Model generation method, label marking method, iris image quality evaluation method and device
Generating an iris image quality evaluation model through deep convolutional neural network, solving the accuracy of iris image quality evaluation of different devices, realizing accurate quality evaluation of images collected by multiple devices, and improving the stability and accuracy of the recognition results .
Patent Information
- Application Number
- CN202010222328.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-03-26
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2040-03-26
AI Technical Summary
The prior art is difficult to accurately evaluate the quality of iris images acquired by different devices, resulting in bias and inaccuracy of identification results.
Deep convolutional neural network is used to generate an iris image quality evaluation model. By obtaining the training sample sets collected by multiple devices, selecting samples with a difference greater than the threshold, updating the label, iterating the model until the difference between the model output and the label is smaller than the threshold, and generating an accurate iris Image quality evaluation model.
Accurate quality evaluation of iris images collected by a variety of different devices is achieved, reducing the workload of manual labeling, and improving the accuracy and stability of recognition.
Smart Images

Figure CN111553880B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of iris recognition, and in particular to a model generation method, a label marking method, and an iris image quality evaluation method and device. Background Art
[0002] The uniqueness, stability, non-contactness, and anti-counterfeiting of the iris give iris recognition technology an outstanding advantage in biometric identification. With the continuous development of iris recognition technology and the widespread application of iris recognition products in different scenarios, the requirements for the accuracy and stability of iris recognition are also constantly increasing. The iris recognition process is generally divided into five steps: iris image acquisition, image quality judgment, iris segmentation, feature extraction, and sample comparison. Iris image quality judgment is an important prerequisite for ensuring the effectiveness of iris recognition. By evaluating the quality of iris images and selecting images with good quality for recognition, the error rate of iris recognition can be effectively reduced. Therefore, how to accurately evaluate the quality of iris images has become a key issue that needs to be solved. Summary of the invention
[0003] The present invention provides a model generation method, a label marking method, an iris image quality evaluation method and a device to accurately evaluate the quality of an iris image.
[0004] In order to achieve the above object, the present invention adopts the following scheme:
[0005] According to one aspect of an embodiment of the present invention, a method for generating an iris image quality assessment model is provided, comprising:
[0006] Acquire a training sample set, wherein each training sample in the training sample set includes an iris image and a corresponding quality label, and the iris images in the training samples in the training sample set are taken by iris acquisition devices of different purposes and correspond to different parameter values of the same number of types of iris quality parameters, and the number of types of iris quality parameters is one or more;
[0007] Inputting each training sample in the training sample set into an initial quality assessment model respectively, obtaining a quality assessment result of the iris image of each training sample, and calculating the difference between the quality assessment result of the iris image of each training sample and its quality label; wherein the initial quality assessment model belongs to a deep convolutional neural network;
[0008] Selecting at least one training sample from the training sample set, wherein the difference between the current quality evaluation result of the iris image of the selected training sample and its quality label is greater than a set difference threshold;
[0009] receiving updated quality labels of each selected training sample to update the corresponding training sample, thereby updating the training sample set; and inputting each training sample in the updated training sample set into the current initial quality evaluation model to obtain the quality evaluation result of the iris image of each training sample in the updated training sample set, and calculating the difference between the current quality evaluation result of the iris image of each training sample in the updated training sample set and its current quality label;
[0010] When the differences between the current quality evaluation results of the iris images of each training sample in the updated training sample set and their current quality labels are less than or equal to the set difference threshold, the iris image quality evaluation model is obtained according to the latest initial quality evaluation model.
[0011] According to another aspect of an embodiment of the present invention, a method for marking an iris image quality label is provided, comprising:
[0012] Acquire a training sample set, wherein each training sample in the training sample set includes an iris image and a corresponding quality label, and the iris images in the training samples in the training sample set are taken by iris acquisition devices of different purposes and correspond to different parameter values of the same number of types of iris quality parameters, and the number of types of iris quality parameters is one or more;
[0013] Inputting each training sample in the training sample set into an initial quality assessment model respectively, obtaining a quality assessment result of the iris image of each training sample, and calculating the difference between the quality assessment result of the iris image of each training sample and its quality label; wherein the initial quality assessment model belongs to a deep convolutional neural network;
[0014] Selecting at least one training sample from the training sample set, wherein the difference between the current quality evaluation result of the iris image of the selected training sample and its quality label is greater than a set difference threshold;
[0015] receiving updated quality labels of each selected training sample to update the corresponding training sample, thereby updating the training sample set; and inputting each training sample in the updated training sample set into the current initial quality evaluation model to obtain the quality evaluation result of the iris image of each training sample in the updated training sample set, and calculating the difference between the current quality evaluation result of the iris image of each training sample in the updated training sample set and its current quality label;
[0016] When the differences between the current quality evaluation results of the iris images of each training sample in the updated training sample set and their current quality labels are less than or equal to the set difference threshold, the quality label annotated for the corresponding iris image is determined according to the latest quality label in the training sample.
[0017] In some embodiments, obtaining a training sample set includes:
[0018] Acquire a plurality of different iris images, wherein the plurality of different iris images are taken by iris acquisition devices of different purposes and correspond to different parameter values of the same number of types of iris quality parameters, where the number of types of iris quality parameters is one or more;
[0019] The quality label of each iris image is determined according to the parameter value of each iris quality parameter, and the obtained quality label is associated with the corresponding iris image to obtain a training sample. The training samples corresponding to all iris images form a training sample set.
[0020] In some embodiments, the iris quality parameters include one or more of occlusion level, clarity, brightness, strabismus or emmetropia, and pupil dilation rate.
[0021] In some embodiments, the iris quality parameters include occlusion level, clarity, brightness, strabismus or emmetropia, and pupil dilation rate;
[0022] The quality label of each iris image is determined according to the parameter values of each iris quality parameter, including:
[0023] Calculating a matching score between each acquired iris image and a standard iris image of the corresponding iris, calculating a brightness score of each acquired iris image, and calculating a clarity score of each acquired iris image; wherein the standard iris image meets a set clarity requirement, a set occlusion degree requirement, a set brightness requirement, and an iris emmetropia requirement;
[0024] The quality label is calculated based on the matching score, brightness score and clarity score of each iris image obtained.
[0025] In some embodiments, acquiring a plurality of different iris images includes:
[0026] Acquire a plurality of different initial images, wherein the plurality of different initial images are taken by iris acquisition devices of different purposes and correspond to different parameter values of the same number of iris quality parameters, where the number of the types of iris quality parameters is one or more;
[0027] The iris position in each of the initial images is detected, and an image of a set size is deducted with the detected iris position as the center to obtain an iris image, wherein each iris image contains data of one iris.
[0028] In some embodiments, the quality label is calculated based on the matching score, brightness score and clarity score of each acquired iris image, including:
[0029] Substitute the matching score, brightness score and clarity score of the iris image into the formula q=min(100,max(0,q1-abs(q2-128) / 255*100+q3 / 2)) to calculate the quality label, where q represents the quality label, q1 represents the matching score between the iris image and the standard iris image of the corresponding iris, q2 represents the brightness score of the iris image, q3 represents the clarity score of the iris image, and abs(q2-128) represents the absolute value of q2-128.
[0030] In some embodiments, calculating the matching score between each acquired iris image and a standard iris image of the corresponding iris includes:
[0031] Calculate the scores obtained by matching each acquired iris image with a plurality of standard iris images of the corresponding iris, wherein different standard iris images have different pupil expansion and contraction rates;
[0032] The scores obtained by matching multiple standard iris images with the corresponding iris are averaged to obtain a matching score corresponding to each acquired iris image.
[0033] In some embodiments, calculating the brightness score of each acquired iris image includes:
[0034] The brightness mean of each pixel of each iris image is calculated, and the brightness score of the corresponding iris image is obtained according to the brightness mean of each pixel of each iris image.
[0035] In some embodiments, calculating the clarity score of each acquired iris image includes:
[0036] The overall clarity of each iris image is counted using the Laplacian of Gaussian kernel, and a clarity score is obtained based on the overall clarity of each iris image.
[0037] In some embodiments, selecting at least one training sample from the training sample set includes:
[0038] All training samples whose differences between the current quality evaluation results and their quality labels are greater than a set difference threshold are selected from the training sample set.
[0039] In some embodiments, each quality assessment result and each quality label belong to the same score range.
[0040] In some embodiments, the updated quality label of each selected training sample is determined according to its current quality evaluation result.
[0041] In some embodiments, the deep convolutional neural network includes an input layer, at least two convolutional pooling units, and a fully connected convolutional layer; each of the convolutional pooling units includes a convolutional layer, a BatchNorm layer, an activation layer, and a pooling layer.
[0042] In some embodiments, before inputting each training sample in the training sample set into the initial quality evaluation model to obtain the quality evaluation result of the iris image of each training sample, the method further includes:
[0043] A deep convolutional neural network for classification is trained using a public data set, the output number of the trained deep convolutional neural network for classification is modified according to the quality evaluation method corresponding to the initial quality evaluation model, and the trained deep convolutional neural network for classification after the output number is modified is fine-tuned using training samples containing iris images and corresponding quality labels to obtain an initial quality evaluation model.
[0044] In some embodiments, the deep convolutional neural network is a ResNet, VGG or AlexNet network.
[0045] In some embodiments, the method further comprises:
[0046] In a case where the difference between the current quality evaluation result and the current quality label of the iris image of at least part of the training samples in the updated training sample set is greater than the set difference threshold, selecting at least one training sample from the updated training sample set again, wherein the difference between the current quality evaluation result and the current quality label of the iris image of the reselected training sample is greater than the set difference threshold;
[0047] Receiving the updated quality labels of each reselected training sample to update the corresponding training sample, thereby updating the training sample set again; and inputting each training sample in the re-updated training sample set into the current initial quality evaluation model to obtain the quality evaluation result of the iris image of each training sample in the re-updated training sample set, and calculating the difference between the quality evaluation result of the iris image of each training sample in the re-updated training sample set and its current quality label;
[0048] When the differences between the quality evaluation results of the iris images of each training sample in the re-updated training sample set and their current quality labels are less than or equal to the set difference threshold, the iris image quality evaluation model is obtained according to the latest initial quality evaluation model.
[0049] According to another aspect of an embodiment of the present invention, there is provided a method for evaluating iris image quality, comprising:
[0050] Obtaining an iris image to be evaluated;
[0051] The quality of the iris image to be evaluated is evaluated by using the iris image quality evaluation model obtained by the iris image quality evaluation model generation method according to any one of the above embodiments.
[0052] According to another aspect of an embodiment of the present invention, there is provided an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for generating an iris image quality assessment model described in any one of the above embodiments or the steps of the method for assessing iris image quality described in any one of the above embodiments are implemented.
[0053] According to another aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the method for generating an iris image quality assessment model described in any of the above embodiments or the steps of the iris image quality assessment method described in any of the above embodiments is implemented.
[0054] The iris image quality assessment model generation method, iris image quality label marking method, iris image quality assessment method, electronic device and computer-readable storage medium of the embodiments of the present invention can facilitate accurate assessment of iris images collected by a variety of different devices. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:
[0056] Figure 1 is a flow chart of a method for generating an iris image quality assessment model according to an embodiment of the present invention;
[0057] Figure 2 is a flow chart of an iris image quality evaluation method according to a specific embodiment of the present invention;
[0058] Figure 3 is a schematic diagram of a process for obtaining a quality label in one embodiment of the present invention;
[0059] Figure 4 is a flow chart of a training model in a specific embodiment of the present invention;
[0060] Figure 5 is a schematic diagram of the structure of a deep convolutional neural network in a specific embodiment of the present invention;
[0061] Figure 6 The figure is a flow chart of a method for labeling iris image quality labels according to an embodiment of the present invention. DETAILED DESCRIPTION
[0062] To make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings. Here, the exemplary embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.
[0063] When evaluating the quality of iris images, the existing method is to calculate each score separately and finally weight them to give the total score of the iris image quality. However, since each indicator has a different impact on the iris recognition result, it is difficult to give a suitable weighting coefficient. In addition, some indicators are based on the results of iris segmentation (for example, the degree of occlusion). Once the segmentation is wrong, the judgment of the iris image quality will be biased. Moreover, the calculation results of the clarity of the iris image in the existing method may be inconsistent with the subjective judgment of the person. For example, the light spots of glasses and inappropriate light may cause this to happen.
[0064] In addition, the quality of iris acquisition devices for different purposes (such as iris images acquired by mobile devices and iris images acquired by security equipment) varies greatly, and the same standards cannot be used to evaluate the quality of iris images. Therefore, different iris image quality evaluation products are needed to evaluate iris images acquired by different devices, otherwise inaccurate quality evaluation results will be obtained.
[0065] In order to solve the above problems, an embodiment of the present invention provides a method for generating an iris image quality assessment model. The iris image quality assessment model generated by the method can be used to accurately evaluate iris images collected by different devices.
[0066] Figure 1 FIG. 1 is a flow chart of a method for generating an iris image quality assessment model according to an embodiment of the present invention. Figure 1 As shown, the method for generating an iris image quality assessment model in this embodiment may include the following steps S110 to S150.
[0067] The specific implementation of steps S110 to S150 will be described in detail below.
[0068] Step S110: Obtain a training sample set, wherein each training sample in the training sample set includes an iris image and a corresponding quality label, and the iris images in the training samples in the training sample set are taken by iris acquisition devices for different purposes and correspond to different parameter values of the same number of types of iris quality parameters, and the number of types of iris quality parameters is one or more.
[0069] In step S110, the training sample set may include a large number of training samples, such as tens of thousands. An iris image may only include one iris, which may be obtained from an image captured by an iris acquisition device. For a training sample, an iris image is associated with a quality label, which may be a quality evaluation of the iris image based on human senses, for example, which may be expressed in the form of a score, grade, etc. In addition, the specific scoring system (range) or number of grades may be determined as needed.
[0070] The training sample set contains a large number of iris images, which can be collected by a variety of different devices. For example, some iris images can be collected by one device (such as a mobile device), and some iris images can be collected by another device (such as a security device). Of course, some iris images can also be collected by other devices. In this way, these iris images can cover a variety of different devices.
[0071] In addition, for an iris image collected by a device, it can be evaluated by one or more types of iris quality parameters, and different iris images may have different parameter values of one or more types of iris quality parameters. Among them, the iris quality parameters may include one or more parameters such as occlusion degree, clarity, brightness, strabismus or emmetropia, and pupil dilation rate. Specifically, for example, the iris quality parameters may include parameters such as occlusion degree, clarity, brightness, strabismus or emmetropia, and pupil dilation rate. Then, for those iris images collected by a device, any two iris images may have at least one different parameter value of the iris quality parameter, for example, different clarity, or different clarity and brightness, or different occlusion degree, clarity and brightness. For the iris images collected by each device, various combinations of various parameter values of all iris quality parameters can be covered as much as possible, so that the quality distribution of iris images in the training sample set can be more uniform.
[0072] In the above step S110, obtaining the training sample set may be directly reading or receiving a prepared training sample set that meets the requirements; or, the required training sample set may be obtained by on-site processing.
[0073] Exemplarily, in step S110, obtaining a training sample set may specifically include the following steps: S111, obtaining a plurality of different iris images, wherein the plurality of different iris images are taken by iris acquisition devices of different purposes and correspond to different parameter values of the same number of iris quality parameters, and the number of types of iris quality parameters is one or more; S112, determining a quality label of each iris image according to the parameter value of each iris quality parameter, and associating the obtained quality label with the corresponding iris image to obtain a training sample, so that the training samples corresponding to all iris images form a training sample set.
[0074] In step S111, the prepared iris images that meet the requirements can be directly read or received, wherein the requirements can refer to that the iris images are taken by one or more devices, and the iris images taken by the same devices can cover the combination of various parameter values of various iris quality parameters as much as possible, covering various quality combinations. In step S112, the parameter values of various iris quality parameters can be manually calibrated for each iris image, and then the scores or values obtained can be calculated according to the set rules based on these parameter values as the quality labels, so that the quality labels can be obtained in an objective way.
[0075] Further, in some specific embodiments, the above-mentioned step S111, i.e., obtaining a plurality of different iris images, more specifically, may include the steps: S1111, obtaining a plurality of different initial images, wherein the plurality of different initial images are taken by iris acquisition devices of different purposes and correspond to different parameter values of the same number of iris quality parameters, and the number of types of iris quality parameters is one or more; S1112, detecting the iris position in each of the initial images, and deducting an image of a set size with the detected iris position as the center to obtain an iris image, wherein each iris image contains data of one iris.
[0076] In step S1111, the initial image may be an image of various qualities directly obtained by an iris acquisition device, which may include not only the iris but also other parts (such as the eyeball part outside the iris), and may include not only one iris but also multiple irises. Multiple different initial images cover various quality combinations, and of course, specific parameter values are not required to have been obtained. In step S1112, the iris in the initial image may be located using existing iris detection technology, for example, to find the center position of the iris, and then the iris may be deducted according to a set size so that one iris image contains one iris, wherein the set size may be determined based on the actual iris size, the image scaling ratio, etc.
[0077] Some existing iris image quality judgment methods can only give usable or unusable results. However, for devices with different purposes, the standards for usable and unusable are different. For example, the quality of images collected by mobile devices and security equipment is quite different, and the same standards cannot be used. As a result, different devices require different judgment methods. In order to better use the same standards to evaluate the quality of iris images collected by different devices, the quality label can be expressed in the form of scores.
[0078] In some specific embodiments, when the iris quality parameters include occlusion degree, clarity, brightness, strabismus or emmetropia, and pupil expansion / contraction rate, the above-mentioned step S112, i.e., determining the quality label of each iris image according to the parameter value of each iris quality parameter of each iris image, more specifically, may include the steps of: S1121, calculating the matching score of each acquired iris image with the standard iris image of the corresponding iris, calculating the brightness score of each acquired iris image, and calculating the clarity score of each acquired iris image; wherein the standard iris image meets the set clarity requirement, the set occlusion degree requirement, the set brightness requirement, and the iris emmetropia requirement; S1122, calculating the quality label according to the matching score, brightness score and clarity score of each acquired iris image.
[0079] For example, in the above step S1121, a matching score can be obtained by comparing an iris image with a standard image; or, a matching score component can be obtained by comparing an iris image with multiple standard images respectively, and then the matching score corresponding to the iris image can be obtained by combining the matching score components. The specific method of comparing an iris image with a standard image may include the steps of: detecting the position of the iris in the iris image and the standard image respectively; obtaining the iris inner circle and iris outer circle of the iris in the iris image and the standard image; performing feature extraction on the area between the iris inner circle and the iris outer circle for the iris image and the standard image respectively; performing feature comparison on the feature extraction result of the iris image and the feature extraction result of the standard image to obtain a score, and directly using the score as the matching score, or obtaining the matching score after processing (such as normalization).
[0080] For another example, in the above step S1121, the matching scores of each acquired iris image and the standard iris image of the corresponding iris are calculated, which may specifically include the steps of: calculating the scores obtained by matching each acquired iris image with multiple standard iris images of the corresponding iris; wherein different standard iris images have different pupil dilation rates; averaging the scores obtained by matching with multiple standard iris images of the corresponding iris, to obtain the matching score corresponding to each acquired iris image. In this way, the comparison result of the iris image and the standard image can take into account the influence of the pupil dilation rate.
[0081] In other embodiments, different standard iris images may have different parameter values in other iris quality parameters. In addition, the scores obtained by matching the standard iris images may be processed in other ways to reflect the comprehensive conditions of these standard iris images.
[0082] For example, in the above step S1121, calculating the brightness score of each acquired iris image may specifically include the steps of: calculating the brightness mean of each pixel of each acquired iris image, and obtaining the brightness score of the corresponding iris image according to the brightness mean of each pixel of each iris image.
[0083] For example, in the above step S1121, the clarity score of each acquired iris image is calculated, which may specifically include the steps of: using the Gaussian Laplace kernel to count the overall clarity of each iris image, and obtaining a clarity score according to the overall clarity of each iris image. The clarity of each acquired iris image may be counted with reference to the national standard 29794-6, and the overall clarity of each iris image may be counted using the Gaussian Laplace kernel. The overall clarity may be directly used as the clarity score, or the clarity score may be obtained after processing.
[0084] In addition, in the above step S1121, the definition requirement may be set to be above the definition (score), the occlusion requirement may be set to be below the occlusion ratio, the brightness requirement may be set to be within the brightness (score) range, and the iris emmetropia requirement means that the iris is emmetropia rather than strabismus in the standard iris image. Alternatively, the above various requirements (the definition requirement, the occlusion requirement, the brightness requirement, and the iris emmetropia requirement) may be the results of sensory evaluation by the human eye or statistical results.
[0085] In the above step S1122, the directly obtained matching score, brightness score and clarity score may conform to the same scoring system; or, when it is necessary to calculate the quality label, they may be unified into the same scoring system. The calculated quality label and the matching score, brightness score and clarity score or the unified scoring system may have the same scoring system, for example, all within the range of 0 to 100, or the calculated quality label may be converted to the same scoring system. After the matching score, brightness score and clarity score of each iris image are obtained, the quality label may be calculated according to a set rule or a set calculation formula.
[0086] For example, the above-mentioned step S1122, i.e., calculating the quality label according to the obtained matching score, brightness score and clarity score of each iris image, may specifically include the steps of: substituting the matching score, brightness score and clarity score of the iris image into the formula q=min(100,max(0,q1-abs(q2-128) / 255*100+q3 / 2)), and calculating the quality label, wherein q represents the quality label, q1 represents the matching score between the iris image and the standard iris image of the corresponding iris, q2 represents the brightness score of the iris image, q3 represents the clarity score of the iris image, and abs(q2-128) represents the absolute value of q2-128.
[0087] Among them, the formula q=min(100,max(0,q1-abs(q2-128) / 255*100+q3 / 2)) can limit the calculated quality label to the range of 0 to 100. In this formula, 128 represents the grayscale value of the grayscale image, and 255 represents the grayscale value of white. In addition, the matching score, brightness score and clarity score of the iris image can comply with the unified scoring system.
[0088] Step S120: Input each training sample in the training sample set into the initial quality evaluation model respectively, obtain the quality evaluation result of the iris image of each training sample, and calculate the difference between the quality evaluation result of the iris image of each training sample and its quality label; wherein the initial quality evaluation model belongs to a deep convolutional neural network.
[0089] In step S120, the deep convolutional neural network may include an input layer, at least two convolutional pooling units, and a fully connected convolutional layer; each of the convolutional pooling units includes a convolutional layer, a BatchNorm layer, an activation layer, and a pooling layer. Alternatively, the deep convolutional neural network may be an existing network, for example, a ResNet, VGG or AlexNet network. Each quality evaluation result and each quality label may belong to the same score range, for example, a range of 0 to 100. After the training sample is input into the initial quality evaluation model, the model can output the result of this training after calculation.
[0090] After the training samples (iris images and quality labels) are input into the initial quality assessment model, the initial quality assessment model of the deep convolutional neural network can be trained to obtain quality assessment results that are as consistent as possible with most quality labels. If the difference between the quality assessment results and labels corresponding to one or some training samples is larger than the difference between the quality assessment results and labels corresponding to other training samples, it is possible that the quality label setting is unreasonable. If it is unreasonable, the adjusted quality label can be received through the subsequent step S140.
[0091] By using a deep convolutional network as the basis for iris image quality evaluation, a better evaluation effect can be achieved. Moreover, the deep convolutional network can easily achieve the purpose of not needing to calculate each score of image quality separately, and can give a total score of image quality, thus avoiding the problem of difficult weight selection.
[0092] In order to reduce the number of training samples that need to be prepared, or to achieve better training results when the number of training samples is small, existing data sets can be used for pre-training.
[0093] Exemplarily, before step S120, that is, before inputting each training sample in the training sample set into the initial quality evaluation model to obtain the quality evaluation result of the iris image of each training sample, Figure 1 The method shown may also include the step: S160, using a public data set to train a deep convolutional neural network for classification, modifying the output number of the trained deep convolutional neural network for classification according to the quality evaluation method corresponding to the initial quality evaluation model, and using training samples containing iris images and corresponding quality labels to fine-tune the trained deep convolutional neural network for classification after the output number is modified to obtain an initial quality evaluation model.
[0094] Among them, the public data set may refer to a currently known data set, for example, the ImageNet data set. When the quality evaluation method (such as scoring) corresponding to the initial quality evaluation model is different from the method of the deep convolutional neural network used for classification, the initial quality evaluation model suitable for this solution can be obtained by appropriately modifying the scoring method. For example, when the quality evaluation method corresponding to the initial quality evaluation model is a scoring method, the number of outputs of the trained deep convolutional neural network for classification can be modified to 1, and at the same time, the loss layer of the trained deep convolutional neural network for classification can be modified to an algorithm for calculating the distance; if the quality evaluation method is multiple levels, the number of outputs of the trained deep convolutional neural network for classification can be modified to the number of levels, and the loss layer can be modified to other algorithms accordingly; then, the prepared training samples can be used for correction or training through step S120, etc. A small amount of training (such as once) is performed on the modified network using training samples similar to the embodiment of the present invention to fine-tune the initial quality evaluation model that can be used in the embodiment of the present invention.
[0095] Step S130: selecting at least one training sample from the training sample set, wherein the difference between the current quality evaluation result of the iris image of the selected training sample and its quality label is greater than a set difference threshold.
[0096] In step S130, at least one training sample that meets the requirements is selected from the training sample set, and a training sample with a large difference between the currently obtained quality evaluation result and its current quality label can be selected. After selecting the training sample, it can be manually determined whether the current quality label needs to be adjusted. If necessary, the quality label can be manually adjusted, and then the adjusted quality label and the original corresponding iris image form an updated training sample. The quality label can be adjusted according to its current quality evaluation result. For example, if the quality label needs to be adjusted, the current quality evaluation result or the result calculated or determined based on it can replace the original quality label.
[0097] Exemplarily, in step S130, selecting at least one training sample from the training sample set may specifically include: selecting all training samples from the training sample set whose difference between the current quality evaluation result and its quality label is greater than a set difference threshold, so as to speed up the model training.
[0098] Step S140: receiving updated quality labels of each selected training sample to update the corresponding training sample, thereby updating the training sample set; and inputting each training sample in the updated training sample set into the current initial quality evaluation model to obtain the quality evaluation result of the iris image of each training sample in the updated training sample set, and calculating the difference between the current quality evaluation result of the iris image of each training sample in the updated training sample set and its current quality label.
[0099] In step S140, the quality label can be manually adjusted or updated, and the adjusted or updated quality label can be input through a human-computer interaction device. After the quality label is updated, the corresponding training sample is updated, and the training sample set is updated. The updated training sample set can include the training samples updated this time and the training samples not updated this time. All training samples in the updated training sample set can be input into the current initial quality evaluation model for training, and the quality evaluation results corresponding to each training sample can be obtained.
[0100] All training samples in the updated training sample set can be respectively input into the previously trained initial quality evaluation model for retraining, and new differences can be obtained.
[0101] Step S150: when the differences between the current quality evaluation results of the iris images of each training sample in the updated training sample set and their current quality labels are less than or equal to the set difference threshold, an iris image quality evaluation model is obtained according to the latest initial quality evaluation model.
[0102] In step S150, when the initial quality result and the quality label belong to the same range (such as the same scoring system), the sizes of the two can be directly compared. For example, if the difference between the initial quality result and the quality label is within the set range (set difference threshold), the model can be considered to be trained. Each time all samples in the training sample set are sent to the initial quality evaluation model, after multiple iterations of deep convolutional neural network learning, the training loss no longer decreases significantly, and the training is completed. The parameters in the quality evaluation result can be adaptively adjusted in one training. Therefore, the "current initial quality evaluation model" described in this article can refer to the initial quality evaluation model after the most recent training, the "current quality evaluation result" can refer to the quality evaluation result obtained after the most recent training, and the "current quality label" can refer to the original quality label before the first training is completed, and can refer to the most recently updated quality label afterwards, wherein the "most recent" can be a time point when a certain model or label executes a certain step, or a time point when a certain quality evaluation result executes a certain step.
[0103] In some other embodiments, when the set difference threshold condition is not met, a method similar to the above steps S130 to S140 may be used to iteratively perform training until the difference between the quality evaluation result and the quality label meets the requirement. One or more trainings may be performed before or after the above steps S130 to S140, that is, multiple groups of steps similar to the above steps S130 to S140 may be included. Exemplarily, Figure 1 The method shown may further include the steps of:
[0104] S170, when the difference between the current quality evaluation result of the iris image of at least part of the training samples in the updated training sample set and its current quality label is greater than the set difference threshold, select at least one training sample from the updated training sample set again, wherein the difference between the current quality evaluation result of the iris image of the re-selected training sample and its current quality label is greater than the set difference threshold;
[0105] S180, receiving the updated quality labels of each re-selected training sample to update the corresponding training sample, thereby updating the training sample set again; and inputting each training sample in the re-updated training sample set into the current initial quality evaluation model to obtain the quality evaluation result of the iris image of each training sample in the re-updated training sample set, and calculating the difference between the quality evaluation result of the iris image of each training sample in the re-updated training sample set and its current quality label;
[0106] S190, when the differences between the quality evaluation results of the iris images of each training sample in the training sample set updated again and its current quality label are all less than or equal to the set difference threshold, obtain an iris image quality evaluation model according to the latest initial quality evaluation model.
[0107] In the above step S170, the training samples corresponding to the large differences that do not meet the requirements are selected, and the quality labels of these training samples can be adjusted. In addition, if the difference between the current quality evaluation result of the iris image of each training sample in the training sample set updated again and its current quality label is greater than the set difference threshold, the training can continue.
[0108] In addition, embodiments of the present invention further provide a method for evaluating iris image quality. The method for evaluating iris image quality in these embodiments may include the following steps:
[0109] S210, obtaining an iris image to be evaluated;
[0110] S220: Use the iris image quality assessment model obtained by the iris image quality assessment model generation method according to any one of the above embodiments to perform quality assessment on the iris image to be assessed.
[0111] In step S210, the iris image to be evaluated can be an iris image collected by various devices. In step S220, the iris image quality evaluation model generation method can be implemented with reference to the iris image quality evaluation model generation method of the above embodiments, which will not be described in detail.
[0112] In addition, based on Figure 1 With the same inventive concept as the method for generating an iris image quality evaluation model shown in the figure, an embodiment of the present invention further provides a method for marking an iris image quality label. Figure 6 FIG. 1 is a flow chart of a method for labeling iris image quality according to an embodiment of the present invention. Figure 6 As shown, the iris image quality labeling method of these embodiments may include:
[0113] Step S310: obtaining a training sample set, wherein each training sample in the training sample set includes an iris image and a corresponding quality label, and the iris images in the training samples in the training sample set are taken by iris acquisition devices of different purposes and correspond to different parameter values of the same number of iris quality parameters, and the number of the types of iris quality parameters is one or more;
[0114] Step S320: inputting each training sample in the training sample set into an initial quality assessment model to obtain a quality assessment result of the iris image of each training sample, and calculating the difference between the quality assessment result of the iris image of each training sample and its quality label; wherein the initial quality assessment model belongs to a deep convolutional neural network;
[0115] Step S330: selecting at least one training sample from the training sample set, wherein the difference between the current quality evaluation result of the iris image of the selected training sample and its quality label is greater than a set difference threshold;
[0116] Step S340: receiving the updated quality labels of each selected training sample to update the corresponding training sample, thereby updating the training sample set; and inputting each training sample in the updated training sample set into the current initial quality evaluation model to obtain the quality evaluation result of the iris image of each training sample in the updated training sample set, and calculating the difference between the current quality evaluation result of the iris image of each training sample in the updated training sample set and its current quality label;
[0117] Step S350: When the differences between the current quality evaluation results of the iris images of each training sample in the updated training sample set and their current quality labels are less than or equal to the set difference threshold, determine the quality label for the corresponding iris image based on the latest quality label in the training sample.
[0118] The specific implementation methods of the above steps S310 to S340 can be respectively performed with reference to the specific implementation methods of the above steps S110 to S140, and will not be described in detail. In the above step S350, after the difference meets the requirements, the final quality label in the training sample can be determined as the quality label annotated by the corresponding iris image, or the final quality label in the training sample can be processed as required and then determined as the quality label annotated by the corresponding iris image. The final annotated quality label and the corresponding iris image can be stored in association and can be used for other purposes, for example, training other models, such as other iris image quality evaluation models. Since the quality label is updated and has a better accuracy, it can enable other models to obtain better training effects, and for other iris image quality evaluation models, better quality evaluation results can be obtained.
[0119] In addition, based on the same inventive concept as the method of the above embodiment, the embodiment of the present invention further provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the iris image quality assessment model generation method described in any of the above embodiments or the steps of the iris image quality assessment method described in any of the above embodiments when executing the program. The electronic device for implementing the iris image quality assessment model generation method may be a computer, a server, etc. The electronic device for implementing the iris image quality assessment method may be a computer, a server, an iris acquisition device, an iris recognition device, etc.
[0120] In addition, an embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the method for generating an iris image quality assessment model described in any of the above embodiments or the steps of the iris image quality assessment method described in any of the above embodiments are implemented.
[0121] In order to enable those skilled in the art to better understand the present invention, the implementation modes of the present invention will be described below with reference to specific embodiments.
[0122] In a specific embodiment, the iris image quality evaluation method is implemented based on a deep convolutional network. The deep convolutional network can achieve excellent results in image recognition. By using a deep convolutional network, it is possible to avoid calculating each score of image quality separately and directly obtain the total score of image quality.
[0123] Figure 2 FIG. 1 is a flow chart of a method for evaluating iris image quality according to a specific embodiment of the present invention. Figure 2 The iris image quality evaluation method of this embodiment includes a model training process and a model application process. The model training process may include: data acquisition, data labeling (labeling quality labels), and model training and label updating.
[0124] Data acquisition:
[0125] In order to make the trained model more robust and universal, iris images of different qualities can be obtained from a variety of devices with different uses, including mobile devices with poor quality and security devices with good quality. The distribution range of iris image quality can be very wide, for example, including one or more combinations of different occlusions, different sharpness, different brightness, different pupil expansion and contraction rates, strabismus and emmetropia.
[0126] After obtaining an iris image (such as a human eye image), the iris position can be detected by iris detection technology, and then an image of, for example, 224*224 size is deducted with the iris position as the center to obtain an iris image containing data of only one iris.
[0127] Based on the quality of the iris image, a quality label is given to each iris image. The quality label may include three parts: comparison score (matching score), brightness score, and clarity score. Figure 3 shown.
[0128] 1. Comparison score
[0129] If an iris image is compared with a standard iris image of the corresponding iris (such as unobstructed, clear, straight, and properly lit), the comparison score obtained can basically represent the quality of the iris image. In this way, the comparison score can be linked to the quality of the iris image.
[0130] Specifically, 3 to 5 iris images with good quality are selected for each iris as standard iris images. Standard iris images, for example, meet the requirements of clarity, less than 10% occlusion, normal vision, and appropriate lighting. At the same time, different standard iris images differ in pupil expansion rate. Other iris images (of various qualities) of the iris are compared with the standard iris image, and the average of the comparison scores of each iris image compared with multiple standard iris images is taken as the final comparison score of the iris image, which can be expressed as q1.
[0131] 2. Brightness score
[0132] Due to the diversity of iris textures, irises with different texture richness have different amounts of available information on the image under the same lighting conditions, especially when the lighting is too bright or too dark. In order to prevent the comparison score from not fully representing the image quality when the lighting is not appropriate, the overall brightness item of the statistical image is added, which can be expressed as q2, and the calculation method can be to calculate the mean brightness of the entire image.
[0133] 3. Clarity score
[0134] Also, due to the different richness of iris texture, the comparison score with the standard image may not be sufficient to express the clarity of the image. Here, we can refer to the statistics of clarity in national standard 29794-6 and use the Gaussian Laplace kernel to calculate the overall clarity of the image. The clarity can be expressed as q3.
[0135] After obtaining the comparison score q1, brightness score q2, and clarity score q3, the quality label q of the iris image can be calculated according to the formula q=min(100,max(0,q1-abs(q2-128) / 255*100+q3 / 2)). Then, the quality label can be associated with the iris image and stored, corresponding to a training sample.
[0136] Next, the training samples can be used to train the initial quality assessment model and update the quality labels.
[0137] The iris image quality assessment model can be trained based on an initial quality assessment model. The initial quality assessment model can belong to a deep convolutional neural network. Figure 5 As shown, the network structure of the model may include: a connected input layer, multiple convolutional pooling units and residual calculation, and a fully connected layer. Among them, the convolutional pooling unit may include: a connected convolutional layer, a BatchNorm layer, an activation layer (Relu), and a pooling layer. Usually, 5 to 10 convolutional pooling units can be used for training. The number of convolutional pooling units can be selected according to the number of iris images used for training (the number of training samples) and the size of the iris images. If the number of iris images is small and the size of the iris images is slightly larger, it can be appropriately adjusted to use fewer units for training. For example, 8 pooling units can be used. In addition, if there are few available training samples, a classification network can be trained using a public data set (such as the ImageNet data set), and then. Fine-tune the quality judgment network based on this classification network. Among them, the network used can be a resnet, VGG, AlexNet and other networks, so that a better quality evaluation model can be obtained.
[0138] An iterative method is used to update the quality label, that is, first use the above iris images and quality labels to train the initial quality assessment model, then use the model training effect (prediction results or quality assessment results) to screen out iris images whose model quality assessment results are significantly different from the quality labels. Manually determine and modify the quality labels, and then retrain the model until the model prediction results of all iris images are close to the quality labels.
[0139] Since the quality labels are generated according to established rules, the convergence of the model may change from a large oscillation at the beginning to a gradual smoothing process during the gradual manual calibration process. The oscillation of the model may be caused by inappropriate image labels. As the quality labels are gradually corrected, the convergence of the model is getting better and better. For example, the correction of the quality labels can be completed after 5 times.
[0140] The quality score of the iris image can range from 0 to 100. If the difference between the model output score and the label of each image exceeds 5 points, it is manually determined whether the label needs to be adjusted. After the adjustment is completed, all images are sent to the model training and the above steps are repeated until the difference between the model score and the label score is within 5 points. Then the training is stopped and a usable model is obtained. The process is as follows: Figure 4 shown.
[0141] By training and updating labels through iterative methods, it is no longer necessary to manually label the quality labels of all iris images, thus reducing the workload. At the same time, when correcting the quality labels, only images with large differences between the model output results and the label results can be manually confirmed.
[0142] In this embodiment, the training data includes images collected by multiple devices, and the image quality range is widely distributed. When using, different thresholds can be selected according to the actual situation of the device. Specifically, different image quality thresholds can be selected to filter images according to the requirements of different devices for recognition performance, so as to balance the proportion of images filtered out by the quality threshold and the image recognition performance that passes the evaluation, and select the appropriate image quality threshold for the device. The model can be used by multiple devices. The label generation process can ensure less manual workload while ensuring accuracy, and can achieve efficient iris image quality model learning.
[0143] In summary, the iris image quality assessment model generation method, iris image quality labeling method, iris image quality assessment method, electronic device and computer-readable storage medium of the embodiments of the present invention can make the model available for a variety of iris acquisition devices by using training samples of a training sample set containing iris images acquired by devices with different purposes. Moreover, by iteratively training the model and updating the quality labels according to the model output results, it is possible to avoid correcting all quality labels, thereby reducing the workload. In addition, by adopting a deep convolutional network, a more accurate iris image quality assessment result can be obtained. Therefore, the model training of the embodiments of the present invention is more convenient, and the evaluation method of the embodiments of the present invention can accurately evaluate iris images acquired by a variety of different devices.
[0144] In the description of this specification, the description with reference to the terms "one embodiment", "a specific embodiment", "some embodiments", "for example", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. The order of steps involved in each embodiment is used to schematically illustrate the implementation of the present invention, and the order of steps is not limited and can be appropriately adjusted as needed.
[0145] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0146] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0147] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0148] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.
[0149] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for generating an iris image quality assessment model, characterized in that: include: Acquire a training sample set, wherein each training sample in the training sample set includes an iris image and a corresponding quality label, and the iris images in the training samples in the training sample set are taken by iris acquisition devices of different purposes and correspond to different parameter values of the same number of types of iris quality parameters, and the number of types of iris quality parameters is one or more; Inputting each training sample in the training sample set into an initial quality assessment model respectively, obtaining a quality assessment result of the iris image of each training sample, and calculating the difference between the quality assessment result of the iris image of each training sample and its quality label; wherein the initial quality assessment model belongs to a deep convolutional neural network; Selecting at least one training sample from the training sample set, wherein the difference between the current quality evaluation result of the iris image of the selected training sample and its quality label is greater than a set difference threshold, wherein the set difference threshold is a threshold suitable for the iris acquisition device corresponding to the current training sample; Receive the updated quality labels of each selected training sample to update the corresponding training sample, thereby updating the training sample set; and input each training sample in the updated training sample set into the current initial quality evaluation model to obtain the quality evaluation result of the iris image of each training sample in the updated training sample set, and calculate the difference between the current quality evaluation result of the iris image of each training sample in the updated training sample set and its current quality label; wherein the updated quality label of each selected training sample is determined according to its current quality evaluation result; When the differences between the current quality evaluation results of the iris images of each training sample in the updated training sample set and their current quality labels are all less than or equal to the set difference threshold, obtaining an iris image quality evaluation model according to the latest initial quality evaluation model; Wherein, the quality label is obtained by: Determine a quality label of each iris image according to the parameter value of each iris quality parameter; The step of determining the quality label of each iris image according to the parameter value of each iris quality parameter of each iris image comprises: Calculating a matching score between each acquired iris image and a standard iris image of the corresponding iris, calculating a brightness score of each acquired iris image, and calculating a clarity score of each acquired iris image; wherein the standard iris image meets a set clarity requirement, a set occlusion degree requirement, a set brightness requirement, and an iris emmetropia requirement; The quality label is calculated based on the matching score, brightness score and clarity score of each iris image obtained.
2. A method for labeling iris image quality, characterized in that: include: Acquire a training sample set, wherein each training sample in the training sample set includes an iris image and a corresponding quality label, and the iris images in the training samples in the training sample set are taken by iris acquisition devices of different purposes and correspond to different parameter values of the same number of types of iris quality parameters, and the number of types of iris quality parameters is one or more; Inputting each training sample in the training sample set into an initial quality assessment model respectively, obtaining a quality assessment result of the iris image of each training sample, and calculating the difference between the quality assessment result of the iris image of each training sample and its quality label; wherein the initial quality assessment model belongs to a deep convolutional neural network; Selecting at least one training sample from the training sample set, wherein the difference between the current quality evaluation result of the iris image of the selected training sample and its quality label is greater than a set difference threshold, wherein the set difference threshold is a threshold suitable for the iris acquisition device corresponding to the current training sample; Receive the updated quality labels of each selected training sample to update the corresponding training sample, thereby updating the training sample set; and input each training sample in the updated training sample set into the current initial quality evaluation model to obtain the quality evaluation result of the iris image of each training sample in the updated training sample set, and calculate the difference between the current quality evaluation result of the iris image of each training sample in the updated training sample set and its current quality label; wherein the updated quality label of each selected training sample is determined according to its current quality evaluation result; When the differences between the current quality evaluation results of the iris images of each training sample in the updated training sample set and its current quality label are less than or equal to the set difference threshold, determining the quality label annotated for the corresponding iris image according to the latest quality label in the training sample; Wherein, the quality label is obtained by: Determine a quality label of each iris image according to the parameter value of each iris quality parameter; The step of determining the quality label of each iris image according to the parameter value of each iris quality parameter of each iris image comprises: Calculating a matching score between each acquired iris image and a standard iris image of the corresponding iris, calculating a brightness score of each acquired iris image, and calculating a clarity score of each acquired iris image; wherein the standard iris image meets a set clarity requirement, a set occlusion degree requirement, a set brightness requirement, and an iris emmetropia requirement; The quality label is calculated based on the matching score, brightness score and clarity score of each iris image obtained.
3. The method according to claim 1 or 2, characterized in that Get the training sample set, including: Acquire a plurality of different iris images, wherein the plurality of different iris images are taken by iris acquisition devices of different purposes and correspond to different parameter values of the same number of types of iris quality parameters, where the number of types of iris quality parameters is one or more; The quality label of each iris image is determined according to the parameter value of each iris quality parameter, and the obtained quality label is associated with the corresponding iris image to obtain a training sample. The training samples corresponding to all iris images form a training sample set.
4. The method according to claim 3, characterized in that The iris quality parameters include one or more of occlusion level, clarity, brightness, strabismus or emmetropia, and pupil dilation rate.
5. The method according to claim 4, characterized in that Iris quality parameters include occlusion, clarity, brightness, strabismus or emmetropia, and pupil dilation rate.
6. The method according to claim 3, characterized in that Get multiple different iris images, including: Acquire a plurality of different initial images, wherein the plurality of different initial images are taken by iris acquisition devices of different purposes and correspond to different parameter values of the same number of iris quality parameters, where the number of the types of iris quality parameters is one or more; The iris position in each of the initial images is detected, and an image of a set size is deducted with the detected iris position as the center to obtain an iris image, wherein each iris image contains data of one iris.
7. The method according to claim 1, characterized in that The quality label is calculated based on the matching score, brightness score and clarity score of each iris image, including: Substitute the matching score, brightness score and clarity score of the iris image into the formula q=min(100,max(0,q1-abs(q2-128) / 255*100+q3 / 2)) to calculate the quality label, where q represents the quality label, q1 represents the matching score between the iris image and the standard iris image of the corresponding iris, q2 represents the brightness score of the iris image, q3 represents the clarity score of the iris image, and abs(q2-128) represents the absolute value of q2-128.
8. The method according to claim 1, characterized in that Calculating the matching score between each acquired iris image and the standard iris image of the corresponding iris, including: Calculate the scores obtained by matching each acquired iris image with a plurality of standard iris images of the corresponding iris, wherein different standard iris images have different pupil expansion and contraction rates; The scores obtained by matching multiple standard iris images with the corresponding iris are averaged to obtain a matching score corresponding to each acquired iris image.
9. The method according to claim 1, characterized in that Calculate the brightness score of each acquired iris image, including: The brightness mean of each pixel of each iris image is calculated, and the brightness score of the corresponding iris image is obtained according to the brightness mean of each pixel of each iris image.
10. The method according to claim 1, characterized in that Calculate the clarity score of each acquired iris image, including: The overall clarity of each iris image is counted using the Laplacian of Gaussian kernel, and a clarity score is obtained based on the overall clarity of each iris image.
11. The method according to claim 1 or 2, characterized in that: Selecting at least one training sample from the training sample set includes: All training samples whose differences between the current quality evaluation results and their quality labels are greater than a set difference threshold are selected from the training sample set.
12. The method according to claim 1 or 2, characterized in that: Each quality evaluation result and each quality label belongs to the same score range.
13. The method according to claim 1 or 2, characterized in that: The deep convolutional neural network includes an input layer, at least two convolutional pooling units, and a fully connected convolutional layer; each of the convolutional pooling units includes a convolutional layer, a BatchNorm layer, an activation layer, and a pooling layer.
14. The method according to claim 13, characterized in that Before inputting each training sample in the training sample set into the initial quality evaluation model to obtain the quality evaluation result of the iris image of each training sample, the method further includes: A deep convolutional neural network for classification is trained using a public data set, the output number of the trained deep convolutional neural network for classification is modified according to the quality evaluation method corresponding to the initial quality evaluation model, and the trained deep convolutional neural network for classification after the output number is modified is fine-tuned using training samples containing iris images and corresponding quality labels to obtain an initial quality evaluation model.
15. The method according to claim 1 or 2, characterized in that: The deep convolutional neural network is a ResNet, VGG or AlexNet network.
16. The method according to claim 1 or 2, characterized in that: Also includes: In a case where the difference between the current quality evaluation result and the current quality label of the iris image of at least part of the training samples in the updated training sample set is greater than the set difference threshold, selecting at least one training sample from the updated training sample set again, wherein the difference between the current quality evaluation result and the current quality label of the iris image of the reselected training sample is greater than the set difference threshold; Receiving the updated quality labels of each reselected training sample to update the corresponding training sample, thereby updating the training sample set again; and inputting each training sample in the re-updated training sample set into the current initial quality evaluation model to obtain the quality evaluation result of the iris image of each training sample in the re-updated training sample set, and calculating the difference between the quality evaluation result of the iris image of each training sample in the re-updated training sample set and its current quality label; When the differences between the quality evaluation results of the iris images of each training sample in the re-updated training sample set and their current quality labels are less than or equal to the set difference threshold, the iris image quality evaluation model is obtained according to the latest initial quality evaluation model.
17. A method for evaluating iris image quality, characterized in that: include: Obtaining an iris image to be evaluated; The quality of the iris image to be evaluated is evaluated by using the iris image quality evaluation model obtained according to the iris image quality evaluation model generation method according to any one of claims 1, 3 to 16.
18. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 17 are implemented.
19. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 17 are implemented.
Citation Information
Patent Citations
Iris image quality judging method
CN101241550A
Classifier training method and device
CN105320957A
Sample category label correction method and device
CN107451597A
Fingerprint quality evaluation method based on ridge quality expert visual cognitive machine learning
CN109003259A
Iris image quality evaluation method and device and electronic equipment
CN109712134A
