A method for training an attack material identification model, a method for identifying an attack material

Abstract

Embodiments of the present application provide a method for training an attack material recognition model and a method for recognizing attack materials. The method for training an attack material recognition model comprises: obtaining a training data set, wherein the training data set comprises mixed images corresponding to each of at least one object, the mixed images being obtained by mixing a plurality of acquisition images of different qualities, and the plurality of acquisition images being obtained by acquiring the same object; and training an attack material recognition network according to data in the training data set to obtain a target attack material recognition network. Embodiments of the present application greatly enhance the diversity of training images in the training data set by the image mixing method, reduce the influence of low-quality training images on the neural network (for example, when too many low-quality images are used to train the model, the trained network model may have an overfitting problem), and ultimately improve the quality of the obtained attack material recognition network.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of attack material identification, and in particular, the embodiments of the present application relate to a method for training an attack material identification model and a method for identifying attack materials. BACKGROUND

[0002] The attack material (for example, in the field of fingerprint identification, the attack material is a 3D finger mold made by pressing a real person's fingerprint on a flat piece of paper or a fake real person's finger) identification of the related art is realized based on an image classification algorithm, for example, a fingerprint unlocking anti-counterfeiting (for example, an under-screen fingerprint unlocking anti-counterfeiting) scheme is realized by an image classification algorithm based on data driving. This method includes the following processes: first, a large number of fingerprint images of real people and all attack scenes (i.e., material images obtained by image acquisition of attack materials, which include fake fingerprints made on the attack materials) are collected, then an algorithm framework is built (for example, a convolutional neural network is constructed), then the collected images are used to train the constructed convolutional neural network to obtain an attack material identification model, and the trained attack material identification model should have the function of identifying whether the input fingerprint belongs to an entity fingerprint or a fake fingerprint (i.e., identifying attack materials).

[0003] It is not difficult to understand that the key to determining the performance of the attack material identification model is the quality and quantity of the collected images and the method of training the neural network. Since the attack material (for example, 2.5D material) is high in cost and difficult to make, it is easy to result in a small number and low quality of such images collected, and further result in poor performance of the anti-counterfeiting identification model obtained by training.

[0004] Therefore, how to improve the number and quality of pictures included in the training data set has become a technical problem to be solved. SUMMARY

[0005] The purpose of the embodiments of the present application is to provide a method for training an attack material identification model and a method for identifying anti-counterfeiting materials. The embodiments of the present application greatly enhance the diversity of training images in the training data set by the method of image mixing, at the same time, reduce the influence of low-quality training images on the neural network (for example, too many low-quality images in the training data set will result in overfitting problem of the network model obtained by training), and finally improve the quality of the obtained attack material identification network.

[0006] In a first aspect, some embodiments of the present application provide a method for training an attack material recognition model, the method comprising: obtaining a training data set, wherein the training data set comprises mixed images corresponding to respective objects in at least one object, the mixed images being obtained by mixing a plurality of acquisition images of different qualities, the plurality of acquisition images being obtained by capturing the same object; and training an attack material recognition network according to data in the training data set to obtain a target attack material recognition network.

[0007] Some embodiments of the present application enrich the diversity and quantity of training images in the training data set by mixing acquisition images of different qualities, and ultimately improve the recognition ability of the attack material recognition model trained using these data for such attack materials.

[0008] In some embodiments, the obtaining of the training data set comprises: obtaining N acquisition images by capturing the same object; obtaining quality categories to which respective acquisition images belong to obtain at least two image sets, wherein the two image sets comprise a first image set and a second image set, and the quality of all images in the first image set is higher than the quality of respective images in the second image set; and mixing acquisition images included in the first image set with acquisition images included in the second image set to obtain the mixed images, wherein the training data set comprises images in the first image set and the mixed images.

[0009] Some embodiments of the present application use mixed images obtained by mixing acquisition images of different qualities for the same object as training data, which can avoid overfitting of the network model caused by using low-quality acquisition images as training images, compared with the prior art which directly discards such images (i.e., images in the second image set). This not only increases the quantity of training data, but also improves the utilization rate of each acquisition image, effectively solving the problem of overfitting of the model caused by directly using low-quality images (i.e., images in the second image set) to train the model in related art.

[0010] In some embodiments, the first image set and the second image set are obtained by an image segmentation model, wherein the acquisition images in the first image set are images whose captured fingerprint area is greater than or equal to a set area threshold, and the acquisition images in the second image set are images whose captured fingerprint area is less than the set area threshold.

[0011] Some embodiments of the present application propose a scheme of using an image segmentation model to clean up the collected images, that is, some embodiments of the present application train a fingerprint line segmentation model (i.e., an image segmentation model), determine the segmentation area through the segmentation mask output by the model, and regard the images with a segmentation area (i.e., the area of the fingerprint region obtained by segmentation) less than a specified threshold as low-quality collected images (images in the second image set), and regard the images greater than the threshold as normal collected images (images in the first image set), thereby improving the accuracy of image quality classification and ultimately improving the training effect of the fingerprint recognition network model.

[0012] In some embodiments, the first image set includes an ith collected image, and the second image set includes a jth collected image, where the ith collected image is any image in the first image set, the jth collected image is any image in the second image set, i and j are integers greater than or equal to 1; and the mixing processing of the collected images included in the first image set and the collected images included in the second image set to obtain the mixed image includes: performing weighted summation on pixel values of corresponding pixel points in the ith collected image and the jth collected image according to a mixing coefficient matrix to obtain the mixed image, where the mixing coefficient matrix has the same size as the collected images.

[0013] Some embodiments of the present application perform weighted summation on two images with different qualities according to a mixing coefficient matrix to realize the mixing of images, that is, the mixing operation of two images is realized through a mixing coefficient matrix.

[0014] In some embodiments, the mixing coefficient matrix includes a first mixing coefficient matrix, where the weight values of each pixel point included in the first mixing coefficient matrix are real numbers greater than or equal to zero and less than or equal to 1, and the weight values of each pixel point are the same.

[0015] Some embodiments of the present application perform mixing processing on two images through a constant mixing matrix, which can quickly expand the number of effective training samples under the premise of simple method and low cost.

[0016] In some embodiments, the mixing coefficient matrix includes a second mixing coefficient matrix, where the weight values of each pixel point corresponding to a target region on the second mixing coefficient matrix are a first value, and the weight values of each pixel point included in the remaining regions except the target region on the second mixing coefficient matrix are a second value, and the target region is one or more regions randomly selected from the jth collected image.

[0017] Some embodiments of the present application replace the pixel values of each pixel point in the corresponding region on the normal image with the pixel values of the random region on the low-quality image, so as to quickly expand the number of effective training samples under the premise of simple method and low cost.

[0018] In some embodiments, the mixing coefficient matrix comprises a third mixing coefficient matrix, wherein the third mixing coefficient matrix is obtained according to the intensity of the fingerprint signal on the ith acquisition image.

[0019] Some embodiments of the present application mix the low-quality image and the normal image through the segmentation mask of the high-quality image, so as to quickly expand the number of effective training samples under the premise of simple method and low cost.

[0020] In some embodiments, the image segmentation model is configured to output a segmentation mask corresponding to the ith acquisition image, and the third mixing coefficient matrix is the segmentation mask.

[0021] In some embodiments, the attack material recognition network comprises an input layer and a shallow convolutional layer configured to extract shallow features, wherein a step length of a convolution kernel corresponding to the shallow convolutional layer is 1.

[0022] Compared with the related art technical solution in which the step length of the shallow convolution kernel is 2, the embodiments of the present application increase the perception ability of image details, because the fingerprint anti-counterfeiting signal relies on local features.

[0023] In some embodiments, the shallow convolutional layer is a first convolutional layer adjacent to the input layer.

[0024] Some embodiments of the present application make the model pay more attention to shallow features by reducing the step length of the first convolutional layer in a manner similar to increasing the receptive field, thereby improving the extraction of shallow features by the obtained attack material recognition model and enabling the model to better recognize counterfeit fingerprints.

[0025] In some embodiments, the attack material recognition network is obtained by deepening the number of channels of the convolutional layers included in the convolutional neural network.

[0026] Some embodiments of the present application increase the diversity of the features extracted by the attack material recognition model by increasing the number of channels, thereby ultimately improving the extraction ability of the target fingerprint recognition model for diverse features.

[0027] In some embodiments, training the attack material recognition network according to the data in the training data set to obtain a target attack material recognition network comprises: training the attack material recognition network model according to the data in the training data set to obtain an initial attack material recognition network model; and performing distillation processing on the initial attack material recognition network model to obtain the target attack material recognition model.

[0028] The distillation operation of the trained model provided in some embodiments of this application can improve the quality of the fingerprint recognition network while reducing the network size, and enable the target attack material recognition model to be deployed on devices with insufficient resources.

[0029] Secondly, some embodiments of this application provide a method for identifying attack materials, the method comprising: acquiring an image to be identified; inputting the image to be identified into a target attack material identification network obtained by using the method described in any of the embodiments of the first aspect above; and determining, through the target attack material identification network, whether the image to be identified belongs to an image collected from attack materials.

[0030] Thirdly, some embodiments of this application provide an apparatus for training an attack material identification model. The apparatus includes: a training data acquisition module configured to acquire a training dataset, wherein the training dataset includes mixed images corresponding to each object in at least one object, the mixed images being obtained by mixing multiple acquired images of different qualities, the multiple acquired images being acquired from the same object; and a training module configured to train an attack material identification network based on data in the training dataset to obtain a target attack material identification network.

[0031] Fourthly, some embodiments of this application provide an apparatus for identifying attack materials, the apparatus comprising: an image acquisition module configured to acquire an image to be identified; and an anti-counterfeiting fingerprint recognition module configured to input the image to be identified into a target attack material identification network obtained by the method described in any of the embodiments of the first aspect above, and to determine whether the image to be identified belongs to an image collected from attack materials through the target attack material identification network.

[0032] Fifthly, some embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the methods described in any of the embodiments of the first or second aspect.

[0033] In a sixth aspect, some embodiments of this application provide an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, can implement the method as described in any embodiment of the first or second aspect.

[0034] In a seventh aspect, some embodiments of this application provide a computer program product comprising a computer program that, when executed by a processor, implements the method described in any of the embodiments of the first or second aspect. Attached Figure Description

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments of the present application. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be regarded as a limitation on the scope, and for those of ordinary skill in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0036] Figure 1 The schematic diagram for collecting a fingerprint image provided for the embodiments of the present application;

[0037] Figure 2 One of the method flowcharts for training an attack material recognition model provided for the embodiments of the present application;

[0038] Figure 3 The second method flowchart for training an attack material recognition model provided for the embodiments of the present application;

[0039] Figure 4 One of the weight diagrams of the first mixing coefficient matrix provided for the embodiments of the present application;

[0040] Figure 5 One of the weight diagrams of the second mixing coefficient matrix provided for the embodiments of the present application;

[0041] Figure 6 The schematic diagram of the data cleaning process provided for the embodiments of the present application;

[0042] Figure 7 The schematic diagram of the data enhancement process provided for the embodiments of the present application;

[0043] Figure 8 The second weight diagram of the first mixing coefficient matrix provided for the embodiments of the present application;

[0044] Figure 9 The second weight diagram of the second mixing coefficient matrix provided for the embodiments of the present application;

[0045] Figure 10 The weight diagram of the third mixing coefficient matrix provided for the embodiments of the present application;

[0046] Figure 11 The architecture schematic diagram of the attack material recognition network provided for the embodiments of the present application;

[0047] Figure 12 The method flowchart for identifying an attack material provided for the embodiments of the present application;

[0048] Figure 13 The component block diagram of the device for training an attack material recognition model provided for the embodiments of the present application;

[0049] Figure 14 A component block diagram of the device for identifying attack materials provided in an embodiment of the present application is shown in FIG. 1.

[0050] Figure 15 A component schematic diagram of the electronic device provided in an embodiment of the present application is shown in FIG. 2. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application.

[0052] It should be noted that similar reference numerals and letters refer to like items in the following drawings, and therefore, once an item is defined in one drawing, it need not be further defined and explained in subsequent drawings. Meanwhile, in the description of the present application, the terms "first", "second", and the like are only used to distinguish description, and cannot be understood as indicating or implying relative importance.

[0053] In recent years, important progress has been made in the research of computer vision, deep learning, machine learning, image processing, image recognition and other technologies based on artificial intelligence. Artificial intelligence (AI) is a new science and technology that studies and develops theories, methods, technologies and application systems for simulating and extending human intelligence. Artificial intelligence is a comprehensive discipline involving chips, big data, cloud computing, the Internet of Things, distributed storage, deep learning, machine learning, neural networks and many other technology categories. Computer vision, as an important branch of artificial intelligence, specifically allows machines to recognize the world. Computer vision technology usually includes face recognition, liveness detection, fingerprint recognition and anti-fraud verification, biometric recognition, face detection, pedestrian detection, object detection, pedestrian recognition, image processing, image recognition, image semantic understanding, image retrieval, character recognition, video processing, video content recognition, behavior recognition, three-dimensional reconstruction, virtual reality, augmented reality, simultaneous localization and mapping (SLAM), computational photography, robot navigation and positioning, and other technologies. With the research and progress of artificial intelligence technology, this technology has been applied in many fields, such as security, city management, traffic management, building management, park management, face passage, face attendance, logistics management, warehouse management, robots, intelligent marketing, computational photography, mobile imaging, cloud services, smart home, wearable devices, unmanned vehicles, autonomous driving, intelligent medical care, face payment, face unlocking, fingerprint unlocking, face and certificate verification, smart screens, smart televisions, cameras, mobile Internet, network live streaming, beauty, makeup, medical cosmetology, intelligent temperature measurement, and other fields.

[0054] The biometric security test process needs to prevent material (e.g., 2.5D material or 3D material) attacks, i.e., to identify attack materials. It can be understood that the collected images for identifying attack materials include but are not limited to fingerprint images, and these images can also be other biometric feature images such as palm print images, iris images, lip print images, or face images, etc. For example, a related company proposed a fingerprint biometric security test process in August 2020, which clearly pointed out that the fingerprint unlocking part needs to increase the detection of 2.5D material attacks (i.e., identify 2.5D attack materials). It should be noted that 2.5D material is an attack material between 2D and 3D, which is made of 2D fingerprint images to display the ridge and valley lines of the fingerprint with concave and convex lines, and compared with 3D material, the 2.5D material lacks depth information of the ridge and valley lines. For experienced attack material makers, they only need to obtain the 2D fingerprint image of the living being to make attack materials (e.g., 2.5D attack materials), so this material making is relatively easy, and it is not difficult to understand that if the existing product cannot effectively identify these attack materials, the attack behavior of the product will cause great security risks. Therefore, improving the detection accuracy of material attacks (e.g., 2.5D material attacks) is of great significance to improve the anti-fake performance of fingerprint unlocking.

[0055] The following will be described in combination with Figure 1 The existing technical solutions for identifying materials have many defects. It can be understood that the fingerprint image acquisition device required by the under-screen fingerprint unlocking scheme can be a capacitive sensor or an ultrasonic sensor or other types of sensing lines, and the embodiments of the present application do not limit the specific type of image acquisition sensor.

[0056] The existing under-screen fingerprint unlocking anti-fake scheme is implemented based on a data-driven image classification algorithm. This method at least includes the following processing process: first, collect a large number of real person and all attack scene (i.e., attack material) fingerprint images, such as Figure 1As shown, a large number of real fingerprint images are collected by using an under-screen capacitive sensor to collect live fingerprints, and material images are collected by using materials; then, an algorithm framework is built (for example, a convolutional neural network is constructed); and then, the collected images are used to train the neural network to obtain a target attack material recognition model. It is not difficult to understand that when these technical solutions are used for anti-counterfeiting identification, the key to determining the performance of the attack material recognition model is the quality and quantity of the collected images in the training data set and the method of training the neural network. The quality of the image can be determined by evaluating the strength and area of the characteristic signal collected by the collected image. For example, taking a fingerprint collection image as an example, the quality of the image can be represented by the strength of the fingerprint line signal collected on the collected image or the area of the collected fingerprint. For example, the stronger the line signal and the larger the area, the better the quality of the collected image. Taking a palm print collection image as an example, the quality of the image can be represented by the strength of the palm print line signal collected on the collected image or the area of the collected palm print line.

[0057] It should be noted that attack materials (for example, 2.5D materials) are high in cost and difficult to make, which can easily lead to a small number of images (for example, 2.5D images) collected and low quality. It is not difficult to understand that only a small amount of training data can significantly reduce the training effect of the anti-counterfeiting model. In addition, high-quality material images (for example, collected 2.5D images) have high similarity with real person images, so it is difficult to train a model with high anti-counterfeiting performance on the basis of ensuring low false negatives of the anti-counterfeiting model.

[0058] Therefore, the inventors of the present application found through the above analysis and research that how to improve the lack of training data and improve the quality of material images has become a technical problem to be solved.

[0059] At least to solve the above technical problems, some embodiments of the present application provide a method of obtaining a mixed image by mixing two or more material images of different shooting qualities and using the mixed image instead of a collection image of poor shooting quality (for example, the collection image types include: a fingerprint image, a palmprint image, an iris image, a lipprint image, or a face image, etc.) as training data (for example, using a semantic segmentation model to segment the fingerprint area of a material image of the same attack material to distinguish low-quality images and high-quality images, and then mixing the two images), in this way, on the one hand, the diversity of training images is greatly enhanced (the number of training images increases after mixing), and on the other hand, the influence of network overfitting caused by directly training the network with low-quality collection images is reduced (since the mixed image instead of the low-quality collection image is used as training data, the problem of overfitting of the target fingerprint recognition network obtained by training is effectively avoided). In order to further improve the performance of the target fingerprint recognition network, some embodiments of the present application propose a "two-step training method" for attack material recognition model training: the first step is to modify the model based on the reference convolutional neural network (adjust the step size and / or increase the number of channels, for details, please refer to the following text) to capture more detailed features, and the second step is to distill a lightweight attack material recognition model from the model trained in the previous step as a teacher model.

[0060] The following will be described in conjunction with Figure 2 An exemplary method for training an attack material recognition model is described, for example, which can be executed by a terminal for image recognition or by a server.

[0061] As Figure 2 shown, some embodiments of the present application provide a method 10 for training an attack material recognition model, the method 10 comprising: S110, obtaining a training data set, wherein the training data set comprises mixed images corresponding to each object in at least one object respectively, the mixed image is obtained by mixing multiple collection images of different qualities, and the multiple collection images are obtained by collecting the same object; and S120, training an attack material recognition network according to the data in the training data set to obtain a target attack material recognition network.

[0062] It can be understood that, by mixing collection images of different qualities, some embodiments of the present application enrich the number of training images in the training data set, and ultimately improve the recognition ability of the model trained by using these data for the objects.

[0063] The above process of obtaining a training data set (corresponding to S110) and model training (corresponding to S120) is described below.

[0064] In some embodiments of the present application, the different quality acquisition images involved in S110 include low quality acquisition images and high quality acquisition images. In some embodiments of the present application, the different quality acquisition images involved in S110 can include more than two types of acquisition images of different quality, for example, the different quality acquisition images involved in S110 include high quality acquisition images, intermediate quality acquisition images and low quality acquisition images, and the corresponding mixing of acquisition images of different quality can include mixing high quality acquisition images and low quality acquisition images, mixing high quality acquisition images and intermediate quality acquisition images, and mixing intermediate quality acquisition images and low quality acquisition images. The present application does not limit the specific number of categories of images of different quality.

[0065] In some embodiments of the present application, the plurality of acquisition images involved in mixing the plurality of acquisition images of different quality in S110 includes two or more. For example, in some embodiments of the present application, the mixed image is obtained by mixing two acquisition images of different quality. In some other embodiments of the present application, the mixed image can be obtained by mixing a plurality of acquisition images of different quality, for example, mixing a first acquisition image and a second acquisition image to obtain an initial mixed image, and then mixing the initial mixed image and a third acquisition image to obtain a mixed image, wherein the first acquisition image, the second acquisition image and the third acquisition image belong to images of different shooting quality.

[0066] In some embodiments of the present application, the same object involved in S110 refers to the same attack material (the material has a fingerprint or palmprint made thereon). For example, the same 2.5D material or 3D material. It should be noted that the plurality of acquisition images in some embodiments of the present application can be a plurality of fingerprint acquisition images or a plurality of palmprint acquisition images, etc., and the present application does not limit the specific type of acquisition images.

[0067] As shown in FIG. 1, in some embodiments of the present application, S110 exemplarily includes: Figure 3

[0068] S111, image acquisition is performed on the same object to obtain N acquisition images, wherein N is an integer greater than 1.

[0069] For example, image acquisition is performed on the same object (i.e. the same attack material, which includes a 2.5D material or a 3D material and has a fingerprint or a palmprint made thereon) to obtain 20 acquisition images (as an example of N acquisition images).

[0070] It can be understood that the device for image acquisition of the same object can include an optical fingerprint sensor, an ultrasonic sensor or a semiconductor sensor, etc., and the present application does not limit the specific type of sensor for acquisition image. ​

[0071] S112, obtaining a quality category to which each of the N captured images belongs, to obtain at least two image sets, wherein the two image sets include a first image set and a second image set, and the quality of each image in the first image set is higher than the quality of each image in the second image set.

[0072] In some embodiments of the present application, the image quality is evaluated by the strength of the captured fingerprint line signal, and in some other embodiments of the present application, the image quality can also be evaluated by the area of the captured fingerprint. For example, the larger the captured fingerprint area is, the better the quality of the corresponding image is, or the stronger the captured fingerprint line signal is, the better the quality of the image is.

[0073] It can be understood that in some embodiments of the present application, the N captured images obtained in S111 are divided into two groups (corresponding to two image sets) or multiple groups of images (corresponding to multiple image sets) with different qualities by using an image processing algorithm. In some embodiments of the present application, the N captured images obtained in S111 can also be divided into two groups or multiple groups of images with different qualities by using a trained image classification model. For example, in some embodiments of the present application, the first image set and the second image set are obtained by using an image segmentation model, wherein the image segmentation model identifies the captured images with a fingerprint area less than a set threshold as the captured images in the second image set, and identifies the captured images with a fingerprint area greater than or equal to the set threshold as the captured images in the first image set. It should be noted that the area of the captured fingerprint is determined by a target segmentation mask, wherein the segmentation mask is a prediction result obtained by inputting the captured image into the image segmentation model. It can be understood that the segmentation mask has the same size as the input captured image, and the weight value of each pixel point on the segmentation mask represents the strength of the fingerprint signal corresponding to the pixel point. In order to obtain the area of the captured fingerprint, a threshold can be set, the value of the pixel point greater than the threshold on the segmentation mask is replaced by a number d (for example, the value of the d can be 1), and the value of the pixel point less than the threshold is replaced by a number e (for example, the value of the e can be 0). It can be understood that the area of the captured fingerprint can be determined according to the new encoding matrix.

[0074] That is, if the collected images are fingerprint collected images, in order to improve the accuracy of the quality categories to which the collected images belong, some embodiments of the present application obtain the quality categories to which the collected images belong according to the sizes of the areas of the fingerprints collected on the collected images. For example, the collected images are divided into two categories or multiple categories according to a set area threshold. It can be understood that the N collected images obtained in S111 can be divided into two categories according to a set area threshold, and the two categories can be named as low-quality images (as an example of the second image set) and high-quality images (as an example of the first image set). The N collected images obtained in S111 can be divided into three categories according to two set area thresholds, and the three categories can be named as low-quality images, intermediate-quality images and high-quality images. The first image set and the second image set are two relative concepts. If the low-quality images and the intermediate-quality images are mixed, the low-quality images are taken as an example of the second image set and the intermediate-quality images are taken as an example of the first image set. If the low-quality images and the high-quality images are mixed, the low-quality images are taken as an example of the second image set and the high-quality images are taken as an example of the first image set. If the intermediate-quality images and the high-quality images are mixed, the intermediate-quality images are taken as the second image set and the high-quality images are taken as the first image set.

[0075] S113, mixing the collected images included in the first image set obtained in S112 and the collected images included in the second image set to obtain the mixed images, wherein the training data set includes the images in the first image set and the mixed images.

[0076] It should be noted that, in some embodiments of the present application, the process of mixing two images in S113 includes mixing one image from each of the two sets obtained in S112 to obtain a mixed image. In some embodiments of the present application, multiple images in the first set obtained in S112 and multiple images in the second set can be mixed to obtain a mixed image. The embodiments of the present application do not limit the specific number of collected images based on which the mixed image is obtained.

[0077] The implementation process of S113 is exemplarily described below by taking the mixing of one image from each of the two sets obtained in S112 as an example.

[0078] In some embodiments of the present application, the first image set in S112 includes an ith acquisition image, and the second image set in S112 includes a jth acquisition image, where the ith acquisition image is any image in the first image set, the jth image is any image in the second image set, i and j are integers greater than or equal to 1; and the corresponding S113 exemplarily includes: performing weighted summation on pixel values of corresponding pixel points in the ith acquisition image and the jth acquisition image according to a mixing coefficient matrix to obtain the mixed image, where the mixing coefficient matrix has the same size as the acquisition image.

[0079] It should be noted that some embodiments of the present application can use different mixing coefficient matrices (for example, the first mixing coefficient matrix, the second mixing coefficient matrix and the third mixing coefficient matrix below) to mix two same images to obtain multiple mixed images. In some embodiments of the present application, different mixing coefficient matrices can be used to mix two different images to obtain multiple different mixed images. For example, when the total number of mixed images obtained by some embodiments of the present application for the same object is K, M mixed images in the K mixed images are obtained by the first mixing coefficient matrix, and L mixed images are obtained by the second mixing coefficient matrix, where the sum of M and N is equal to K.

[0080] The process of mixing images according to different mixing coefficient matrices is exemplarily described below in combination with the ith acquisition image and the jth acquisition image.

[0081] In some embodiments of the present application, corresponding pixel values of two acquisition images (i.e., the ith acquisition image and the jth acquisition image described above) are weighted and summed according to a certain proportion to obtain a mixed image. For example, in some embodiments, the mixing coefficient matrix includes a first mixing coefficient matrix, where the weight values of each pixel point included in the first mixing coefficient matrix are real numbers greater than or equal to zero and less than or equal to 1, and the weight values of each pixel point are the same.

[0082] For example, the first mixing coefficient matrix included in the mixing coefficient matrix described above is as shown in Figure 4 It should be understood that Figure 4 The mixing coefficient matrix shown in the above figure has a size of 5*5 (it should be understood that the size of the mixing coefficient matrix can be adjusted according to the size of the mixed image), and the mixing coefficient matrix can mix two images of the same size. In Figure 4 In the above figure, the weight values of each pixel point are all the number a (the number is a value greater than or equal to zero and less than or equal to 1), that is, the weight values of each pixel point are the same. That is, if the weight values of each pixel point are represented by the same letter on the same figure, the corresponding specific weight values are also the same.

[0083] In some embodiments of the present application, a random region of the jth acquisition image (e.g., an acquisition image divided into a low-quality map) is filled into a corresponding region of the ith acquisition image (e.g., an acquisition image divided into a normal map), to realize mixing of the ith acquisition image and the jth acquisition image. For example, the mixing coefficient matrix described above includes a second mixing coefficient matrix, wherein the weight values of the pixel points corresponding to the target region on the second mixing coefficient matrix are first numerical values, and the weight values of the pixel points included in the remaining regions other than the target region on the second mixing coefficient matrix are second numerical values, and the target region is one or more regions randomly selected from the jth acquisition image.

[0084] For example, the second mixing coefficient matrix included in the mixing coefficient matrix described above is as shown in Figure 5 It can be understood that Figure 5 The mixing coefficient matrix shown in the figure has a size of 5*5, and the mixing coefficient matrix can mix two images of the same size. In Figure 5 , the weight values of the pixel points in the target region 321 are all the number b (corresponding to the first numerical value, which is a number greater than or equal to zero and less than or equal to 1), and the weight values of the pixel points in the remaining regions are the number c (corresponding to the second numerical value), wherein the specific numerical values corresponding to the number b and the number c are different. That is, the specific values of the weight values of the corresponding pixel points represented by different letters on the same figure are also different. For example, in Figure 5 , the number b takes the value of 1, and the number c takes the value of 0. It should be noted that different letters on different figures (i.e., Figure 4 and Figure 5 ) can represent the same value. For example, Figure 4 , the number a takes the value of 1, and Figure 5 , the number b can also take the value of 1.

[0085] In some embodiments, the mixing coefficient matrix includes a third mixing coefficient matrix, wherein the third mixing coefficient matrix is obtained according to the intensity of the fingerprint signal on the ith acquisition image. For example, in some embodiments of the present application, the jth acquisition image and the segmentation result of the ith acquisition image output by the image segmentation model are used as the mixing coefficient matrix. It can be understood that the segmentation result (i.e., the segmentation mask) is a clearness weight of a pixel level of a fingerprint image, which can be a soft mask of any real number between 0 and 1. That is, in some embodiments of the present application, the image segmentation model is used to output a segmentation mask corresponding to the ith acquisition image, and the third mixing coefficient matrix is the segmentation mask.

[0086] Some embodiments of the present application can process low-quality images into images that can be used for training by mixing different quality of collected images (e.g., fingerprint collection images) taken of the same object, which not only increases the amount of training data but also improves the utilization of each collected image compared to the prior art solution of directly discarding such images (i.e., images in the second image set), and effectively solves the problem of overfitting of the model caused by directly training the model with low-quality images (i.e., images in the second image set).

[0087] The architecture of the attack material recognition network constructed by some embodiments of the present application is exemplarily described below.

[0088] To increase the ability to perceive image details, in some embodiments of the present application, the attack material recognition network (e.g., a fingerprint recognition network) includes an input layer and a shallow convolutional layer for extracting shallow features, wherein the step size of the shallow convolutional layer is 1. For example, in some embodiments, the shallow convolutional layer is the first convolutional layer adjacent to the input layer. It can be understood that, compared to the prior art solution of a shallow convolutional layer with a step size of 2, the embodiments of the present application increase the ability to perceive image details, because the fingerprint anti-counterfeiting signal relies more on local features.

[0089] Some embodiments of the present application make the model pay more attention to shallow features by reducing the step size of the first convolutional layer, which is similar to increasing the receptive field, thereby improving the extraction of shallow features by the obtained fingerprint recognition model and enabling the model to better recognize counterfeit fingerprints.

[0090] To improve the diversity of features extracted by the attack material recognition model, in some embodiments of the present application, the attack material recognition network is obtained by deepening the channel number of the convolutional layers included in a reference convolutional neural network (e.g., ResNet, MobileNet, and ShuffleNet network models). That is, some embodiments of the present application can improve the diversity of features extracted by the attack material recognition model by increasing the channel number, thereby ultimately improving the extraction ability of the target anti-counterfeiting plastic recognition model for diverse features.

[0091] How some embodiments of the present application train the constructed attack material recognition network model is exemplarily described below.

[0092] To obtain high-quality recognition results while reducing the size of the network model, in some embodiments of the present application, S120 includes: training the attack material recognition network model according to the data in the training data set to obtain an initial attack material recognition network model; and performing distillation processing on the initial attack material recognition network model to obtain the target attack material recognition model.

[0093] That is, the operation of distilling the trained model provided by some embodiments of the present application can improve the quality of the attack material recognition network while reducing the network size, so that the target fingerprint recognition model can be deployed on devices with insufficient resources.

[0094] The above method of training an attack material recognition model is exemplarily described below using 2.5D material (as one of the attack materials) and low-quality images (as an example of the second image set) and normal images (as an example of the first image set). It should be noted that in the following examples, the attack material recognition model is specifically implemented as a fingerprint recognition model.

[0095] The method of training a fingerprint recognition model (as a specific example of an attack material recognition model) provided by some embodiments of the present application includes data cleaning, data enhancement, large model training, and model distillation. Specifically, the following steps are included:

[0096] The first step is data cleaning, that is, classifying the collected images according to the quality difference. For example, the quality difference can be determined by the area of the collected fingerprint, or the quality difference can be determined according to the signal strength of the collected fingerprint area. For example, the strength of the fingerprint area signal can be determined by the peak signal-to-noise ratio.

[0097] Since the 2.5D attack material triggers the screen to be difficult to capture Figure 10 When capturing 2.5D fingerprint images (i.e., obtaining material images), many low-quality images without fingerprints will be captured. Such images are noise data, and training a neural network model based on these images will increase the risk of overfitting of the neural network, which is harmful to the training of the neural network. Therefore, some embodiments of the present application propose to use an image segmentation model to clean 2.5D data (i.e., collected images).

[0098] The device for implementing data cleaning in some embodiments of the present application is shown in Figure 6 The figure includes an image segmentation model 100 and an image quality classification module 200, wherein the image segmentation model 100 is a trained network model (for example, the architecture of the image segmentation model can adopt the architecture of a semantic segmentation model).

[0099] For example, as shown in Figure 6As shown, some embodiments of the present application input the first collected image (any one material image) into the trained network model (i.e. image segmentation model) to obtain a segmentation mask corresponding to the input image (the segmentation mask has the same size as the input image, but the value of each pixel point on the segmentation mask is the collected fingerprint intensity value). The segmentation mask is input into the image quality classification module 200, which is configured to obtain the segmentation area (i.e. the area of the segmented fingerprint region) according to the segmentation mask. For images with a segmentation area less than a set area threshold, they are classified as low-quality images (a plurality of low-quality images form Figure 6 the illustrated low-quality image set 222), and images with a segmentation area greater than the set area threshold are classified as normal images (a plurality of normal images form Figure 6 the illustrated normal image set 221). It can be understood that in addition to classifying images according to the fingerprint area, in some other embodiments of the present application, the image quality classification module 200 can also classify images into multiple categories according to the strength of the fingerprint line signal.

[0100] Second step, data augmentation, i.e. the process of mixing different quality collected images.

[0101] Since the number of collected 2.5D images is relatively small, and there are many low-quality images with weak fingerprint signals, some embodiments of the present application do not directly discard this part of the images, but make better use of the low-quality images. For example, some embodiments of the present application use an image mixing method to mix the collected fingerprint-free images or low-quality images with less fingerprint area with normal images with fingerprints, which greatly expands the number of training images, thereby improving the performance of the model trained using these data.

[0102] For example, in some embodiments of the present application, the image mixing method uses the following mixing formula to calculate the mixed image:

[0103] I fusion (x, y) = I norm (x, y)M(x, y) + I low , (x, y) (1-M(x, y)),

[0104] where x, y represent the pixel position of any pixel point on the collected image, I fusion , I norm and I low represent the enhanced mixed image (i.e. mixed image), normal image (as an example of the first image set) and low-quality image (as an example of the second image set) respectively, M(x, y) represents the mixing coefficient matrix, and "1-M(x, y)" is the complementary matrix of the mixing coefficient.

[0105] The following will be described in combination with Figure 7This will illustrate the calculation process of the above mixed formula. Figure 7 The images to be mixed include the i-th acquired image 401 mentioned above (this acquired image belongs to...) Figure 6 (One image from the obtained normal image set) and the j-th acquired image 501 mentioned above (this acquired image belongs to) Figure 6 The first image to be mixed (not shown in the figure) is obtained by weighting and summing the weight values ​​of corresponding pixels in the first mixing coefficient weight map 411 (i.e., the image corresponding to any mixing coefficient matrix) of the i-th acquired image and the first mixing coefficient complementary weight map 412 (i.e., the image corresponding to any complementary mixing coefficient matrix). The second image to be mixed (not shown in the figure) is obtained by weighting and summing the weight values ​​of corresponding pixels in the first and second mixed images. Then, the values ​​of corresponding pixels in the first and second mixed images are added together to obtain the mixed image 601. It is easy to understand that the quality of the mixed image 601 is better than that of the j-th acquired image. Therefore, using the mixed image 601 as the training dataset can significantly improve the overfitting problem of the trained target fingerprint recognition model compared to directly using the j-th acquired image 501 as the training image.

[0106] The following example illustrates the type of the mixing coefficient matrix.

[0107] (1) As described above, in some embodiments of this application, the mixing coefficient matrix includes a first mixing coefficient matrix M(x,y), which is a random number uniformly sampled within the range of (0,1), meaning that the values ​​of each pixel in the mixing coefficient matrix are the same. Figure 8 As shown (as) Figure 4 (A specific example) In this first mixing coefficient matrix, the value of the number 'a' for each pixel is 0.2. It is not difficult to understand that... Figure 7 The weight value of all pixels in the complementary mixing coefficient matrix corresponding to this mixing coefficient matrix is ​​0.8.

[0108] (2) As described above, some embodiments of this application include a second mixing coefficient matrix, which is used to mix the j-th acquired image (which belongs to) Figure 6 A random region from a low-quality image in the resulting low-quality image set is filled into the i-th acquired image (which belongs to...). Figure 6 The corresponding region of a normal image from the obtained set of normal images is used to obtain a mixed image.

[0109] For example, the formula for determining the value of the second mixing coefficient matrix M(x,y) is as follows:

[0110]

[0111] Where (x1, y1) are the coordinates of the top-left corner of the random region, and (x2, y2) are the coordinates of the bottom-right corner of the random region, and satisfy the following two inequalities:

[0112]

[0113]

[0114] In the above formula, W and H represent the width and height of the j-th acquired image.

[0115] For example, the weight values ​​of each pixel in the second mixing coefficient matrix are as follows: Figure 9 As shown (as) Figure 5 A specific example), in the diagram with Figure 5 The specific value of the corresponding number b is 1, and... Figure 5 The corresponding number c takes the value 0. This means the target area, the area to be replaced, is the region comprised of all pixels with the value 1.

[0116] (3) As described above, the mixing coefficient matrix of some embodiments of this application includes a third mixing coefficient matrix.

[0117] For example, the third mixing coefficient matrix can be obtained by using Figure 6 The image segmentation model is used for the i-th acquired image (which belongs to...) Figure 6 The segmentation mask obtained from a normal image in the obtained set of normal images is the mixing coefficient matrix, which is the segmentation result of the normal image. The segmentation result is a fingerprint image pixel-level clarity weight, which can be a soft mask with any real number between 0 and 1. It can be understood that the weight value of each pixel on the segmentation mask is used to characterize the fingerprint strength of the corresponding pixel.

[0118] For example, in some embodiments of this application, by Figure 6 The segmentation result (i.e., segmentation mask) output by the image segmentation model for the i-th acquired image is a soft mask, which can be used as a third mixing coefficient matrix. For example, this soft mask is characterized as follows:

[0119] M(x,y)=norm(seg_score(x,y))

[0120] Where seg_score(x, y) represents the output of the image segmentation model at the (x, y) position in the image, and its value ranges from (0, 255). The larger the value, the stronger the fingerprint signal. Therefore, it needs to be normalized. The normalization formula is:

[0121] Third step, large model training. That is, the constructed fingerprint identification model (as a specific example of the attack material identification model) is trained according to the enhanced image and the normal image pair obtained after the second step processing.

[0122] Since the high-frequency signal of the fingerprint image is an important basis for anti-counterfeiting, unlike ordinary deep learning models, the fingerprint identification model of some embodiments of the present application needs to pay more attention to the shallow features when designing. Therefore, for any convolutional neural network model, the embodiments of the present application need to double the resolution of the corresponding feature map of the shallow layer and double the number of convolution kernels of each convolution layer.

[0123] It should be noted that any reference convolutional neural network model can include ResNet, MobileNet, and ShuffleNet network models, and unlike these prior art network models, some embodiments of the present application set the step of the first convolutional layer (used to extract shallow features) included in these network models to 1 and double the number of channels of each convolutional layer. As shown in Figure 11 , the related art reference convolutional neural network 611 (for example, the network includes ResNet, MobileNet, or ShuffleNet network model) and the improved fingerprint identification network or target fingerprint identification network (as a specific example of the attack material identification network) of some embodiments of the present application are shown, wherein for the first convolutional layer 612, the embodiments of the present application adjust the step of this layer from the prior art 2 to 1 (i.e. increase the resolution of a type of feature map), and in Figure 11 , the convolutional channel number of each convolutional layer is doubled (i.e. each convolutional layer increases the number of convolution kernels by one). It can be understood that, Figure 11 , the thickness of each layer represents the number of channels of the corresponding layer.

[0124] Fourth step, model distillation. That is, the network model obtained after the third step training is distilled to obtain a smaller model. In this way, the small-scale model can be deployed on a resource-constrained device.

[0125] Since the large model does not meet the memory and speed requirements for deployment on the terminal platform, some embodiments of the present application use the model distillation method to distill the large model obtained in the third step to a small model that meets the requirements.

[0126] It should be noted that after the above third step or fourth step processing, a network classifier is obtained, which can identify whether the input collected image corresponds to an entity fingerprint or an attack material. This network classifier is trained by the image processed by the second step data enhancement.

[0127] The method of identifying attack materials performed by a terminal or the like is exemplarily described below.

[0128] As shown in Figure 12 some embodiments of the present application provide a method of identifying attack materials, which comprises: S210, acquiring an image to be identified; and S220, inputting the image to be identified into a target attack material identification network obtained by the method of training a fingerprint identification model, and determining whether the image to be identified belongs to an image obtained by collecting attack materials through the target attack material identification network.

[0129] It should be noted that the training process of the target attack material identification network can refer to the description above, and will not be described in detail here to avoid repetition.

[0130] Some embodiments of the present application provide a computer-readable storage medium having a computer program stored thereon, and the program is executed by a processor to implement the method described in any one of the embodiments of the method of training a fingerprint identification model or the method of identifying a counterfeit fingerprint.

[0131] Please refer to Figure 13 , Figure 13 a device for training an attack material identification model is shown, and it should be understood that the device 20 corresponds to the above Figure 2 method embodiments, and can perform each step involved in the above method embodiments. The specific functions of the device 20 can be referred to the description above, and the detailed description will be appropriately omitted here to avoid repetition. The device 20 includes at least one software functional module stored in the form of software or firmware in the memory or solidified in the operating system of the device 20. The device 20 includes a training data acquisition module 301 and a training module 302.

[0132] The training data acquisition module 301 is configured to acquire a training data set, wherein the training data set includes a mixed image corresponding to each object in at least one object, and the mixed image is obtained by mixing a plurality of acquisition images of different qualities, and the plurality of acquisition images are obtained by collecting the same object.

[0133] The training module 302 is configured to train an attack material identification network according to the data in the training data set to obtain a target attack material identification network.

[0134] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the device 20 described above can refer to the corresponding process in the foregoing method of training a fingerprint identification model, and will not be described in detail here.

[0135] Please refer to Figure 14 ,Figure 14 It is shown that the device for identifying attack materials provided by the embodiments of the present application, it should be understood that the device 30 corresponds to the above-mentioned Figure 12 method embodiments, can perform each step involved in the above-mentioned method embodiments, and the specific functions of the device 30 can be referred to the description in the foregoing, to avoid repetition, the detailed description is appropriately omitted here. The device 30 includes at least one software function module that can be stored in the form of software or firmware in the memory or solidified in the operating system of the device 20, the device 30, comprising: an image to be identified acquisition module 311 and a counterfeit fingerprint identification module 312.

[0136] The image to be identified acquisition module 311 is configured to acquire an image to be identified.

[0137] The counterfeit fingerprint identification module 312 is configured to input the image to be identified into a target attack material identification network obtained by using the method for training the fingerprint identification model, and determine whether the image to be identified belongs to an image obtained by collecting attack materials through the target attack material identification network.

[0138] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the device 30 described above can refer to the corresponding process in the foregoing method for identifying counterfeit fingerprints, and will not be described in more detail here.

[0139] Some embodiments of the present application provide a computer program product, the computer program product includes a computer program, the computer program is executed by the processor to implement the method described in any embodiment of the method for training the fingerprint identification model or the method for identifying the counterfeit fingerprint.

[0140] As Figure 15 shown, some embodiments of the present application provide an electronic device 500, the electronic device 500 includes a memory 510, a processor 520, a bus 530, and a computer program stored in the memory 510 and executable on the processor 520, wherein the processor 520 reads the program from the memory 510 through the bus 530 and executes the program to implement the method described in any embodiment of the method for training the fingerprint identification model or the method for identifying the counterfeit fingerprint.

[0141] The processor 520 can process digital signals and can include various computing structures. For example, a complex instruction set computer structure, a reduced instruction set computer structure, or a structure that implements a combination of multiple instruction sets. In some examples, the processor 520 can be a microprocessor.

[0142] The memory 510 can be used to store instructions executed by the processor 520 or data related to the instructions execution process. The instructions and / or data can include code for implementing some or all of the functions of one or more modules described in the embodiments of the present application. The processor 520 of the embodiments of the present disclosure can be used to execute the instructions in the memory 510 to implement the method shown in Figure 2 、 Figure 3 or Figure 12 The memory 510 includes dynamic random access memory, static random access memory, flash memory, optical memory, or other memory well known to those skilled in the art.

[0143] In several embodiments provided in the present application, it should be understood that the disclosed apparatus and method can also be implemented by other means. The apparatus embodiments described above are only schematic, for example, the flowcharts and block diagrams in the drawings show the possible implementation architectures, functions and operations of the apparatus, method and computer program product according to the embodiments of the present application. In this regard, each block in the flowcharts or block diagrams can represent a module, a program segment or a part of code, which contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in a different order than that noted in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and the combination of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0144] In addition, each functional module in the embodiments of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0145] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts of the prior art that make contributions or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0146] The above merely provides an example of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application. It should be noted that similar reference numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0147] The above merely provides an example of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application. It should be noted that similar reference numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0148] It should be noted that, in this document, the terms such as first and second are used merely to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the processes, methods, articles or devices including a series of elements not only include those elements, but also include other elements not explicitly listed or inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or device including the element.

Claims

1. A method of training an attack material identification model, the method comprising: The method comprises: obtaining a training data set, wherein the training data set comprises mixed images corresponding to each of at least one object respectively, the mixed images are obtained by mixing a plurality of acquisition images of different qualities, the plurality of acquisition images are obtained by acquiring the same object, the mixed images comprise images obtained by mixing images with a fingerprint area greater than or equal to a set area threshold and images with a fingerprint area less than the set area threshold, the images with a fingerprint area greater than or equal to the set area threshold and the images with a fingerprint area less than the set area threshold are obtained by an image segmentation model; training an attack material recognition network according to data in the training data set to obtain a target attack material recognition network.

2. The method of claim 1, wherein obtaining a training data set comprises: acquiring N acquisition images of the same object, wherein N is an integer greater than 1; obtaining quality categories of each acquisition image in the N acquisition images to obtain at least two image sets, wherein the two image sets comprise a first image set and a second image set, the quality of all images in the first image set is higher than the quality of each image in the second image set; mixing acquisition images included in the first image set with acquisition images included in the second image set to obtain the mixed images, wherein the training data set comprises images in the first image set and the mixed images.

3. The method of claim 2, wherein, obtaining the first image set and the second image set by an image segmentation model, wherein the acquisition images in the first image set are images with a fingerprint area greater than or equal to a set area threshold, and the acquisition images in the second image set are images with a fingerprint area less than the set area threshold.

4. The method according to any one of claims 2-3, wherein, The first image set includes an ith acquisition image, and the second image set includes a jth acquisition image, wherein the ith acquisition image is any image in the first image set, the jth acquisition image is any image in the second image set, i and j are integers greater than or equal to 1; mixing acquisition images included in the first image set with acquisition images included in the second image set to obtain the mixed images comprises: weighting and summing pixel values of corresponding pixel points in the ith acquisition image and the jth acquisition image according to a mixing coefficient matrix to obtain the mixed images, wherein the mixing coefficient matrix has the same size as the acquisition images.

5. The method of claim 4, wherein, The mixing coefficient matrix comprises a first mixing coefficient matrix, wherein the weight values of each pixel point included in the first mixing coefficient matrix are real numbers greater than or equal to zero and less than or equal to 1, and the weight values of each pixel point are the same.

6. The method of claim 4, wherein, The mixing coefficient matrix comprises a second mixing coefficient matrix, wherein a weight value of each pixel point corresponding to a target region on the second mixing coefficient matrix is a first numerical value, and a weight value of each pixel point included in a remaining region except the target region on the second mixing coefficient matrix is a second numerical value, and the target region is one or more regions randomly selected from the jth acquisition image.

7. The method of claim 4, wherein, The mixing coefficient matrix comprises a third mixing coefficient matrix, wherein the third mixing coefficient matrix is obtained according to the intensity of the fingerprint signal on the ith acquisition image.

8. The method of claim 7, wherein, The image segmentation model is used to output a segmentation mask corresponding to the ith acquisition image, and the third mixing coefficient matrix is the segmentation mask.

9. The method of claim 1, wherein, The attack material recognition network comprises an input layer and a shallow convolutional layer for extracting shallow features, wherein a step length of a convolution kernel corresponding to the shallow convolutional layer is 1.

10. The method according to any one of claims 1 to 3 or 9, wherein The attack material recognition network is obtained by deepening the channel number of a convolutional layer included in a reference convolutional neural network.

11. The method of any one of claims 1-3 or 9, wherein, The training of the attack material recognition network according to the data in the training data set comprises: training the attack material recognition network according to the data in the training data set to obtain an initial attack material recognition network model; performing distillation processing on the initial attack material recognition network model to obtain the target attack material recognition model.

12. A method of identifying an attacking material, characterized by, The method comprises: obtaining a to-be-recognized image; inputting the to-be-recognized image into the target attack material recognition network obtained by the method of any one of claims 1-11, and determining whether the to-be-recognized image belongs to an image obtained by collecting an attack material through the target attack material recognition network.

13. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the method of any one of claims 1-12.

14. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein, The processor executes the program to implement the method of any one of claims 1-12.

15. A computer program product, characterised in that, The computer program product comprises a computer program, and the computer program is executed by the processor to implement the method of any one of claims 1-12.

Images