Elliptical iris positioning method, system and equipment based on deep convolutional neural network

Abstract

The embodiment of the invention provides an elliptic iris positioning method, system and device based on a deep convolutional neural network. The method is applied to the technical field of iris recognition, and comprises the following steps: marking a training data set to obtain ellipse parameter information of an ellipse frame in the training data set; designing elements of the deep neural network model, and constructing the deep neural network model; using the training data set to train the deep neural network model; performing forward reasoning on the iris image to be processed by using the trained deep neural network model to obtain a reasoning result, the reasoning result including ellipse parameter information of a prediction frame in the iris image; and performing post-processing on the reasoning result, and analyzing ellipse parameters corresponding to the iris and pupil boundary in the iris image according to the ellipse parameter information of the prediction frame, thereby realizing high-precision iris positioning. In this way, the technical problems that in the prior art, the anti-interference capacity is low, and the iris positioning precision needs to be further improved can be solved.

Description

Technical Field

[0001] The present disclosure relates to the field of iris recognition technology, and in particular, to an elliptical iris localization method, system, and device based on a deep convolutional neural network. Background Art

[0002] Iris recognition technology is a technology that uses the iris in the human eye for identity authentication and belongs to a type of human biometric recognition technology. Iris recognition has characteristics such as uniqueness, stability, non-contact, and high security, and is recognized as the most accurate and convenient biometric recognition technology. It has been widely applied to various scenarios that require precise identity authentication, such as finance, security, checkpoints, access control, and insurance. Iris localization refers to accurately determining the outer boundary of the iris and the outer boundary of the pupil in the iris image of the human eye, and it is one of the key technologies in iris recognition technology. It is generally considered that the boundary shapes of the pupil and the iris are close to circular, and many iris localization algorithms are also implemented based on the circular hypothesis. In fact, the more realistic boundary shapes of the iris and the pupil are elliptical, especially when the human eye is in a non-frontal gaze state. Therefore, an iris localization method based on the hypothesis that the boundary shapes of the iris and the pupil are elliptical can more realistically reflect the boundary shapes of the iris and the pupil, thereby improving the iris localization accuracy.

[0003] However, the existing technology has obvious deficiencies. For iris localization algorithms based on the elliptical hypothesis, in the early stage, image processing means were used to first detect the outer boundaries of the iris and the pupil, and then the elliptical parameters were fitted by a method similar to the least squares approximation; this method has very low anti-interference ability, and the actual effect is actually worse than that of iris localization algorithms based on the circular hypothesis. Therefore, it has not been actually applied and promoted. In addition, with the development of deep learning technology, object detection models have become mature, making it possible to implement a detection model with an elliptical shape as the target. Summary of the Invention

[0004] The present disclosure provides an elliptical iris localization method, system, and device based on a deep convolutional neural network, which can implement iris localization based on the elliptical hypothesis in a high-precision and high-efficiency manner, greatly promoting the development of iris localization technology. It solves the technical problems such as low anti-interference ability and insufficient iris localization accuracy in the existing technology.

[0005] According to the first aspect of the present disclosure, there is provided an elliptical iris localization method based on a deep convolutional neural network, including: Obtaining a training data set containing iris images, and marking the training data set to obtain elliptical parameter information of the elliptical frames in the training data set; Design the structure of the deep neural network model, data preprocessing, candidate boxes, calculation of the IOU of elliptical boxes, encoding of position information, selection of positive and negative samples, and loss design elements, and construct a deep neural network model. Use the training dataset to train the deep neural network model; Use the trained deep neural network model to perform forward inference on the iris image to be processed, and obtain the inference result. The inference result contains the elliptical parameter information of the predicted box in the iris image; Perform post-processing on the inference result. According to the elliptical parameter information of the predicted box, parse out the elliptical parameter corresponding to the boundary between the iris and the pupil in the iris image, and achieve high-precision iris localization.

[0006] According to the second aspect of the present disclosure, there is provided an elliptical iris localization system based on a deep convolutional neural network, including: A marking module for marking the training dataset to obtain the ground truth. The training dataset contains the iris images to be trained, and the ground truth contains the elliptical parameter information of the elliptical boxes in the training dataset; A construction module for constructing a deep neural network model; A training module for using the training dataset to train the deep neural network model; An inference module for using the trained deep neural network model to perform forward inference on the iris image to be processed, and obtain the inference result. The inference result contains the elliptical parameter information of the predicted box in the iris image; A post-processing module for performing post-processing on the inference result and parsing the elliptical parameter information corresponding to the boundary between the iris and the pupil in the iris image according to the position information of the predicted box.

[0007] According to the third aspect of the present disclosure, there is provided an electronic device. The electronic device includes: a memory and a processor. A computer program is stored on the memory, and when the processor executes the program, the method according to the first aspect of the present disclosure is implemented.

[0008] Compared with the prior art, the advantages and positive effects achieved by the present disclosure are: The present disclosure proposes an effective method for detecting the outer boundaries of the iris and the pupil based on an elliptical hypothesis. An elliptical box is marked on the iris image in the training dataset to complete model training, and the elliptical parameter information of the boundary between the iris and the pupil is parsed according to the elliptical parameter information of the predicted box given by the model, which greatly improves the accuracy and anti-interference ability of iris localization.

[0009] It should be understood that the content described in the invention content section is not intended to limit the key or important features of the embodiments of the present disclosure, nor is it used to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. Description of the Drawings

[0010] In combination with the accompanying drawings and with reference to the following detailed description, the above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent. The accompanying drawings are used to better understand the solution and do not limit the present disclosure. In the drawings, the same or similar reference numerals denote the same or similar elements, where: Figure 1 shows a flowchart of an elliptical iris localization method based on a deep convolutional neural network according to an embodiment of the present disclosure; Figure 2 shows a schematic diagram of information marking according to an embodiment of the present disclosure; Figure 3 shows a schematic structural diagram of a deep neural network model according to the present disclosure; Figure 4 shows a schematic structural diagram of a convolutional block ConvBlock according to the present disclosure; Figure 5 shows a schematic structural diagram of a basic block BaseBlock according to the present disclosure; Figure 6 shows a schematic structural diagram of a downsampling block DownBlock according to the present disclosure; Figure 7 shows a schematic diagram of a CTEBlock module according to the present disclosure; Figure 8 shows a schematic diagram of an InceptionBlock module according to the present disclosure; Figure 9 shows a block diagram of an elliptical iris localization system based on a deep convolutional neural network according to an embodiment of the present disclosure; Figure 10 shows a block diagram of an exemplary electronic device capable of implementing the embodiments of the present disclosure. Detailed implementation manners

[0011] To make the objectives, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure. Obviously, the described embodiments are some but not all of the embodiments of the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present disclosure without creative efforts shall fall within the scope of protection of the present disclosure.

[0012] In addition, the term "and / or" in this document is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this document generally represents an "or" relationship between the associated objects before and after.

[0013] Figure 1 The flowchart of an elliptical iris localization method 100 based on a deep convolutional neural network in an embodiment of the present disclosure is shown. As Figure 1 shown, the method 100 includes: S110: Obtain a training data set containing iris images, and label the training data set to obtain the elliptical parameter information of the elliptical frames in the training data set.

[0014] Optionally, in some embodiments, the elliptical parameter information of the elliptical frames in the training data set is the ground truth (GT); the information for labeling the training data set includes the elliptical frame parameter information (loc) corresponding to the boundary shapes of the iris and the pupil in an iris image, and the classification information (class) of the iris and the pupil.

[0015] It should be noted that in the embodiment, the elliptical frame parameter information corresponds to the elliptical parameters corresponding to the boundary shapes of the iris and the pupil. The elliptical parameters include , representing the corresponding elliptical center coordinates of the iris or the pupil, representing the lengths of the two axes of the ellipse, being the axis of the ellipse and the clockwise angle between the axes.

[0016] The classification information of the iris and the pupil, 1 represents the iris, and 0 represents the pupil.

[0017] According to the above description, the ground truth expression of a complete eye is , where represents the elliptical parameters of the iris, and '1' indicates that the category is the iris; represents the elliptical parameters of the pupil, and '0' indicates that the category is the pupil; a complete eye should contain and only contain one valid iris information and one valid pupil information. The information marking method of the ground truth refers to Appendix Figure 2 , Appendix Figure 2 where the label 1 represents the iris boundary and the label 2 represents the pupil boundary.

[0018] In the embodiments of the present application, the accuracy of data annotation and model training can provide high-quality annotation data for machine learning or deep learning models by accurately annotating the iris and pupil boundaries in iris images and recording the elliptical parameter information (such as center coordinates, axis lengths, rotation angles, etc.). These annotation data are the basis of model training and directly affect the performance of the model. Accurate annotation data can help the model better learn the features of the iris and pupil, thereby improving the accuracy and robustness of the model in tasks such as iris recognition and pupil detection. This is particularly important for biometric technologies (such as iris recognition) because iris recognition usually requires extremely high precision. Standardization of elliptical parameter information, the standardized annotation of elliptical parameter information (such as center coordinates, axis lengths, rotation angles, etc.) enables the iris and pupil boundaries in different images to be represented in a unified manner. Standardization helps the model maintain consistency when processing different images and avoid errors caused by different annotation formats. The standardized elliptical parameter information not only simplifies the model training process but also improves the interpretability of the model; researchers and developers can more intuitively understand the output of the model, facilitating optimization and debugging. Clarity of classification information, by labeling the iris and pupil as 1 and 0 respectively, clearly distinguishes the category information of the two. Classification information helps the model distinguish the features of the iris and pupil during the learning process and avoid confusion. Clear classification information can improve the classification accuracy of the model in iris recognition tasks, especially in complex image environments (such as lighting changes, occlusions, etc.), and the model can more accurately identify the boundaries of the iris and pupil. Expression of complete eye information, by combining the elliptical parameter information of the iris and pupil into a complete GT (Ground Truth) expression, ensures that the annotation information of each eye contains and only contains one iris and one pupil; this expression makes the annotation information of each eye more complete and standardized. The complete GT expression helps the model consider the features of both the iris and pupil simultaneously when processing a single eye, thereby improving the comprehensive performance of the model in tasks such as iris recognition and pupil detection; in addition, this expression also facilitates subsequent multi-task learning (such as performing iris recognition and pupil detection simultaneously).

[0019] In summary, the iris images in the training dataset of this embodiment are accurately labeled, and the standardized expressions of the elliptical parameter information and classification information provide a high-quality data basis for model training. The labeled data can help the model better learn the features of the iris and pupil, thereby improving the performance of the model in tasks such as iris recognition and pupil detection. It not only improves the accuracy and robustness of iris recognition technology, but also provides strong support for the development of biometric technology. Accurate iris recognition technology has broad application prospects in fields such as security authentication and identity recognition, and can bring higher security and convenience to society. Generally speaking, the above steps lay a solid foundation for the further development of iris recognition technology through accurate data annotation and standardized information expression, and have important technical significance and application value.

[0020] S120: Design elements such as the structure of the deep neural network model, data preprocessing, candidate boxes, calculation of the IOU of elliptical boxes, encoding of position information, selection of positive and negative samples, and loss design, and construct the deep neural network model.

[0021] Optionally, in some embodiments, the construction of the deep neural network model structure refers to Appendix Figure 3 ; For specific content, refer to Table 1, Figures 4 - 6 ; where Figure 4 Convolution block ConvBlock, Figure 5 Basic block BaseBlock, Figure 6 Downsampling block DownBlock; Figure 7 is the (multi-scale) fusion output block CTEBlock; Figure 8 is the feature fusion block InceptionBlock; Table 1 Deep neural network model structure

[0022] It should be noted that in the embodiment, the structure of the deep neural network model specifically includes: The first stage: includes multiple consecutive downsampling processes. Its function is to extract features during the rapid image degradation process. In this embodiment, 4 downsampling processes are used, the DownBlock structure is used, and the downsampling rate is 2.

[0023] The second stage: includes multiple consecutive feature fusion processes. Its function is to fuse the initially extracted features. In this embodiment, 3 fusion processes are used, the InceptionBlock structure is used, and no downsampling is performed during the fusion process.

[0024] The third stage: includes a process of outputting prediction box information at multiple scales, and its function is to output prediction results, including the position information of the prediction box, the classification information of the prediction box, etc. In this embodiment, 3 different scale channels are used.

[0025] The first stage (Stage 1): Downsampling and feature extraction. The purpose of this stage is to accurately capture the iris pupil boundary features in an efficient manner. Therefore, in the design of the model structure, it is necessary to balance the relationship between accuracy and efficiency simultaneously.

[0026] In this stage, the DownBlock downsampling module (such as Figure 6 ) is used, which includes 2 consecutive BaseBlocks (such as Figure 5 ), and each BaseBlock includes 2 consecutive basic ConvBlocks (such as Figure 4 ). Its expression is: ConvBlock = Conv + BN + RELU BaseBlock = ConvBlock(3 3) + ConvBlock(1 ) DownBlock = BaseBlock(s = 1) + BaseBlock(s = 2) It constitutes a Quad-Conv convolutional structure containing 4 convolutions. Among them, the BaseBlock is composed of 3 3 + 1 1 convolutions, which can ensure the accuracy of feature extraction, increase the depth of the model, and at the same time, the 1 1 convolution is also beneficial to improving efficiency. DownBlock first uses a convolution with step = 1, and then uses a convolution with setp = 2 to achieve downsampling, which can further deepen the fusion degree between neighborhood features and improve the accuracy of feature extraction.

[0027] Considering that the Quad-Conv convolutional structure has a strong fine-grained feature extraction ability, in order to improve the execution efficiency of the model, the number of channels of the convolution is appropriately restricted. The initial channels are quickly expanded to 64 channels to ensure that enough original information enters, but the subsequent channels do not adopt the 2-fold expansion strategy, but are restricted within 128 channels to improve the execution efficiency of the model.

[0028] The second stage (Stage 2): Feature fusion and receptive field adjustment. In this example, 3 consecutive InceptionBlocks (such as Figure 8 ) are used to achieve the required functions.

[0029] The used InceptionBlock includes 4 different sub-channels, and each sub-channel can perform feature extraction. After being combined by 3 consecutive InceptionBlocks, 64 different feature fusion methods can be formed.

[0030] At the same time, each InceptionBlock can form 3 different receptive fields. After combining 3 consecutive InceptionBlocks, 7 different receptive fields can be formed, which greatly improves the model's object detection ability for different granularity scales.

[0031] In addition, there is a 1×1 convolutional channel in the InceptionBlock, which plays a role similar to the residual jump connection and can more effectively transmit gradient information to ensure the trainability of the deep model. 1×1 convolutional channel, which plays a role similar to the residual jump connection and can more effectively transmit gradient information to ensure the trainability of the deep model.

[0032] The third stage (Stage 3): multi-scale output and ellipse parameter prediction, including the ellipse prediction result and the classification prediction result of the iris image. The prediction box of the iris image needs to be represented in the form of an ellipse, and the output layer is specially designed; it also includes: Ellipse parameter regression head: Design an ellipse parameter regression head to directly predict the five parameters of the ellipse:

[0033] In the formula, is the ellipse regularization module, is the regularization parameter.

[0034] Multi-scale prediction module: In this example, feature fusion outputs of 3 different scales are used to support target predictions of multiple different scales. The output points of the 3 output scales are respectively after the InceptionBlock, after one downsampling, and after another downsampling. Then, the different-scale outputs of the 3 output points are fused through the CTEBlock to obtain the final multi-scale output result.

[0035]

[0036] Among them, are the feature maps of 3 scales respectively.

[0037] It should be noted that in the embodiments of this application, data preprocessing includes preprocessing of iris image data and preprocessing of labeled data.

[0038] The preprocessing of iris image data includes: The iris image data undergoes various data augmentation processes. Operations such as deformation, cropping, horizontal flipping, or rotation that may affect the position information of the labeled data are performed, and the corresponding labeled position information is adjusted simultaneously.

[0039] The image after data augmentation contains at least one complete labeled eye image.

[0040] The final input image is scaled to a single-channel grayscale image of the specified size (width 640 × height 480), and after being normalized, it is input into the deep neural network model to be trained.

[0041] Normalization formula:

[0042] In the formula, represents the pixel value of the iris image after normalization; represents the j-th pixel value in the iris image before processing; represents the statistical average value of the j-th pixel value in the training dataset; represents the standard deviation of the j-th pixel value in the training dataset; N represents the total number of pixels in the iris image; represents the natural exponential function, which is used to enhance the smoothness of the pixel value distribution; represents the normalization constant to ensure the mathematical integrity of the formula. First, each pixel value is normalized, that is, subtracting its average value and dividing by the standard deviation. This step ensures that all pixel values are centered around zero; a Gaussian weighting function is introduced. By calculating the difference between each pixel value and the average value and using it as the input of the Gaussian function, a smooth weight value can be obtained; all weighted pixel values are divided by the normalization constant to ensure the mathematical integrity of the entire formula and make the normalized pixel values conform to the standard normal distribution.

[0043] The preprocessing of the labeled data includes the following steps: Cooperate with the augmentation processing operation that may affect the position information of the labeled data, and adjust the corresponding position information.

[0044] After the position information is finally normalized, it is used as the new Ground Truth information and input into the network model to be trained.

[0045] The normalization method is ; where is the new coordinate point after normalization, is the coordinate point before processing; is the new ellipse axis length after normalization, is the ellipse axis length before processing; width is the pixel width of the corresponding image, and hight is the pixel height of the corresponding image.

[0046] It should be noted that in the embodiment of the present application, the candidate ellipse box design: There are 3 width sizes, 1 aspect ratio, and 3 densities for the candidate ellipse box design, which are represented by different step sizes.

[0047] The marking method of the candidate ellipse box is where is the center coordinate of the candidate ellipse box after normalization, is the axis length of the ellipse box after normalization, is the ellipse box axis and the clockwise angle between the axes.

[0048] It should be noted that in the embodiments of the present application, the position information encoding includes: The input is the normalized ellipse parameter information Loc of an ellipse box , where represents the center point coordinate of the ellipse box, represents the axis length of the ellipse, is the ellipse box axis and the clockwise angle between the axes, expressed in radians.

[0049] The output is , where represents the center encoding result of the ellipse box, represents the axis length encoding result of the ellipse box, is the rotation angle encoding result of the ellipse box.

[0050] Among them, , , is the normalized center coordinate of the candidate ellipse box, is the normalized ellipse axis length of the candidate ellipse box, is the encoding coefficient.

[0051] ,

[0052] is the encoding coefficient; , is the normalized ellipse rotation angle of the candidate box.

[0053] The decoding process is the inverse process of the encoding process.

[0054] It should be noted that in the embodiments of the present application, the IOU calculation of the ellipse box includes the following steps: Use a polygon (such as a 16-sided polygon) to represent an ellipse, greatly reducing redundant information and improving accuracy.

[0055] Calculate the intersection polygon of the two polygons.

[0056] Calculate the area of the intersection polygon.

[0057] Calculate the intersection over union of the two polygons to obtain the IOU value.

[0058] In iris image processing, the calculation of the IOU (Intersection over Union) of the elliptical box is a crucial step for evaluating the similarity of two elliptical regions; a 16-sided polygon will be used to approximate the ellipse.

[0059] Approximate the ellipse with a 16-sided polygon (regular hexadecagon), and let the center of the ellipse be , the major axis radius be , the minor axis radius be , and the rotation angle be , then the parametric equation of the ellipse is: ;

[0060] In the formula, is the parameter angle, and the value range is ; Divide into 16 equal parts to obtain 16 vertex coordinates:

[0061] The coordinate of each vertex is:

[0062] Calculation of the intersection polygon. Suppose there are two ellipses, which are represented by 16-sided polygons and respectively; it is necessary to calculate the intersection polygon of these two polygons; Find all the intersection points of and ; Divide the polygon into multiple sub-polygons according to the intersection points; Select the sub-polygons located inside both polygons to form .

[0063] Calculation of the area of the intersection polygon :

[0064] In the formula, are the vertex coordinates of the polygon, is the number of vertices, and , .

[0065] The calculation of the area of the union polygon is the sum of the areas of the two polygons minus the intersection area:

[0066] In the formula, and are the areas of the two polygons respectively.

[0067] The final IOU value is the ratio of the intersection area to the union area:

[0068] Through the above steps, the IOU value of the ellipse box can be accurately calculated in iris image processing, providing a reliable basis for matching and recognition.

[0069] It should be noted that in the embodiments of this application, the selection of positive and negative samples includes the following steps: Samples with IOU > 0 between the predicted box and the candidate box and ranked among the top N are positive samples; the others are candidate negative samples.

[0070] Calculate the classification losses of all candidate negative samples and sort them.

[0071] Select the most difficult negative samples, that is, the M candidate negative samples with the largest losses, as the finally selected negative samples.

[0072] Among them , α is the negative sample magnification factor, and α = 10.

[0073] It should be noted that in the embodiments of this application, in the design of the loss function, the training loss of the deep neural network model is designed as a combined loss, using a combination of two losses with different weights, including the loss function and the classification function.

[0074] The loss function is used to regress the position information and uses smoothL1. The calculation formula is: ; Among them, is the normalized encoded numerical difference value between the predicted box and the true box.

[0075] The classification function uses the cross-entropy loss function to confirm whether the predicted box is an iris or a pupil. The calculation method is as follows:

[0076] Among them, is the total number of samples, is the total number of categories; represents the true sample the th label of the category, using one-shot encoding; represents the sample the th predicted probability of the category. In this embodiment, K = 3, representing the iris, pupil, and background categories respectively.

[0077] In summary, the above two losses are superimposed according to different weights. The calculation method is:

[0078] Among them, is the calculation weight of the th loss; is the calculated value of the th loss; in this embodiment , .

[0079] In the embodiment of the present application, for the model structure design, the design of each module is to better extract image features while reducing the computational complexity. The convolutional block is usually composed of a convolutional layer, an activation function (such as ReLU), and a normalization layer (such as BatchNorm). Its main function is to extract local features. Through the sliding window operation of the convolutional kernel, it captures details such as textures and edges in the image; the design of the convolutional block enables the model to efficiently process high-dimensional data. At the same time, through multi-layer stacking, higher-level semantic features are gradually extracted; multi-level features are extracted, gradually abstracting from low-level to high-level; the number of parameters is reduced, and the computational efficiency is improved; non-linearity is introduced through the activation function to enhance the expressive ability of the model. The convolutional block is the basis of the deep neural network and determines the feature extraction ability of the model; by reasonably designing the convolutional kernel size, stride, and number of channels, the performance of the model can be significantly improved. The basic block (BaseBlock, Figure 5 ) is an extension of the convolutional block. The downsampling block (DownBlock, Figure 6 ) is composed of convolutional layers with strides. Its function is to reduce the resolution of the feature map, reduce the computational amount, and at the same time expand the receptive field to capture more global features; reduce the size of the feature map and reduce the computational complexity; enhance the robustness of the model to input size changes. The downsampling block is the key to the efficient operation of the model. By reasonably designing the downsampling strategy, the computational cost can be reduced while ensuring performance. The multi-branch structure (Inception module) uses a variety of different branch structures, which can adaptively adjust the weights of different branches and enhance the adaptability of feature extraction; by combining multiple multi-branch structures, not only can the receptive field be expanded, but also more combined branches can be formed. Each combined branch channel has a different receptive field, which is more suitable for multi-scale object detection.

[0080] Data preprocessing is an important part of model training, which directly affects the convergence speed and final performance of the model; it is the basis of model training, and the preprocessing strategy can significantly improve the performance and stability of the model.

[0081] The IOU calculation between the candidate box and the elliptical box calculates the ratio of the intersection and union of the rectangular boxes, which is used to evaluate the accuracy of the prediction box. The IOU calculation of the elliptical box is more complex and usually requires numerical integration or approximation algorithms, which is applicable to special scenarios. IOU is the core metric in the object detection task and directly affects the localization accuracy of the model. By optimizing the IOU calculation, the detection performance of the model can be improved.

[0082] Position information encoding is to convert the position of the target box (such as the center coordinates, width and height) into a format that the model can learn. It is the key to the object detection task and can improve the localization accuracy and generalization ability of the model.

[0083] Positive and negative sample selection In the object detection task, the selection of positive and negative samples directly affects the training effect of the model. The selection of positive and negative samples is the core of model training, which can accelerate model convergence and improve detection performance.

[0084] The loss function is the goal of model optimization, usually including two parts: classification loss and regression loss. The classification loss measures the difference between the predicted class and the true class (such as cross-entropy loss); the regression loss measures the position difference between the predicted box and the true box (such as Smooth L1 loss). The design of the loss function directly affects the optimization direction of the model and can improve the detection accuracy and robustness of the model.

[0085] In summary, in this embodiment, by reasonably designing the structure of the deep neural network model, data preprocessing, candidate box, IOU calculation of the elliptical box, position information encoding, positive and negative sample selection, and loss function, an efficient object detection model can be constructed. Attached Figure 3 , Table 1, Figures 4 - 7 Each component in realizes different functions and works together to finally achieve the optimization and performance improvement of the model.

[0086] S130: Train the deep neural network model using the training data set.

[0087] Optionally, in some embodiments, the labeled iris images and data are split into two sets: a training data set and a validation data set, with a ratio of 9:1.

[0088] Use the training data set for training and the validation data set for validating the deep neural network model and screening parameters.

[0089] Input the preprocessed iris images and labeled data of the training data set into the deep neural network, and the deep neural network performs forward inference on the input iris images to obtain the current inference result.

[0090] Use the current inference result and the corresponding labeled data to calculate the classification loss and regression loss to obtain the current combined loss.

[0091] The current loss is propagated back into the deep neural network through reverse gradient propagation, and together with other parameters such as the learning rate, the weights of the current deep neural network are adjusted.

[0092] Repeat the above process until the training ends.

[0093] The condition for the end of training is: completing the specified number of rounds (e.g., epoch = 120).

[0094] The screening conditions for the deep neural network model parameters are: the metrics of the validation dataset are the best results of the current training and meet the expectations. The metrics in this embodiment include the IOU metric [IOU > 0.95] and the F1Score metric [F1Score > 0.99].

[0095] It should be noted that in the embodiment of the present application, during the forward inference process of the deep neural network on the input iris image, the core lies in gradually extracting discriminative features from the iris image through multi-level feature extraction and information transformation, and finally generating an inference result, which specifically includes the following steps: After the input iris image is preprocessed, it enters the primary feature extraction stage, and basic structural information such as texture, edges, and local details is extracted from the iris image; the specific process is as follows: Local detail capture, through pixel-by-pixel local calculations, extracts the minute changes and detail features in the iris image, manifested as the bright or color mutation regions in the iris texture; Spatial relationship modeling, based on the local details, further captures the spatial relationships between adjacent pixels to form preliminary structured features; for example, the spatial distribution between the stripes and spots in the iris texture.

[0096] Intermediate feature extraction, based on the primary features, further models the complex patterns in the iris image; more semantically meaningful features such as texture combinations, the shapes and spatial distributions of local regions, etc. are extracted; the specific process is as follows: Feature combination and abstraction: Combine and abstract the primary features to form a higher-level feature representation; describe the more complex visual patterns in the iris image, such as the periodic changes in texture or the shape features of local regions.

[0097] Context information fusion: By introducing the context information of the local region, enhance the representational ability of the features; for example, correlate the texture features of a certain region with the features of its surrounding regions to form a richer feature description.

[0098] Advanced feature extraction: Based on the intermediate features, further models the global information in the iris image; features that can describe the overall semantics of the iris, such as the overall texture distribution, structural relationships, and category information of the iris, etc. are extracted. The specific process is as follows: Global information aggregation: The intermediate features are globally aggregated to form a feature representation that can describe the overall semantics of the iris, with high abstraction and discriminability.

[0099] Semantic relationship modeling: Based on the global information, the semantic relationship between different regions in the iris image is captured to form a complete semantic description; for example, the relationship between the global distribution of iris texture and local details.

[0100] Inference result generation: After high-level feature extraction is completed, the inference result generation stage converts high-level features into final inference results; based on the extracted features, task-related outputs (such as classification labels or target boxes) are generated. The specific process is as follows: Feature mapping and transformation: Map high-level features to task-related output spaces to generate preliminary reasoning results.

[0101] Result optimization and screening: Optimize and screen the preliminary inference results to ensure that they meet the task requirements; for example, remove redundant or low-confidence results through methods such as threshold screening or non-maximum suppression.

[0102] Hierarchical information transmission and feedback, the specific process is as follows: Information transfer: The feature extraction results of each level will be used as the input of the next level to ensure the layer-by-layer transmission and accumulation of information.

[0103] Feedback mechanism: After the inference result is generated, the result is compared with the labeled data through the feedback mechanism to generate a loss signal, and the feature extraction process at each level is adjusted through back propagation to further improve the accuracy of the inference result.

[0104] In the embodiment of the present application, the forward reasoning process gradually extracts discriminative features from the iris image through hierarchical information extraction and conversion, and finally generates reasoning results. It is highly creative and hierarchical, can effectively improve the performance of deep neural networks in iris image tasks, and meets the definition of creativity in patent law.

[0105] S140: Using the trained deep neural network model, forward reasoning is performed on the iris image to be processed to obtain a reasoning result, where the reasoning result includes ellipse parameter information of the prediction box in the iris image.

[0106] S150: Post-process the inference result, and parse the ellipse parameters corresponding to the boundary between the iris and the pupil in the iris image according to the ellipse parameter information of the prediction box to achieve high-precision iris positioning.

[0107] Optionally, in some embodiments, the inference result is first decoded; the decoded data is processed by non-maximum suppression to screen out the best-matching prediction boxes. The IOU threshold used for non-maximum suppression is set to 0.7, and the classification threshold is set to 0.8; the screened prediction boxes are de-normalized to obtain the elliptical parameter information of the prediction boxes.

[0108] In the embodiments of the present application, post - processing is performed on the inference results. Specifically, through steps such as decoding, non - maximum suppression (NMS), and inverse normalization, the elliptical parameters corresponding to the boundaries of the iris and the pupil in the iris image are parsed. This process has important technical effects and practical significance in the iris localization task. Through decoding processing and inverse normalization operations, the predicted box information output by the deep neural network can be converted into actual elliptical parameters (such as center coordinates, major axis, minor axis, and rotation angle, etc.), so as to accurately describe the boundaries of the iris and the pupil; non - maximum suppression (NMS) can effectively eliminate redundant or low - confidence prediction results by screening out the best - matching predicted boxes, further improving the localization accuracy. In complex scenarios (such as light changes, blurred iris texture, or occlusion, etc.), NMS and threshold screening can filter out inaccurate predicted boxes to ensure that the finally output elliptical parameters have high reliability; through the dual screening of the IOU threshold (0.7) and the classification threshold (0.8), the problems of false detection and missed detection can be effectively avoided, improving the robustness of the model. The NMS processing can significantly reduce the number of redundant predicted boxes, reduce the subsequent calculation overhead, and thus improve the overall processing efficiency; the inverse normalization operation converts the predicted box parameters from the normalized space to the actual image space, avoiding unnecessary calculations and further optimizing the calculation efficiency; by parsing the elliptical parameter information of the predicted box, the boundaries of the iris and the pupil can be accurately fitted, providing high - quality input data for tasks such as iris recognition and identity verification. Achieved significance: High - precision iris localization is a key step in the iris recognition system. Through post - processing operations, the accuracy of iris boundary extraction can be significantly improved, thereby improving the overall performance of the iris recognition system; accurate iris localization can reduce errors in subsequent feature extraction and matching processes, improving the recognition rate and reliability; in the fields of security, finance, healthcare, etc., the iris recognition technology has attracted much attention due to its high security and uniqueness. High - precision iris localization technology can promote the wide application of iris recognition technology in these fields; for example, in scenarios such as mobile device unlocking, access control systems, and identity verification, high - precision iris localization can significantly improve the user experience and system security. In practical applications, iris images may be affected by factors such as light, occlusion, and blur. Through post - processing operations such as NMS and threshold screening, stable iris localization can be achieved in complex scenarios, providing reliable technical support for practical applications; the above - mentioned post - processing method combines the advantages of deep learning and traditional image - processing technologies, providing an innovative solution for the iris localization task. It can provide reference for other similar tasks (such as face key - point detection, object tracking, etc.), promoting the further development of deep learning technology in the field of computer vision.

[0109] In summary, through post-processing operations such as decoding, non-maximum suppression, and inverse normalization on the inference results, this embodiment can achieve high-precision iris localization, improving the accuracy, robustness, and computational efficiency of the model. It can not only promote the development of iris recognition technology but also provide a reliable solution for iris localization in complex scenarios, facilitating the innovation and application of deep learning technology in the field of computer vision.

[0110] It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that this disclosure is not limited by the described action sequence because, according to this disclosure, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this disclosure.

[0111] The above is the introduction to the method embodiments. The following further illustrates the solution of this disclosure through system embodiments.

[0112] Figure 9 The block diagram of an elliptical iris localization system 200 based on a deep convolutional neural network according to an embodiment of the present disclosure is shown. As Figure 9 shown, the system 200 includes: A marking module 210 for marking the training data set to obtain the ground truth. The training data set contains iris images to be trained, and the ground truth contains the elliptical parameter information of the elliptical frames in the training data set.

[0113] A construction module 220 for constructing a deep neural network model; A training module 230 for training the deep neural network model using the training data set; An inference module 240 for performing forward inference on the iris image to be processed using the trained deep neural network model to obtain an inference result; the inference result contains the elliptical parameter information of the predicted frame in the iris image; A post-processing module 250 for post-processing the inference result and parsing the elliptical parameter information corresponding to the iris and pupil boundaries in the iris image according to the position information of the predicted frame.

[0114] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the described modules can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated here.

[0115] Optionally, in some embodiments, the marking information of the marking module 210 for the training data set includes: the elliptical parameter information corresponding to the iris and pupil boundaries in the iris image, and the classification information of whether it is an iris or a pupil.

[0116] The building block 220 includes: a preprocessing unit, a design unit, an encoding unit, a selection unit, and a calculation unit.

[0117] The preprocessing unit is used to preprocess the image and the marker data.

[0118] The design unit is used to design an elliptical frame for the eye images in the training dataset.

[0119] The encoding unit is used to encode the position information of the elliptical frame.

[0120] The selection unit is used to select positive and negative samples from the training dataset.

[0121] The calculation unit is used to calculate the combined loss.

[0122] It should be noted that in the embodiments of the present application, the calculation unit is specifically used for: calculating the classification loss; calculating the regression loss; performing a weighted average on the obtained classification loss and regression loss to obtain the combined loss.

[0123] In some embodiments, the building block 220 further includes: a building unit.

[0124] The building unit is used to build a network model structure including three stages, and the three stages include: a downsampling process, a feature fusion process, and a multi-scale prediction box information output process.

[0125] In some embodiments, the training module 230 includes: an inference unit, a loss calculation unit, an adjustment unit, and a repeated execution unit.

[0126] The inference unit is used to input the preprocessed image and marker data of the training dataset into the network, and the network performs forward inference on the input image to obtain the current inference result.

[0127] The loss calculation unit is used to calculate the classification loss and the regression loss using the current inference result and the corresponding marker data to obtain the current combined loss.

[0128] The adjustment unit is used to backpropagate the loss into the network through reverse gradients and adjust the current network weights together with other parameters.

[0129] The repeated execution unit is used to repeat the above process until the training ends.

[0130] It should be noted that in the embodiments of the present application, the current inference result includes multiple adjacent result values.

[0131] In some embodiments, the post-processing module 250 includes: a decoding unit, a suppression unit, an inverse normalization unit, and a judgment unit.

[0132] The decoding unit is used to first perform decoding processing on the inference result.

[0133] The suppression unit is used to perform non-maximum suppression processing on the data after decoding processing to screen out the best-matching prediction boxes.

[0134] The inverse normalization unit is used to perform inverse normalization on the screened prediction boxes to obtain the elliptical parameter information of the prediction boxes.

[0135] The judgment unit is used to perform iris and pupil judgment according to the set classification threshold.

[0136] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device.

[0137] Figure 10 A schematic block diagram of an electronic device 300 that can be used to implement the embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as, for example, a laptop computer, a desktop computer, a workbench, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as, for example, a personal digital processor, a cellular phone, a smart phone, a wearable device, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0138] The electronic device 300 includes a computing unit 301, which can execute various appropriate actions and processes according to a computer program stored in the ROM 302 or a computer program loaded from the storage unit 308 into the RAM 303. In the RAM 303, various programs and data required for the operation of the electronic device 300 can also be stored. The computing unit 301, the ROM 302, and the RAM 303 are connected to each other via a bus 304. The I / O interface 305 is also connected to the bus 304.

[0139] Multiple components in the electronic device 300 are connected to the I / O interface 305, including: an input unit 306, such as a keyboard, a mouse, etc.; an output unit 307, such as various types of displays, speakers, etc.; a storage unit 308, such as a disk, an optical disc, etc.; and a communication unit 309, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 309 allows the electronic device 300 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0140] The computing unit 301 may be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 301 executes the various methods and processes described above, such as the elliptical iris localization method based on a deep convolutional neural network. For example, in some embodiments, the elliptical iris localization method based on a deep convolutional neural network may be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 308. In some embodiments, part or all of the computer program may be loaded and / or installed onto the electronic device 300 via the ROM 302 and / or the communication unit 309. When the computer program is loaded into the RAM 303 and executed by the computing unit 301, one or more steps of the elliptical iris localization method based on a deep convolutional neural network described above may be executed. Alternatively, in other embodiments, the computing unit 301 may be configured to execute the elliptical iris localization method based on a deep convolutional neural network by any other suitable means (e.g., by means of firmware).

[0141] The various embodiments of the systems and techniques described above in this document may be implemented in digital electronic circuitry, integrated circuit systems, field-programmable gate arrays (FPGA), application-specific integrated circuits (ASIC), application-specific standard products (ASSP), system-on-a-chip (SOC), complex programmable logic devices (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include: implemented in one or more computer programs that may be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit the data and instructions to the storage system, the at least one input device, and the at least one output device.

[0142] The program code for implementing the methods of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when the program codes are executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program codes may be executed entirely on the machine, partially on the machine, as an independent software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0143] In the context of this disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0144] To provide for interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can also be used to provide for interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic, speech, or tactile input).

[0145] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which the user can interact with an implementation of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of a communication network include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0146] A computer system can include a client and a server. The client and the server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server can be a cloud server, a server of a distributed system, or a server incorporating a blockchain.

[0147] It should be understood that the various forms of processes shown above can be used, with steps reordered, added or deleted. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and no limitations are imposed herein.

[0148] The above specific embodiments do not constitute a limitation on the protection scope of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub - combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the protection scope of this disclosure.

Claims

1. An elliptical iris positioning method based on deep convolutional neural network, characterized in that: include: Acquire a training data set containing iris images, label the training data set, and obtain ellipse parameter information of the ellipse frame in the training data set; Design the deep neural network model structure, data preprocessing, candidate box, elliptical box IOU calculation, position information encoding, positive and negative sample selection and loss design elements, build a deep neural network model and use the training data set to train the deep neural network model; Using the trained deep neural network model, forward reasoning is performed on the iris image to be processed to obtain the reasoning result, which includes the ellipse parameter information of the prediction box in the iris image; The inference results are post-processed, and the ellipse parameters corresponding to the iris and pupil boundary in the iris image are parsed according to the ellipse parameter information of the prediction box to achieve high-precision iris positioning.

2. The elliptical iris positioning method based on deep convolutional neural network according to claim 1 is characterized in that: The ellipse parameter information of the ellipse frame in the training data set is the truth; the information for marking the training data set includes the ellipse frame parameter information corresponding to the iris and pupil boundary shape in an iris image, and the classification information of the iris and the pupil; The ellipse frame parameter information corresponds to the ellipse parameters corresponding to the shape of the iris and pupil boundary. The ellipse parameters include , Represents the corresponding ellipse center coordinates of the iris or pupil, Represents the lengths of the two axes of the ellipse, For ellipse Axis and The clockwise angle of the axis; Iris and pupil classification information, 1 represents iris, 0 represents pupil; The truth of a complete eye is expressed as ,in Represents the ellipse parameter of the iris, 1 means the category is iris; Represents the pupil ellipse parameters, 0 means the category is pupil; a complete eye should contain and only contain one valid iris information and one valid pupil information.

3. The elliptical iris positioning method based on deep convolutional neural network according to claim 1, characterized in that: The structure of the deep neural network model includes: The first stage: includes multiple consecutive downsampling processes, which is used to extract features during rapid image degradation; The second stage: includes multiple continuous feature fusion processes, which is used to fuse the initially extracted features; The third stage: includes the process of multi-scale output prediction box information, which is used to output the prediction results, including the location information of the prediction box and the classification information of the prediction box.

4. The elliptical iris positioning method based on deep convolutional neural network according to claim 3 is characterized in that: The first stage includes: Use a specially designed Quad-Conv structure.

5. The elliptical iris positioning method based on deep convolutional neural network according to claim 3 is characterized in that: The second stage includes: feature fusion. The feature fusion of iris images needs to balance local details and global structures, which specifically includes: Use multiple InceptionBlocks to achieve feature fusion and receptive field expansion.

6. The elliptical iris positioning method based on deep convolutional neural network according to claim 3, characterized in that: The third stage of multi-scale output and ellipse parameter prediction, the prediction box of the iris image needs to be expressed in the form of an ellipse, and the output layer is specially designed; including: Ellipse parameter regression head, used to design the ellipse parameter regression head to directly predict the five parameters of the ellipse; Multi-scale prediction module, used to combine feature maps of different scales.

7. The elliptical iris positioning method based on deep convolutional neural network according to claim 1, characterized in that: Preprocessing of iris image data, including: The iris image data is subjected to various data augmentation processes, including deformation, cropping, horizontal flipping, i.e., rotation, which affects the position information of the marker data, and adjusting the corresponding marker position information; The data-augmented image contains at least one complete labeled eye image; The final input image is scaled to a single-channel grayscale image of a specified size and then input into the deep neural network model to be trained after normalization. Marker data preprocessing, including: adjusting corresponding position information; After the location information is finally normalized, it is used as the new Ground Truth information and input into the network model to be trained; The candidate ellipse box is marked as ,in is the normalized center coordinate of the candidate ellipse frame, is the axis length of the normalized ellipse, For oval frame Axis and The clockwise angle of the axis.

8. The elliptical iris positioning method based on deep convolutional neural network according to claim 1, characterized in that: The method of training the deep neural network model using the training data set includes: Split the labeled iris images and data into two sets: training dataset and validation dataset; Use the training dataset for training and the validation dataset for deep neural network model validation and parameter screening; The iris image and the labeled data preprocessed in the training data set are input into the deep neural network, and the deep neural network performs forward reasoning on the input iris image to obtain the current reasoning result; Use the current inference result and the corresponding labeled data to calculate the classification loss and regression loss to obtain the current joint loss; The current loss is propagated to the deep neural network through the reverse gradient, and together with the learning rate parameter, the current deep neural network weight is adjusted; Repeat until the training is completed; The conditions for the end of training are: the specified rounds of training are completed; In the process of the deep neural network performing forward reasoning on the input iris image, the core lies in gradually extracting discriminative features from the iris image through multi-level feature extraction and information conversion, and finally generating reasoning results; First, decode the inference results; The decoded data is processed through non-maximum suppression to filter out the best matching prediction box; The filtered prediction box is denormalized to obtain the ellipse parameter information of the prediction box.

9. An elliptical iris positioning system based on deep convolutional neural network, characterized in that: include: A labeling module is used to label the training data set to obtain the truth, where the training data set contains the iris images to be trained, and the truth contains the ellipse parameter information of the ellipse frame in the training data set; Building blocks for building deep neural network models; A training module, used to train a deep neural network model using a training dataset; The inference module is used to use the trained deep neural network model to perform forward inference on the iris image to be processed to obtain an inference result; the inference result includes the ellipse parameter information of the prediction box in the iris image; The post-processing module is used to post-process the inference results and parse the ellipse parameter information corresponding to the iris and pupil boundary in the iris image according to the position information of the prediction box.

10. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 8.

Images