Iris Image Anti-Counterfeiting Detection Method Based on Circular Attention Mechanism

Abstract

An embodiment of the present invention provides an iris image anti-counterfeiting detection method based on a cyclic attention mechanism, including: inputting a global iris image into an iris anti-counterfeiting detection network model; processing the global iris image through a first-level network structure unit to obtain a prediction probability and key region parameters at a first-level feature scale; performing, through each level of network structure unit: based on the input image at the current feature scale input to the current-level network structure unit and the key region parameters at the current-level feature scale output by the current-level network unit, obtaining an input image at the next-level feature scale, and processing the input image at the next-level feature scale by the next-level network structure unit to obtain a prediction probability and key region parameters at the next-level feature scale corresponding to the next-level network structure unit; fusing the prediction probabilities at at least two different-level feature scales to obtain a determination result as to whether the global iris image is a real iris image or a forged iris image.

Description

Technical Field

[0001] The present invention belongs to the technical field of biometric recognition. Specifically, it relates to an iris image anti-counterfeiting detection method based on a cyclic attention mechanism. Background Technique

[0002] Iris recognition is a high-precision identity authentication method due to its advantages such as uniqueness, stability, and non-contact. Due to the light transmission characteristics of contact lenses, textured contact lenses can obtain textured iris images on iris sensors, thus hiding or even forging iris textures. Wearing textured contact lenses is currently the best-hidden and most difficult-to-detect iris forgery method, which will seriously affect the security of iris recognition systems. Therefore, detecting textured contact lens irises is an important link to ensure the security of biometric recognition.

[0003] Traditional iris anti-counterfeiting detection methods mainly detect based on the optical characteristics or texture features of the iris, and require manual design and extraction of feature descriptors, with poor real-time performance and accuracy. Currently, deep learning-based iris anti-counterfeiting detection networks require preprocessing of iris images and manual localization of key regions, affecting the real-time performance, accuracy, and generalization of detection. Summary of the Invention

[0004] Aiming at the problems in the existing deep learning-based iris anti-counterfeiting detection methods, the purpose of the present invention is to provide an iris image anti-counterfeiting detection method based on a cyclic attention mechanism, including the following steps:

[0005] S1, input the global iris image to be detected into the iris anti-counterfeiting detection network model;

[0006] Wherein, the iris anti-counterfeiting detection network model includes a cyclic network structure formed by sequentially connecting multi-level network structure units. Each level of the network structure unit is composed of a feature classification network and an attention network, and the feature classification network and the attention network within the same level of network structure unit share a feature extraction network. The feature classification network obtains the prediction probability that the image input to the network structure unit belongs to a forged iris image based on the features extracted by the feature extraction network, and the attention network locates the key region in the image input to the network structure unit based on the features extracted by the feature extraction network to obtain key region parameters;

[0007] S2, process the global iris image through the first-level network structure unit in the iris anti-counterfeiting detection network model to obtain the prediction probability and the key region parameters at the first-level feature scale;

[0008] S3. Perform the following operations through each network structure unit in the iris anti-counterfeiting detection network model: Based on the input image at the current feature scale input to the current-level network structure unit and the key region parameters at the current-level feature scale output by the current-level network unit, obtain the input image at the next-level feature scale, and have the next-level network structure unit process the input image at the next-level feature scale to obtain the prediction probability and the key region parameters at the next-level feature scale corresponding to the next-level network structure unit;

[0009] S4. Fuse the prediction probabilities at at least two different-level feature scales to obtain a determination result as to whether the global iris image is a real iris image or a forged iris image.

[0010] In some alternative embodiments, the feature classification network includes the feature extraction network and the first fully connected layer; the attention network includes the feature extraction network and the second fully connected layer.

[0011] In some alternative embodiments, the second fully connected layer includes a double-layer fully connected layer.

[0012] In some alternative embodiments, the feature extraction network is MobileNetV2.

[0013] In some alternative embodiments, the iris anti-counterfeiting detection network model includes three levels of the network structure units. Among them, the key region corresponding to the first-level network structure unit includes the iris region, and the key region corresponding to the second-level network structure unit includes the texture region of the iris.

[0014] In some alternative embodiments, the fusing the prediction probabilities at at least two different-level feature scales to obtain a determination result as to whether the global iris image is a real iris image or a forged iris image includes:

[0015] Taking the average value or weighted sum of the prediction probabilities at the second-level feature scale and the prediction probabilities at the third-level feature scale to obtain a fusion probability, and obtaining the determination result according to the fusion probability.

[0016] In some alternative embodiments, the structures of the feature classification networks in each level of the network structure units are the same, and the structures of the attention networks in each level of the network structure units are the same.

[0017] In some alternative embodiments, the key region parameters include the center coordinates and the side length of the key region.

[0018] In some alternative embodiments, obtaining the input image of the next feature scale based on the input image at the current feature scale input to the current-level network structure unit and the key region parameters of the current-level feature scale output by the current-level network unit includes:

[0019] Generating an image mask for locating the key region based on the key region parameters of the current-level feature scale, performing a cropping operation of multiplying the image mask and the input image at the current feature scale pixel by pixel, and upsampling the cropped regional image by bilinear interpolation to obtain the input image of the next feature scale.

[0020] In some alternative embodiments, the key region parameters include the center coordinates and the region side length of the key region; the image mask is generated in the following manner:

[0021] Obtaining the upper left corner coordinates (x tl , y tl ) and the lower right corner coordinates (x br , y br ) of the key region according to the center coordinates and the region side length of the key region.

[0022] Determining the mask M as:

[0023] M(·) = [σ(x - x tl ) - σ(x - x br )] · [σ(y - y tl ) - σ(y - y br )]

[0024] where σ(·) represents the sigmoid function.

[0025] The method provided in the above embodiments applies the cyclic attention mechanism for solving the fine-grained image classification problem to the detection of iris images to improve the detection accuracy of forged iris images, can perform unsupervised localization on the key regions of iris images that can distinguish real iris images and forged textures, and fuse multi-level scale features for discrimination, improving the real-time performance of detection. In some embodiments, a lightweight convolutional neural network is used in the feature extraction part, reducing the computing power and memory threshold of hardware devices, enabling the method to run on miniaturized and low-power devices (such as embedded edge computing devices), and better balancing the detection accuracy and the computing cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] By reading the detailed description of the non-limiting embodiments with reference to the following drawings, other features, objects, and advantages of the present application will become more apparent:

[0027] Figure 1Schematic flowchart of an embodiment of an iris image anti-counterfeiting detection method based on a cyclic attention mechanism.

[0028] Figure 2 Exemplary framework diagram of the iris anti-counterfeiting detection network model according to an embodiment of the present invention. In the figure, 201 is a feature classification network; 202 is an attention network; 203 is a network structure unit.

[0029] Figure 3 A, Figure 3 B, Figure 3 Schematic diagrams of the image center coordinates and regional side length position parameters of the iris region, iris texture region, and interpolated iris texture region in sequence.

[0030] Figure 4 A and Figure 4 B are respectively the schematic diagram of the image mask of the iris region and the schematic diagram of the image mask of the iris texture region. Detailed implementation manners

[0031] In order to enable those skilled in the art to better understand the solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0032] As Figure 1 shown, an embodiment of the present application provides an iris image anti-counterfeiting detection method based on a cyclic attention mechanism. This method can run on electronic devices with image / graphic processing capabilities such as computers, laptops, servers, etc. In this embodiment, the above method process includes the following steps:

[0033] S1, input the global iris image to be detected into the iris anti-counterfeiting detection network model.

[0034] In this embodiment, an iris anti-counterfeiting detection network model based on a cyclic attention mechanism is designed and trained.

[0035] The iris anti-counterfeiting detection network model includes a cyclic network structure formed by sequentially connecting multiple levels of network structure units. Each level of the network structure unit is composed of a feature classification network and an attention network, and the feature classification network and the attention network within the same level of network structure unit share a feature extraction network. The feature classification network obtains the prediction probability that the image input to the network structure unit belongs to a texture contact lens iris image based on the features extracted by the feature extraction network. The attention network locates the key region in the image input to the network structure unit based on the features extracted by the feature extraction network to obtain key region parameters.

[0036] Among them, the feature classification network and the attention network within the same-level network structure unit sharing the feature extraction network means that within a network structure unit, there is a feature extraction network, and the result of this feature extraction network is respectively output to the classification module of the feature classification network and the regression module of the attention network for further processing, and the obtained processing result is the output result of this level of network structure unit.

[0037] In some embodiments, the classification module in the feature classification network includes a first fully connected layer, that is, the feature classification network includes the feature extraction network and the first fully connected layer; the regression module of the attention network includes a second fully connected layer, that is, the attention network includes the feature extraction network and the second fully connected layer. In this way, probability prediction and calculation of key area parameters are respectively realized through the first fully connected layer and the second fully connected layer.

[0038] Optionally, the second fully connected layer in the attention network includes a double-layer fully connected layer.

[0039] The iris anti-counterfeiting detection network model utilizes the mutual correlation between region localization and feature extraction, and alternately trains the localization ability and extraction ability of the network through multiple cycles, so as to achieve the purpose of mutual reinforcement, and finally enables the network to focus on the tiny feature differences in the key area for detection.

[0040] S2. Process the global iris image through the first-level network structure unit in the iris anti-counterfeiting detection network model to obtain the prediction probability and the key area parameters at the first-level feature scale.

[0041] Input the global iris image to be detected into the first-level network structure unit in S1, and use the feature classification network and the attention network in the first-level network structure unit to process the input iris image to obtain the prediction probability that the input image at the first-level feature scale belongs to a forged iris, the center coordinates of the key area, and the side length of the area. Among them, the feature extraction network in the first-level network structure unit extracts features from the input image, the classification module in the feature classification network classifies the feature map extracted by the feature extraction network to obtain the prediction probability that the input image at the first-level feature scale belongs to a forged iris; the regression module in the attention network performs regression based on the feature map extracted by the feature extraction network to obtain the key area localization result at the first-level feature scale, including the center point coordinates of the key area and the side length of the key area.

[0042] S3. Perform the following operations through each network structure unit in the iris anti-counterfeiting detection network model: Based on the input image at the current feature scale input to the current-level network structure unit and the key region parameters at the current-level feature scale output by the current-level network unit, obtain the input image at the next-level feature scale, and let the next-level network structure unit process the input image at the next-level feature scale to obtain the prediction probability and the key region parameters at the next-level feature scale corresponding to the next-level network structure unit.

[0043] Taking the first-level network structure unit as an example, the specific implementation manner of S3 is described as follows:

[0044] The input image of this-level network structure unit is the global iris image. Locate the iris region according to the input image input to this-level network structure unit in step S2 and the center coordinates and region side lengths of the key region output by this-level network structure unit.

[0045] Specifically, according to the center coordinates and region side lengths (x t , y t , l t ) of the iris image at the current feature scale (i.e., the first-level feature scale), a rectangular region range (x tl , y tl , x br , y br ) can be obtained. Define a continuously differentiable mask function M by this range:

[0046] M(·) = [σ(x - x tl ) - σ(x - x br )]·[σ(y - y tl ) - σ(y - y br )]

[0047] where σ(·) represents the sigmoid function, and its value range is between 0 and 1. This mask function is close to 1 within the rectangular region range and close to 0 in other regions. Therefore, the input image of this-level network structure unit can be multiplied with it pixel by pixel to obtain the cropped iris region image.

[0048] After the first-level network structure unit completes the above processing steps, for the subsequent network structure units, perform the above-described processing steps level by level until after generating the input image of the last-level network structure unit of the entire network model, let the last-level network structure unit process the input image of this level to obtain the prediction probability at the last-level feature scale corresponding to the last-level network structure unit.

[0049] S4. Fuse the prediction probabilities of at least two different levels of feature scales to obtain a determination result on whether the global iris image is an image of a real iris or a forged iris.

[0050] By fusing the prediction probabilities output by multiple network structure units of different levels obtained in step S3, a fused probability can be obtained. Based on the relationship between the fused probability and a threshold, it is determined whether the global iris image is an image of a real iris or a forged iris. For example, a threshold can be set. When the fused probability is greater than the threshold, it is determined as an image of a real iris. When the fused probability is less than the threshold, it is determined as an image of a forged iris.

[0051] Optionally, the way to fuse the prediction probabilities output by the above-mentioned multiple network structure units of different levels can be to calculate the average value or weighted summation.

[0052] The method provided in the above embodiment applies the cyclic attention mechanism for solving the fine-grained image classification problem to the detection of iris images to improve the detection accuracy of forged iris images. It can perform unsupervised localization on the key regions of iris images that can distinguish real iris images from forged textures, and fuse multi-level scale features for discrimination, improving the real-time performance of detection.

[0053] In some optional embodiments, the iris anti-counterfeiting detection network model includes three levels of the network structure units. Among them, the key region corresponding to the first-level network structure unit includes the iris region, and the key region corresponding to the second-level network structure unit includes the texture region of the iris.

[0054] In this embodiment, step S3 can be specifically implemented in the following manner:

[0055] S31. Determine a mask function according to the center coordinates and the side length of the region of the key region located in step S2, and use this mask function to crop the global iris image input to the first-level network structure unit to obtain an iris region image as the input image of the second-level network structure unit;

[0056] S32. Input the input image obtained in S31 into the second-level network structure unit. The feature classification network and the attention network of the second-level network structure unit respectively obtain the prediction probability of being predicted as a forged iris at the second-level feature scale and the center coordinates and the side length of the region of the key region located at the second-level feature scale;

[0057] S33. Determine a mask function according to the center coordinates and the side length of the region of the key region in step S32, and use this mask function to process the input image (iris region image) input to the second-level network result unit to obtain an iris texture region image.

[0058] S34. Input the iris texture region image obtained in step S32 into the third-level network structure unit to obtain the prediction probability at the third-level feature scale.

[0059] S35. Average or perform weighted summation on the prediction probabilities at the second-level (iris region) and third-level (iris texture region) feature scales to obtain a fusion probability, and determine the authenticity discrimination result of the iris image according to the relationship between the fusion probability and a preset threshold.

[0060] The above embodiments achieve adaptive iris texture region localization by gradually locating the iris region and iris texture region from the original global iris image, and adopt a relatively concise model structure, reducing the computing power requirements for the device running this method, enabling the method to be deployed on lightweight devices.

[0061] The following further describes this embodiment in conjunction with the accompanying drawings.

[0062] See Figure 2 、 Figure 3 and Figure 4 , as described in detail below:

[0063] Figure 2 It is a schematic structural diagram of an iris anti-counterfeiting detection network model in the iris anti-counterfeiting detection method provided by an embodiment of the present invention.

[0064] See Figure 2 As shown, the iris anti-counterfeiting detection network model includes a recurrent network structure formed by multiple levels of network structure units 203. Each level of network structure unit consists of a feature classification network 201 and an attention network 202. The feature classification networks at different levels of scales may have the same structure but independent parameters; the attention networks at different levels may have the same structure but independent parameters. Among them, the feature classification network includes a feature extraction network, a fully connected layer, and a Softmax layer; the attention network includes a feature extraction network and two fully connected layers.

[0065] It should be noted that referring to the feature extraction network shown in Table 1 below, in the present invention, the feature extraction network uses the lightweight feature extraction network MobileNetV2, which uses depthwise separable convolution for feature extraction. The feature layer consists of Conv2d and Bottleneck, where Bottleneck uses an inverted residual structure, which can not only increase the reuse rate of features but also improve the calculation efficiency. Therefore, using MobileNetV2 as the feature extraction network significantly reduces the number of network parameters (Params) and the amount of calculation (FLOPs), can reduce the operation cost of the anti-counterfeiting detection network, and is more conducive to the integrated deployment of the anti-counterfeiting detection network in the iris recognition system.

[0066] Table 1: Schematic diagram of the feature layer structure of the MobileNetV2 feature extraction network.

[0067]

[0068] The method of the above embodiments of the present application can be applied to the detection of images of forged irises wearing texture contact lenses. In some embodiments, the above method may further include the training step of the above iris anti-counterfeiting detection network model. Specifically, the training step includes:

[0069] First, establish an experimental data set. The experiment selects two public databases, IIITD CLI and ND series, which contain real iris samples and contact lens iris samples, to train and test the network. The ethnic distribution of the data samples covers Europe, America and Asia, and the contact lenses come from different brands such as Johnson & Johnson, CIBA Vision, and Bausch & Lomb, ensuring the diversity of the data samples. The IIITDCLI database is provided by the Indian Institute of Technology Delhi and contains 6,570 sample images from 101 experimenters, with an image resolution of 640×480 pixels. The ND series database comes from the Computer Vision Laboratory of the University of Notre Dame in the United States, with a total of 18,196 sample images, and the image resolution is also 640×480 pixels. Normalize the sample images of each data set to 448x448 pixels, and then randomly divide the data set into a training set and a test set according to a ratio of 7:3.

[0070] Secondly, set training parameters and evaluation metrics. In the training stage, first load MobileNetV2 pre-trained on ImageNet to initialize the feature classification network in the multi-level network structure unit of the network model, and then use the batch stochastic gradient descent algorithm with momentum (Batch Stochastic Gradient Descent, BSGD) to update the network parameters. During the training process, set the initial value of the learning rate to 0.001; the momentum factor to 0.9; the weight decay to 0.0005; the batch size to 32; and the number of iterations to 50. In the test stage, use the correct classification rate (Correct ClassificationRate, CCR) and the receiver operating characteristic curve (Receiver Operating Characteristic, ROC) as evaluation metrics for measuring the detection accuracy, and use the number of network parameters Params and the number of floating-point operations per second FLOPs as evaluation metrics for measuring the operation cost.

[0071] Then, define the network loss function. The network is optimized using two different loss functions: the classification loss function (Classification Loss) is used for the feature classification network; the pairwise ranking loss function (PairwiseRanking Loss) is used for the attention network. The overall loss function of the network is defined as follows:

[0072]

[0073] where X t represents the t-th iris image; s represents the s-th level of feature scale; L cls (·) represents the classification loss function; L rank (·) represents the ranking loss function. The classification loss is calculated using the cross-entropy loss function Cross Entropy. The calculation formula of L cls (·) is as follows:

[0074] L cls (Y t * ,Y t s ) = [Y t * ·lnY t s +(1 - Y t * )·ln(1 - Y t s )]

[0075] where Y t s represents the predicted value of the sample by the s-th level feature classification network, and Y t * represents the label value of the sample. Using the cross-entropy loss function can avoid the problem of gradient dispersion in the network and make the prediction result of the feature classification network effectively converge to the true label. The ranking loss is calculated by the prediction error of the true label by the feature classification networks of two adjacent levels. The calculation formula of L rank (·) is as follows:

[0076]

[0077] where and represent the predicted probabilities of the true label by the s-th and (s + 1)-th level feature classification networks respectively. margin represents the interval difference, which is set to 0.05. When , the loss is smaller, indicating that the feature region location located by the attention network is more conducive to detecting the subtle feature differences between genuine and fake irises, and can make the prediction of the next-level feature classification network more accurate.

[0078] Finally, according to the above set conditions, the parameters of each layer of the attention network are fixed, and the feature classification network in the multi-level network structure unit is trained until convergence; then the parameters of each layer of the feature classification network are fixed, and the attention network is trained until convergence. Iterative alternation is performed in each training round until the losses of both networks converge.

[0079] See Figure 3 As shown, in the embodiment of the present invention, specifically, an implementation manner of obtaining the input image of the next-level feature scale based on the input image at the current feature scale input to the current-level network structure unit and the key region parameters of the current-level feature scale output by the current-level network unit is as follows:

[0080] Based on the key region parameters of the current-level feature scale, an image mask for locating the key region is generated, a cropping operation of multiplying the image mask and the input image at the current feature scale pixel by pixel is performed, and the input image of the next-level feature scale is obtained by bilinear interpolation upsampling of the cropped regional image.

[0081] Taking the first-level network structure unit as an example, specifically, first, the attention network maps the feature map generated by the feature extraction network into the position parameters x t , y t , l t (t = 1, 2,..., n, representing the t-th iris image), where x t , y t are the center coordinates of the corresponding regions respectively, and l t is half of the side length of the region.

[0082] Secondly, according to the region position parameters, the top-left coordinates (x tl , y tl ) and the bottom-right point coordinates (bottom-right) (x br , y br ) of the iris region are obtained, and the calculation formulas are as follows:

[0083] x tl = x t - l t , y tl = y t - l t

[0084] x br = x t + l t , y br = y t + l t

[0085] Then, according to the vertex coordinates of the iris region, an image mask M is defined, and M is multiplied pixel by pixel with the iris image X t to perform a cropping operation, and the cropped iris region image X t ' can be obtained. Similarly, the cropped iris texture region image X” is obtained by using the second-level attention network t , and the calculation formula is as follows:

[0086] X′ t = X t ⊙ M(x tl , y tl , x br , y br )

[0087] Finally, after bilinear interpolation upsampling of the cropped iris region image and the iris texture region image, iris images with different feature scales (i.e., input images of different-level network structure units) are obtained.

[0088] See Figure 4 shown. It should be noted that the image mask M used for cropping the image is a continuously differentiable mask function to ensure that backpropagation can be performed during the network training optimization. The definition formula is as follows:

[0089] M(·) = [σ(x - x tl ) - σ(x - x br )] · [σ(y - y tl ) - σ(y - y br )]

[0090] where σ(·) represents the sigmoid function, and its value range is between 0 and 1. Only the pixel points located within the iris region (texture region) (satisfying that x is between x tl and x br , and y is between y tl and y br ) can make the result of M(·) approach 1, and the results of other pixel points approach 0, thus generating an approximate binary image mask.

[0091] For the embodiments of the present invention, by retaining the prediction probabilities of the true labels at the iris region scale and the iris texture region scale, and fusing multi-level scale features, a more accurate discrimination result of whether the iris image to be detected is a texture contact lens iris or not is obtained. Experimental verification is carried out on two public databases (IIITD CLI and ND series) containing real iris samples and contact lens iris samples. The results show that the detection accuracy of RAINet is better than that of other anti-counterfeiting detection networks, and the average correct classification rates under different experimental conditions such as the same sensor, cross-sensor, and cross-database reach 99.93%, 97.31%, and 97.86% respectively.

[0092] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0093] 1. The present application provides an iris anti-counterfeiting detection method based on a cyclic attention mechanism, which introduces the cyclic attention mechanism to perform key region localization at multiple levels in sequence. For example, the iris region and the iris texture region of the original iris image are unsupervised located in sequence, enabling the network to continuously focus on the subtle feature differences between the key regions of genuine and fake irises by simulating the visual characteristics of the human eye, and having better accuracy and generalization compared with other iris anti-counterfeiting detection networks.

[0094] 2. MobileNetV2 is used to lightweight the feature classification network, solving the problem of excessive computing cost caused by using a cyclic network structure; in addition, constructing an end-to-end network does not require complex image preprocessing or manual annotation, is suitable for integration into an iris recognition system, and can be deployed on edge computing devices.

[0095] The above are only the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make various improvements or refinements, which do not affect the essence of the present invention.

Claims

1. An iris image anti-counterfeiting detection method based on a cyclic attention mechanism, characterized in that Including the following steps: Input the global iris image to be detected into the iris anti-counterfeiting detection network model. The iris anti-counterfeiting detection network model includes a recurrent network structure formed by sequentially connecting multiple levels of network structure units. Each level of the network structure unit is composed of a feature classification network and an attention network, and the feature classification network and the attention network within the same level of network structure unit share a feature extraction network. The feature classification network obtains the prediction probability that the image input to the network structure unit belongs to a forged iris image based on the features extracted by the feature extraction network, and the attention network locates the key region in the image input to the network structure unit based on the features extracted by the feature extraction network to obtain the key region parameters; Process the global iris image through the first-level network structure unit in the iris anti-counterfeiting detection network model to obtain the prediction probability and the key region parameters at the first-level feature scale; Each level of network structure unit in the iris anti-counterfeiting detection network model performs the following operations: Based on the input image at the current-level feature scale input to the current-level network structure unit and the key region parameters at the current-level feature scale output by the current-level network structure unit, obtain the input image at the next-level feature scale, and have the next-level network structure unit process the input image at the next-level feature scale to obtain the prediction probability and the key region parameters at the next-level feature scale corresponding to the next-level network structure unit; Fuse the prediction probabilities at at least two different-level feature scales to obtain the determination result of whether the global iris image is a real iris image or a forged iris image; Among them, obtaining the input image at the next-level feature scale based on the input image at the current-level feature scale input to the current-level network structure unit and the key region parameters at the current-level feature scale output by the current-level network structure unit includes: Generating an image mask for locating the key region based on the key region parameters at the current-level feature scale, performing a cropping operation of multiplying the image mask and the input image at the current-level feature scale pixel by pixel, and performing bilinear interpolation upsampling on the cropped regional image to obtain the input image at the next-level feature scale; The key region parameters include the center coordinates and the region side length of the key region; the image mask is generated in the following manner: Obtain the upper-left corner coordinates (x tl , y tl ) and the lower-right corner coordinates (x br , y br ) of the key area according to the center coordinates and the side length of the key area, and determine the mask M as: M(·) = [σ(x - x tl ) - σ(x - x br )]·[σ(y - y tl ) - σ(y - y br )] Where σ(·) represents the sigmoid function, and (x, y) represents the coordinates of the pixel points in the input image.

2. The method according to claim 1, wherein The feature classification network includes the feature extraction network and a first fully connected layer; the attention network includes the feature extraction network and a second fully connected layer.

3. The method according to claim 2, wherein The second fully connected layer includes a double-layer fully connected layer.

4. The method according to claim 1, wherein The feature extraction network is MobileNetV2.

5. The method according to claim 1, characterized in that, The iris anti-counterfeiting detection network model includes three levels of network structure units. Among them, the key region corresponding to the first-level network structure unit includes the iris region, and the key region corresponding to the second-level network structure unit includes the texture region of the iris.

6. The method according to claim 5, characterized in that, Fusing the prediction probabilities of at least two different levels of feature scales to obtain a determination result on whether the global iris image is a real iris image or a forged iris image includes: Taking the average value or weighted sum of the prediction probability of the second-level feature scale and the prediction probability of the third-level feature scale to obtain a fusion probability, and obtaining the determination result according to the fusion probability.

7. The method according to claim 1, characterized in that, The structures of the feature classification networks in each level of the network structure units are the same, and the structures of the attention networks in each level of the network structure units are the same.

8. The method according to claim 1, characterized in that The key area parameters include the central coordinates of the key area and the side length of the area.

Images