently among the K nearest neighbors.
[0094] The iris recognition system based on graph neural network and anti-aliasing convolutional neural network of the present invention comprises:
[0095] Iris acquisition module, used for real-time acquisition of iris images and image preprocessing;
[0096] The image segmentation module constructs adaptive edges based on various knowledge to perform feature aggregation and accurately segment the iris.
[0097] The iris classification module extracts the feature vector with texture information and uses the K-nearest neighbor algorithm (KNN) for classification through adaptive threshold coding.
[0098] Figure 2 Schematic diagram of iris anomaly. The iris recognition method of the present invention can be used to perform anomaly recognition.
[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit the present invention. A person skilled in the art should understand that the specific implementation of the present invention can be modified or replaced by equivalents, but these modifications or changes are within the scope of protection of the pending claims.
Claims
1. An iris recognition method based on graph neural network and anti-aliasing convolutional neural network, characterized in that: The iris recognition method comprises the following steps: S1, real-time acquisition of the target person's iris image, pre-processing of the acquired image, conversion to grayscale image, and image filtering; S2. Use graph neural network (GNN) to treat each pixel in the iris image as a node in the graph, and build adaptive edges based on multi-faceted knowledge to achieve effective feature aggregation and accurately segment the iris area. S3, after segmentation, the iris image is normalized using Daugman's linear mapping method; S4. Deep feature extraction based on anti-aliasing convolutional neural network extracts feature vectors with texture information; the feature vectors are encoded through an adaptive threshold coding method to generate binary iris codes with higher variability, and the K-nearest neighbor algorithm KNN is used for matching and recognition.
2. According to claim 1, the iris recognition method based on graph neural network and anti-aliasing convolutional neural network is characterized in that: The image preprocessing and grayscale image conversion and image filtering in step S1 are specifically as follows: The basic formula for converting a color image into a grayscale image is as follows: Gray=0.299×R+0.587×G+0.114×B In the formula, R, G, and B are the brightness of the three color channels RGB in the color image; the image denoising process uses Gaussian filtering to eliminate noise, and its calculation formula includes: Where G(x, y) is the Gaussian kernel function and δ is the standard deviation of the Gaussian distribution.
3. The iris recognition method based on graph neural network and anti-aliasing convolutional neural network according to claim 2 is characterized in that: In step S2, GNN is used to achieve accurate iris segmentation. The principles include: GNN treats individual pixels as nodes in a graph and builds adaptive edges based on multiple aspects of knowledge, including visual similarity, positional correlation, and semantic consistency for feature aggregation; Specifically, we learn visual relations between nodes based on visual similarity and define these relations using a visual adjacency matrix (VAM), which measures visual similarity using cosine similarity in the embedding space: Among them, X i and X j Respectively represent V i and V j The characteristic vector of v is the linear projection matrix that maps nodes into the embedding space; V is VAM, V ij is the element in the i-th row and j-th column of V; the spatial position is another non-ignorable information for iris segmentation; adjacent nodes are more likely to belong to the same eye component than distant nodes; a position adjacency matrix PAM gives the correlation of spatial information between nodes: Among them (h i , w i ) and (h i , w j ) is v i and v j The position vector of ij Represents the elements in the i-th row and j-th column of P; similarly, semantic consistency maps nodes to the semantic space and learns pseudo labels to define relationships based on the consistency of labels. The semantic adjacency matrix SAM follows semantic consistency to represent the relationship between nodes: in and Yes i and v j Pseudo labels of VAM, PAM and SAM are integrated into two basic aggregation modules, namely the encoder block and the decoder block, both of which utilize the propagation rule to aggregate nodes and update their characteristics; the encoder block has rich visual and position features but less semantic information, and only utilizes VAM and PAM to maximize the exploration of visual similarity and position correlation; in the encoder block, the total adjacency matrix of the propagation process can be defined as: A encoder =SoftMax(V+P) Compared with the encoder block, the decoder block integrates all three adjacency matrices for feature aggregation, and its total adjacency matrix of the propagation process is a combination of VAM, PAM and SAM, denoted as: A decoder =SoftMax(V+P+S)。 4. The iris recognition method based on graph neural network and anti-aliasing convolutional neural network according to claim 3 is characterized in that: In step S3, after segmentation, the iris image is normalized using Daugman's linear mapping method, and the calculation formula includes: x=(r-r i ) / (r0-r i )*W y=θ / 2π*H Where x and y are the normalized image coordinates, r is the distance from the origin to the iris, θ is the angle between the point and the center of the iris relative to the horizontal line, and W and H are the width and height of the normalized image.
5. The iris recognition method based on graph neural network and anti-aliasing convolutional neural network according to claim 4 is characterized in that: In step S4, based on the deep feature extraction of the anti-aliasing convolutional neural network CNN, feature vectors with texture information are extracted, these feature vectors are encoded by the adaptive threshold encoding method to generate binary iris codes with higher variability, and the K-nearest neighbor algorithm KNN is used for classification, and the steps are as follows: The feature extraction process uses ResNet50 as the backbone network, in which each convolutional layer outputs a feature map tensor FMT, which contains the representation of the input image at different levels of abstraction: F (l) =σ(K*F (l-1) +b) Among them, K represents the convolution kernel, * represents the convolution operation, b is the bias term, σ is the activation function, and F (l) It is FMT; in the iris recognition system, FMT is captured and stored for subsequent processing; based on the properties of the iris image, a unique mask is generated, which is used to remove irrelevant or noisy areas from the FMT and only retain the parts that are useful for recognition; The feature vector is encoded by an adaptive threshold coding method, which dynamically adjusts the threshold according to the local statistical characteristics of the FMT and converts each element in the FMT into a binary value: where f i is an element in FMT, T is the threshold; the encoding process will focus on the area within the mask, and the result will be classified using the K-nearest neighbor algorithm KNN: Among them, b test is the sample to be tested, b i is a standard sample, and the category is the test sample is classified as the category that appears most frequently among the K nearest neighbors.
6. An iris recognition system based on graph neural network and anti-aliasing convolutional neural network, characterized in that: The iris recognition system adopts the iris recognition method based on graph neural network and anti-aliasing convolutional neural network according to any one of claims 1 to 5 to perform iris recognition, specifically comprising: Iris acquisition module, used for real-time acquisition of iris images and image preprocessing; The image segmentation module builds adaptive edges based on various knowledge to perform feature aggregation and accurately segment the iris; The iris classification module extracts the feature vector with texture information and uses the K-nearest neighbor algorithm KNN for classification through adaptive threshold coding.