A three-dimensional iris reconstruction and unfolding method

Abstract

The present invention discloses a three-dimensional iris reconstruction and expansion method, which relates to the field of intelligent medical care. The method comprises the following contents: step one, generating a three-dimensional iris tag; step two, establishing an iris depth map prediction model; and step three, expanding the iris. The present invention utilizes the anatomical constraint relationship between multiple iris images in a video stream to accurately predict iris depth information, realize three-dimensional iris reconstruction, and expand the iris into a two-dimensional rectangular image and depth map. The method overcomes the defect of insufficient depth of a traditional single image and utilizes multiple pictures in a video to predict iris image depth information. The iris textures between the iris pictures have an anatomical constraint relationship, which can effectively predict iris image depth information, improve the robustness of iris recognition, and is suitable for a variety of application scenarios.

Description

Technical Field

[0001] The present invention relates to intelligent medical treatment, and in particular to a three-dimensional iris reconstruction and expansion method. Background Art

[0002] As a high-precision biometric technology, iris recognition has been widely used in fields such as identity authentication and security monitoring. Traditional iris recognition methods rely mainly on a single two-dimensional iris image, which can only obtain the planar information of the iris and lacks important depth information.

[0003] When iris images are acquired from different angles, the 2D image will appear different due to the iris' inherent 3D structure. Although Daugman's rubber sheet model is commonly used for iris normalization, this model ignores the iris' 3D structure. This results in a significant degradation in 2D iris recognition performance in complex scenarios such as lighting variations, occlusions, and pose changes. The 3D effect of the iris is particularly pronounced in images taken from a side angle.

[0004] In recent years, a 3D iris reconstruction method based on optical coherence tomography (OCT) has been proposed. This method uses structured light to illuminate the eye to obtain a 3D model of the iris, generating a planar image and depth map for training depth prediction models. However, this method is limited in applicability and only works with iris images illuminated by specific structured light. It also fails to fully utilize the anatomical constraints of iris texture in video streams, limiting its universality and practical value.

[0005] With the development of deep learning technology, iris segmentation methods based on convolutional neural networks have emerged. These methods train network models using large amounts of labeled data and directly output segmentation results. However, these methods have the following limitations: 1. They simplify the iris into a two-dimensional structure, failing to consider its actual thickness and height, resulting in oversimplified pupil modeling and a lack of anatomical constraints. 2. Performance degrades under non-ideal conditions, with sensitivity to changes in viewpoint, poor occlusion handling, and inaccurate boundary localization. 3. Manual labeling is time-consuming and labor-intensive, resulting in inconsistent labeling results and a lack of anatomical verification, making data quality difficult to ensure. 4. Recognition accuracy is limited, environmental adaptability is insufficient, medical application value is limited, and system robustness is insufficient.

[0006] Traditional mathematical methods rely on multi-view images (at least two), are based on geometric principles and triangulation, and have a clear computational process and strong interpretability. However, their accuracy depends on image quality and camera calibration, and they are sensitive to noise. Deep learning methods can estimate depth from a single image, using data-driven learning features to implicitly learn depth cues, but their accuracy depends on the quality of the training data.

[0007] Advantages of traditional mathematical methods: solid theoretical foundation, no need for training data, and controllable and interpretable results; Disadvantages: multi-viewpoint requirements, computational complexity, and sensitivity to noise. Advantages of deep learning methods: single-image estimation, fast speed, and strong robustness; Disadvantages: large amounts of training data are required, generalization is limited, and results are difficult to interpret. Depth information obtained from a single image is somewhat dependent on lighting and shadows. Using multiple images to obtain depth information yields better results, but the technical difficulty increases, requiring key point extraction and calibration.

[0008] Recent research has shown that incorporating three-dimensional information can significantly improve the performance of biometric systems. Specifically for iris recognition, existing research has attempted to reconstruct a three-dimensional iris model from multiple two-dimensional images or estimate depth information using specialized equipment. However, these methods often require complex imaging setups or multiple image captures, limiting their practical application.

[0009] The main challenges facing iris-based biometric systems include: 1. Traditional two-dimensional iris images provide limited information and fail to account for the iris' natural curvature and depth variations. 2. Performance degrades under varying lighting conditions, viewing angles, and pupil dilation. 3. Accurate iris segmentation and normalization for feature extraction are difficult. 4. There is a lack of effective methods for reconstructing the iris' three-dimensional structure from limited two-dimensional imaging data.

[0010] In the field of 3D eye modeling, existing research has primarily focused on eye movement modeling and gaze estimation, with less consideration of the iris's 3D structural properties. Traditional eye models typically use a single spherical structure, which fails to accurately reflect the iris' thickness and height, resulting in the loss of important 3D information during iris segmentation and feature extraction.

[0011] Iris 3D reconstruction and unfolding technology has significant application value in medical diagnosis, biometrics, and intelligent medical systems. In medical diagnosis, reconstructing the fine structure of the iris can aid in the diagnosis of conditions such as iris coloboma, iris cysts, and glaucoma (iris morphological abnormalities). High-precision 3D models can be provided for laser iris surgery and artificial iris implantation, improving surgical safety. Iris thickness and vascularity can be quantified to study disease progression (e.g., diabetic iris disease). In biometrics, iris texture is unique and stable, and unfolded standardized images can be used to extract stable features (e.g., Gabor filter coding) for high-precision identity recognition. Reconstructing 3D iris features (e.g., microvascular structure) can distinguish authentic irises from forged images (e.g., contact lens imitations). In the field of intelligent medical system integration, iris recognition can replace traditional IDs in medical settings, ensuring accurate association between patients and electronic medical records. The feature encoding after iris unfolding is irreversible, preventing the leakage of original biometric information and meeting medical data security requirements. However, existing technologies mostly use a single iris image, which only contains two-dimensional information and lacks depth information. Some studies use OCT to obtain a three-dimensional iris model, generate an iris plan view and depth map from the 3D iris model, and train a depth model based on this to predict the depth map information of the iris image. This method uses a specific structured light to illuminate the eye to obtain an iris image. However, its limitation is that it is only applicable to iris images illuminated by structured light, lacks universality, and does not utilize the anatomical constraints of video stream information. The market needs a three-dimensional iris reconstruction and expansion method that can accurately use multiple images in a video to predict the depth information of the iris image, thereby improving recognition accuracy and robustness. Summary of the Invention

[0012] To address the problems in existing iris recognition technology, such as insufficient information in two-dimensional methods and difficulty in coping with complex scenarios, as well as the limited applicability of existing three-dimensional reconstruction methods and insufficient utilization of video information, the purpose of the present invention is to propose a more accurate and robust three-dimensional iris reconstruction and expansion method, aiming to better understand and process the three-dimensional structure of the iris, significantly improve the accuracy and robustness of iris recognition, and thus enhance the application effects in fields such as three-dimensional eye movement detection, iris recognition, security authentication, and medical diagnosis.

[0013] In order to achieve the above objectives, the present invention adopts the following innovative technical solutions:

[0014] A three-dimensional iris reconstruction and unfolding method, comprising the following contents:

[0015] Step 1: Generate 3D iris labels:

[0016] 1) OCT-based 3D iris model: Using OCT technology to obtain a high-precision 3D iris model, simulate and generate iris image video streams and their corresponding depth map labels;

[0017] 2) 3D iris model based on eye movement video: Using the iris segmentation results from the eye movement video, an eyeball model is constructed to obtain parameters such as the eyeball center, iris radius, and eyeball radius, and to calculate the parametric equations for the optical axis and iris circle.

[0018] Step 2: Iris depth map prediction model:

[0019] 1) Build a deep learning model that takes a time-series iris image sequence as input and outputs the corresponding iris depth map. By leveraging the correlation between consecutive frames in a video stream, this model overcomes the limitations of single-image depth estimation.

[0020] 2) A CNN encoder is used to extract multi-level features of each frame of iris image, an LSTM module is used to fuse the temporal information in the video sequence, and a multi-scale attention mechanism is introduced to enable the model to focus on image areas and features that are more important for depth prediction.

[0021] 3) Combined with anatomical constraint loss function: During the model training process, constraints based on the anatomical structure of the eyeball are introduced as part of the loss function, such as prior knowledge of the eyeball center, iris radius, and eyeball radius. This helps to improve the accuracy of depth prediction and the rationality of conforming to the real iris structure, overcoming the prediction results that may not conform to the anatomical structure produced by traditional deep learning methods.

[0022] Step 3: Iris expansion:

[0023] 1) Construct a 3D iris image using the 2D iris image and depth information;

[0024] 2) Expand the 3D iris into a 2D rectangular image and depth map. The specific steps include constructing a 3D coordinate system, calculating the rotation matrix, adjusting the iris to the orthographic position, and performing polar coordinate conversion.

[0025] In the aforementioned 3D iris reconstruction and expansion method, step 1, generating a 3D iris tag specifically includes:

[0026] 1) OCT-based 3D iris model: Using OCT technology to obtain a high-precision 3D iris model, simulate and generate iris image video streams and their corresponding depth map labels;

[0027] 2) 3D iris model based on eye movement video:

[0028] Parameter acquisition: Based on the iris segmentation results of the eye movement video, build the eyeball model and obtain the eyeball center , iris radius r, eyeball radius R and iris center ;

[0029] Optical axis calculation: Optical axis , unit optical axis ;

[0030] Iris circle parametric equation:

[0031] ;

[0032] Feature extraction and matching: The SIFT / SURF algorithm is used to extract iris texture features, calculate the three-dimensional coordinates of feature points, and generate depth map labels; the iris texture features include: pupil, pupillary ring, contraction groove, radial stripes, crypts, iris freckles, and Wolfflin's nodules.

[0033] In the aforementioned three-dimensional iris reconstruction and expansion method, the specific method of step 2, the iris depth map prediction model, includes:

[0034] 1) Build a deep learning model with a time-series iris image sequence as input and a corresponding iris depth map as output;

[0035] 2) A CNN encoder is used to extract temporal iris image features, an LSTM module is used to fuse temporal information, and a deep decoder is used to generate iris depth maps, combining an anatomically constrained loss function and a multi-scale attention mechanism;

[0036] The anatomical constraint loss function is:

[0037] The loss function constrains the predicted iris radius, eyeball center, eyeball radius, and the distance from the iris feature point to the eyeball center, thereby using the geometric structure information of the eyeball to standardize the prediction of the iris depth map; among them, the eyeball center , iris radius and eyeball radius The iris segmentation result based on the eye movement video is estimated, where is the weight used to balance the importance of anatomical constraints;

[0038] Assuming that N iris feature points are extracted, the predicted three-dimensional coordinates of the i-th feature point are , the predicted eye center is , and there is a desired distance ;So, ;

[0039] The multi-scale attention mechanism is as follows: multi-scale features are extracted by using convolution kernels of different sizes, and an attention mechanism is introduced in the decoding process to focus on key iris texture areas. It is first pre-trained on OCT simulation data, and then fine-tuned on real eye movement video data, combined with data enhancement to simulate lighting, occlusion and pupil zoom scenarios.

[0040] In the aforementioned three-dimensional iris reconstruction and expansion method, the temporal iris image sequence is at least 3 frames.

[0041] The aforementioned three-dimensional iris reconstruction and expansion method specifically includes:

[0042] 3D iris reconstruction: Construct a 3D iris image using the 2D images of the x and y axis information and the depth information of the z axis information;

[0043] Unfolding process:

[0044] 1) Construct a three-dimensional coordinate system with the center of the eyeball as the origin.

[0045] 2) Calculate the rotation matrix based on the iris center and eyeball center vectors , adjust the iris to the normal viewing position.

[0046] 3) Apply a rotation matrix to each feature point of the iris texture , adjust to the straight-ahead position.

[0047] 4) Through polar coordinate transformation, the iris circle is expanded into a rectangular texture image and depth map.

[0048] Parametric expansion:

[0049] 1) Texture unwrapping: Map the iris ring texture into a rectangular image.

[0050] 2) Depth Preservation: Synchronously expand the depth information and generate the corresponding depth map.

[0051] The aforementioned three-dimensional iris reconstruction and expansion method is characterized in that the information acquisition module for acquiring OCT images or eye movement videos is a head-mounted device.

[0052] The aforementioned three-dimensional iris reconstruction and unfolding method is applied to eye tracking, biometric authentication, medical diagnosis or computer vision and image processing. Furthermore, it is particularly suitable for applications with high precision and high robustness requirements in the following fields: three-dimensional eye movement detection, high-security biometric authentication, refined ophthalmic medical diagnosis and surgical planning.

[0053] The present invention is beneficial in that:

[0054] The present invention proposes to use multiple pictures in a video to predict the depth information of the iris image. Since the head-mounted device camera and the head are relatively fixed, the eyeball center remains stationary during eye movement, and the iris textures between each iris picture have an anatomical constraint relationship, the depth information of the iris image can be effectively predicted based on this, overcoming the limitation of insufficient depth of a single image.

[0055] The present invention introduces anatomical constraints into the deep learning model to improve the accuracy and robustness of depth prediction.

[0056] The present invention proposes a method for expanding a three-dimensional iris into a two-dimensional rectangular image and a depth map, which facilitates subsequent iris recognition and analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a schematic diagram of a single iris plane image and depth map in the prior art (the top is the iris plane image, and the bottom is the iris depth map);

[0058] Figure 2 The eyeball model and the optical axis iris depth map of the present invention;

[0059] Figure 3 Schematic diagram of iris texture features of the present invention; (Pupil refers to pupil, Collarette refers to pupillary ring, Contraction furrows refers to contraction furrows, Radial streaks refers to radial streaks, Crypts refers to crypts, Iris Freckles refers to iris freckles, Wolfflin nodules refers to Wolfflin nodules);

[0060] Figure 4 This is the iris segmentation result based on the parameters of the 3D eyeball model (Initial irisLabels, 3D Eyeball Model, Iris Annotation, Iris center, Powelloptimiza, Eye ball center, IrisRadius, Initial iris, Iris ellip).

[0061] Figure 5 Schematic diagram of iris texture feature extraction and matching (frontal view position) of the present invention;

[0062] Figure 6 Schematic diagram of iris texture feature extraction and matching (optical axis motion detection) of the present invention;

[0063] Figure 7 It is the three-dimensional iris expansion diagram of the present invention (Pupil / Iris Boundary refers to the pupil / iris boundary, and Iris / Sclera Boundary refers to the iris / sclera boundary);

[0064] Figure 8 It is a three-dimensional diagram of the iris circle of the present invention unfolded into a rectangle. DETAILED DESCRIPTION

[0065] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] A three-dimensional iris reconstruction and unfolding method, comprising the following contents:

[0067] Step 1: Generate 3D iris labels:

[0068] 1.1 OCT-based 3D iris model: Use OCT technology to obtain a high-precision 3D iris model, and simulate and generate iris image video streams and their corresponding depth map labels;

[0069] Based on OCT, a three-dimensional iris model is obtained, such as Figure 2 As shown in the figure, the simulation constructs the iris image video stream and the iris depth map label. The original method only creates a single iris image and the corresponding iris depth map, such as Figure 1 As shown, the upper part is the iris plane map, and the lower part is the iris depth map. The present invention provides high-precision three-dimensional ground truth for supervising the training of depth prediction models.

[0070] 1.2 3D iris model based on eye movement video: Based on the iris segmentation results of the eye movement video, the eyeball model is constructed to obtain parameters such as the eyeball center, iris radius, and eyeball radius, and calculate the parametric equations of the optical axis and iris circle:

[0071] 1.2.1 Parameter acquisition: Based on the iris segmentation results of the eye movement video, build the eyeball model and obtain the eyeball center , iris radius r, eyeball radius R and iris center ;

[0072] Based on the iris segmentation label of the eye movement video, the eyeball model is constructed, including the eyeball center, iris radius, and eyeball radius.

[0073] The eyeball center remains stationary during eyeball movement. Assume that the eyeball center is [x,y,0], the iris radius is r, the eyeball radius is R, and the iris center Iris is (x0,y0,z0). The optical axis can be determined based on the eyeball center and the iris center.

[0074] Optical axis calculation: Optical axis , unit optical axis

[0075] ;

[0076] Iris circle parametric equation:

[0077] ;

[0078] 1.2.2 Eye Movement Modeling

[0079] Modeling is performed based on the following parameters: eye center C=(x,y,0), origin of the coordinate system;

[0080] Iris center I=(x0,y0,z0)

[0081] Optical axis vector

[0082] Iris circle parametric equation:

[0083] ;

[0084] The iris segmentation results based on the parameters of the 3D eyeball model are as follows: Figure 4 shown.

[0085] 1.2.3 Feature extraction and matching: Use SIFT / SURF algorithm to extract iris texture features, calculate the three-dimensional coordinates of feature points, and generate depth map labels; iris texture features include: pupil, pupil ring, contraction groove, radial stripes, crypts, iris freckles and Wolfflin's tubercles, iris texture features such as Figure 3 shown.

[0086] 1.2.3.1 Front view position Figure 5 As shown in the figure, the iris texture extraction and matching diagram and optical axis motion detection are as follows Figure 6 shown.

[0087] 1.2.3.2 Calculation of 3D Coordinates of Iris Feature Points

[0088] Iris feature point set

[0089] The i-th feature point Pi=

[0090] The depth information is calculated based on the rotation matrix between the feature points of each image in the eye movement video and the normal view position.

[0091] 1.2.3.3 Creating an Iris Depth Map Label

[0092] ```Python

[0093] # Pseudocode example: Depth map generation

[0094] for frame in video_frames:

[0095] estimate_pose(eye_center, iris_radius)

[0096] compute_rotation_matrix(alpha, beta)

[0097] project_3d_points_to_depth_map(feature_points)

[0098] Step 2: Iris depth map prediction model:

[0099] 2.2.1 Build a deep learning model with a time-series iris image sequence (at least 3 frames) as input and a corresponding iris depth map as output;

[0100] 2.2.2 A CNN encoder is used to extract image features, and an LSTM module is used to fuse temporal information, combined with an anatomical constraint loss function and a multi-scale attention mechanism;

[0101] The following demonstrates the specific process of inputting three iris plane images and outputting three iris depth maps.

[0102] 2.2.2.1 Network Architecture

[0103] Input: Time-sequential iris image sequence (more than 3 frames)

[0104] Backbone network:

[0105] ```mermaid

[0106] graph TD

[0107] Input1 --> CNN_Encoder

[0108] Input2 --> CNN_Encoder

[0109] Input3 --> CNN_Encoder

[0110] CNN_Encoder --> LSTM_Fusion

[0111] LSTM_Fusion --> Depth_Decoder

[0112] ```

[0113] 2.2.2.2 The anatomical constraint loss function is: ;

[0114] 2.2.2.3 Multi-scale attention mechanism to enhance texture feature extraction:

[0115] Two-stage training: pre-training on OCT simulation data, then fine-tuning on real video data;

[0116] Data enhancement: simulate different lighting, occlusion, and pupil zoom states.

[0117] 3. Step 3: Iris expansion:

[0118] 3.1 Using the 2D iris image (x-axis and y-axis information) and depth information (z-axis information), a 3D iris image can be constructed to obtain 3D iris texture information.

[0119] 3.2 Expand the 3D iris into a 2D rectangular image and depth map. The specific steps include constructing a 3D coordinate system, calculating the rotation matrix, adjusting the iris to the orthographic position, and performing polar coordinate transformation.

[0120] Iris expansion is the process of expanding the iris in three dimensions, including the two-dimensional rectangular image (x-axis and y-axis information) and depth information (z-axis information), to construct a three-dimensional rectangular relief structure. The specific expansion process is as follows:

[0121] 3.2.1 Construct a three-dimensional coordinate system with the center of the eyeball as the origin

[0122] 3.2.2 Calculate the rotation matrix p for rotating to the normal view position based on the iris center and eyeball center vectors.

[0123] 3.2.3 3D iris circle, applying rotation matrix p to the front view position

[0124] 3.2.4 Each feature point vector of iris texture includes (x, y, z) information. Each feature point vector is rotated to the front view position by applying the rotation matrix p.

[0125] 3.2.5 Iris circle expansion, including two-dimensional rectangular image (x-axis and y-axis information) and depth information (z-axis information), as well as three-dimensional rectangular relief, three-dimensional iris expansion diagram as shown Figure 7 and Figure 8 shown.

[0126] Definition of parametric unwrapping: The iris is an annular region, typically located on a 3D surface (such as the surface of the eyeball). The purpose of parametric unwrapping is: Texture unwrapping: Mapping the annular iris texture (the image after 2D projection) from a circular shape to a rectangular image. Depth preservation: During the unwrapping process, the depth information (z-axis value) of each pixel in 3D space is preserved, generating a depth map corresponding to the unwrapped rectangular texture image.

[0127] 4. Usage examples:

[0128] 4.1. Input Data:

[0129] 3D iris data (e.g. point cloud format, containing x, y, z coordinates).

[0130] Iris center coordinates, inner and outer radius, and target size (width and height) of the expanded image.

[0131] 4.2. Projection:

[0132] The 3D iris is projected onto a 2D plane to generate a 2D texture image and a corresponding z-value map.

[0133] 4.3. Expand:

[0134] Perform polar coordinate transformation on the two-dimensional texture image to generate a rectangular texture image.

[0135] Perform the same polar coordinate transformation on the z-value map to generate a rectangular depth map.

[0136] Output:

[0137] The expanded rectangular texture image.

[0138] The corresponding rectangular depth map.

[0139] 4.5 Sample code:

[0140] def unwrap_iris_with_depth(img3D, center, r_min, r_max, width,height):

[0141] img2D = project_3d_to_2d(img3D) # Projection to two dimensions

[0142] z_values ​​= extract_z_values(img3D) # Extract depth value

[0143] texture = cv2.warpPolar(img2D, (width, height), center, r_max,cv2.WARP_INVERSE_MAP + cv2.INTER_CUBIC)

[0144] z_map = cv2.warpPolar(z_values, (width, height), center, r_max,cv2.WARP_INVERSE_MAP + cv2.INTER_CUBIC)

[0145] return texture, z_map.

[0146] 5. Application of this method

[0147] 5.1 Enhanced Biometric Authentication

[0148] The proposed 3D iris reconstruction method significantly improves the efficiency of biometric authentication by:

[0149] 1. Improve recognition accuracy, especially under challenging conditions such as varying lighting and viewing angles

[0150] 2. Enhanced resistance to spoofing attempts that rely on 2D iris representations

[0151] 3. Promote more accurate iris feature matching by incorporating depth information.

[0152] 5.2 Medical Diagnosis

[0153] In the medical field, our approach provides valuable capabilities for:

[0154] 1. More accurate assessment of the iris and anterior segment condition

[0155] 2. Track subtle changes in iris structure that may indicate disease progression

[0156] 3. Support surgical planning and evaluation in ophthalmic surgery

[0157] 4. Computer Vision and Image Processing

[0158] 5.3 Computer Vision and Image Processing

[0159] 3D iris reconstruction technology can be widely used in applications with high accuracy and high robustness requirements in the following fields:

[0160] The method is applied to eye tracking, biometric authentication, medical diagnosis, or computer vision and image processing, and is particularly suitable for applications with high precision and high robustness requirements in the following fields: three-dimensional eye movement detection, high-security biometric authentication, and refined ophthalmic medical diagnosis and surgical planning.

[0161] 1. 3D eye movement detection:

[0162] By reconstructing the true three-dimensional structure of the iris, the deformation effect caused by the traditional two-dimensional method when the viewing angle changes is overcome, and the movement trajectory of the iris in three-dimensional space can be tracked more accurately.

[0163] Combined with the depth prediction of the time-series iris image sequence, the depth changes of eye movements can be analyzed to provide more comprehensive eye movement information.

[0164] The unfolded iris image and depth map provide a standardized data format for subsequent eye movement analysis and modeling.

[0165] 2. High-security biometric authentication:

[0166] By utilizing the reconstructed 3D iris structure and expanded depth information, richer and more robust iris features can be extracted, improving the accuracy and anti-spoofing capabilities of identity recognition. For example, iris microstructure and depth information that are difficult to capture in 2D images can be analyzed.

[0167] The expanded feature encoding is irreversible, ensuring the security of the original biometric information. It is particularly suitable for scenarios with strict requirements on data security.

[0168] 3. Refined ophthalmic medical diagnosis and surgical planning:

[0169] Through high-precision three-dimensional iris reconstruction, the morphological characteristics of the iris can be evaluated more accurately, assisting in the diagnosis of eye diseases related to the three-dimensional structure of the iris, such as iris coloboma, iris cysts, and glaucoma.

[0170] Providing accurate three-dimensional models for surgeries such as iris laser surgery and artificial iris implantation helps doctors make more detailed surgical plans and improve the success rate and safety of the surgery.

[0171] Quantitative analysis of three-dimensional parameters such as iris thickness and vascularity can provide an objective basis for studying the progression of eye diseases and evaluating their therapeutic effects. This paper proposes using multiple video images to predict iris image depth information. Because the head-mounted device camera and head are relatively fixed, the center of the eye remains stationary during eye movement, and the iris textures between iris images have anatomical constraints, effectively predicting iris image depth information. This method overcomes the drawback of insufficient depth in traditional single images, improves the robustness of iris recognition, and is applicable to a variety of application scenarios, demonstrating significant innovation and practical value.

[0172] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any form, and any technical solutions obtained by equivalent replacement or equivalent transformation fall within the scope of protection of the present invention.

Claims

1. A three-dimensional iris reconstruction and unfolding method, characterized in that: Includes the following: Step 1: Generate 3D iris labels: 1) 3D iris model based on frequency domain OCT: Using OCT technology to obtain a high-precision 3D iris model, simulate and generate iris image video streams and their corresponding depth map labels; 2) 3D iris model based on eye movement video: Using iris segmentation results from eye movement videos, an eyeball model is constructed to obtain parameters such as the eyeball center, iris radius, corneal curvature, and eyeball radius. The optical axis and iris circle parameter equations are calculated. Iris texture features are extracted, and the 3D coordinates of feature points are calculated through feature matching to generate depth map labels. Step 2: Iris depth map prediction model: 1) Build a deep learning model with a time-series iris image sequence as input and a corresponding time-series iris depth map as output; 2) A CNN encoder consisting of multiple convolutional and pooling layers is used to extract temporal iris image features. An LSTM module is used to model the motion and deformation information of the iris over time. A deep decoder containing deconvolutional or upsampling layers is used to generate an iris depth map. This is combined with an anatomically constrained loss function and a multi-scale attention mechanism. The anatomical constraint loss function is: The loss function constrains the predicted iris radius, eyeball center, eyeball radius, and the distance from the iris feature point to the eyeball center, thereby using the geometric structure information of the eyeball to standardize the prediction of the iris depth map; among them, the eyeball center , iris radius and eyeball radius The iris segmentation result based on the eye movement video is estimated, where is the weight used to balance the importance of anatomical constraints; Assuming that N iris feature points are extracted, the predicted three-dimensional coordinates of the i-th feature point are , the predicted eye center is , and there is a desired distance ;So, ; The multi-scale attention mechanism extracts multi-scale features using convolution kernels of different sizes and introduces an attention mechanism during the decoding process to focus on key iris texture areas. The algorithm is pre-trained on OCT simulation data and then fine-tuned on real eye movement video data. Data augmentation is also used to simulate lighting, occlusion, and pupil zoom scenarios. Step 3: Iris expansion: 1) Using the 2D image and depth information of the iris, a corresponding depth value is assigned to each pixel of the 2D image to construct a 3D iris image; 2) Expand the 3D iris into a 2D rectangular image and depth map. The specific steps include constructing a 3D coordinate system, calculating the rotation matrix, adjusting the iris to the orthographic position, and performing a polar coordinate transformation based on relaxed geometric constraints.

2. The three-dimensional iris reconstruction and unfolding method according to claim 1, characterized in that: The specific contents of step 1, generating a 3D iris tag, include: 1) OCT-based 3D iris model: Using OCT technology to obtain a high-precision 3D iris model, simulate and generate iris image video streams and their corresponding depth map labels under different pupil sizes and lighting conditions; 2) 3D iris model based on eye movement video: Parameter acquisition: Based on the iris segmentation results of the eye movement video, the eyeball model is constructed to obtain the eyeball center including the corneal curvature , iris radius r, eyeball radius R and iris center ; Optical axis calculation: Optical axis , unit optical axis ; Iris circle parametric equation: ； Feature extraction and matching: The SIFT / SURF algorithm is used to extract iris texture features, and the three-dimensional coordinates of feature points are calculated through feature matching to generate depth map labels. The iris texture features include: pupil, pupillary ring, contraction groove, radial stripes, crypts, iris freckles, and Wolfflin's nodules.

3. The three-dimensional iris reconstruction and unfolding method according to claim 2, characterized in that: The temporal iris image sequence is at least 3 frames, so as to perform more accurate depth information prediction by utilizing the anatomical constraint relationship of iris texture between consecutive frames.

4. The three-dimensional iris reconstruction and unfolding method according to claim 1, characterized in that: The iris expansion method specifically includes: 3D iris reconstruction: Construct a 3D iris image using the x- and y-axis information of the 2D iris image and the z-axis information of the corresponding depth information map. Unfolding process: 1) Construct a three-dimensional coordinate system with the center of the eyeball as the origin; 2) Calculate the rotation matrix based on the iris center and eyeball center vectors , adjust the iris to the emmetropia position; 3) Apply the rotation matrix to each feature point vector of the 3D iris texture , so that it is converted to the front view position; 4) Through polar coordinate transformation, the 3D iris circle at the frontal view position is expanded into a 2D rectangular image containing texture information and the corresponding depth map; Parametric expansion: 1) Texture unfolding: Mapping the iris' annular texture from a two-dimensional projected image to a rectangular image; 2) Depth Preservation: While the texture is being expanded, the depth information of each pixel in three-dimensional space is simultaneously preserved to generate a depth map corresponding to the expanded rectangular texture image; the expanded two-dimensional rectangular image and depth map can be used to construct a three-dimensional rectangular relief structure.

5. The three-dimensional iris reconstruction and unfolding method according to claim 1, characterized in that: The information acquisition module for acquiring OCT images or eye movement videos is a head-mounted device.

6. The three-dimensional iris reconstruction and unfolding method according to claim 1, characterized in that: The described method has applications in eye tracking, biometric authentication, medical diagnostics, or computer vision and image processing.

Images