Iris image generation method and system
By generating adversarial network training, the generator learns to synthesize iris images, solving the problem of insufficient data volume of iris database, generating realistic iris images, and improving the performance of iris recognition.
Patent Information
- Application Number
- CN202211177298.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-09-26
- Publication Date
- 2025-08-29
- Estimated Expiration
- 2042-09-26
AI Technical Summary
In the prior art, the data volume of iris image database is insufficient and cannot support large-scale training of deep learning models.
By generating adversarial network training, the generator learns to synthesize realistic iris images, uses multiple vector groups of hidden space Z to generate intra-class and inter-class images, and generates adversarial network parameters through comparison loss and adversarial loss optimization to establish a large-scale iris database.
A large number of realistic iris images are generated, improving the performance of iris recognition, especially in identity control and recognition capabilities.
Smart Images
Figure CN115620080B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image processing technology, and in particular to an iris image generation method and system. Background Art
[0002] Iris recognition has the advantages of high uniqueness, strong stability, good anti-counterfeiting, and non-contact, and is considered to be one of the most promising biometric recognition technologies. In the past few decades, the rapid development of deep learning has promoted the research work on iris recognition. It is well known that deep learning models require a large amount of data during the training process, and the lack of training data will have a negative impact on the performance of the model. However, due to the hidden and private nature of the iris, collecting a large-scale iris database requires a lot of manpower and financial resources. The public iris datasets currently used in the academic field are relatively small (the largest public database, CASIA-IrisV4-Thousand
[17] , only contains a total of 20,000 iris images from 1,000 subjects), which is not enough to support the large-scale training of deep learning models. Therefore, establishing a large-scale iris database is an urgent problem to be solved. Summary of the Invention
[0003] The present invention provides an iris image generation method and system, which solves the problem of insufficient data volume in the iris image database in the related art.
[0004] The technical solutions of the present invention are as follows:
[0005] In a first aspect, a method for generating an iris image includes:
[0006] Obtain the latent space Z, which includes multiple vector groups, each of which includes vector Z i , vector Z o and vector Z t , vector Z i , vector Z o and vector Z t are all random arrays sampled from Gaussian distribution;
[0007] Perform multiple GAN training operations until the objective function is within the set range; each GAN training operation includes:
[0008] Multiple vector groups are input into the generator to obtain multiple generated images; the multiple generated images are divided into intra-class images and inter-class images. The intra-class images are: vector Z i and vector Z t The images generated by multiple vector groups are all the same, and the inter-class images are: vector Z i and vector Z t Images generated by multiple vector groups that are all different;
[0009] Input every two intra-class images and every two inter-class images into the recognition network respectively, obtain the matching probability of every two intra-class images and every two inter-class images, and calculate the contrast loss based on the matching probability of every two intra-class images and every two inter-class images;
[0010] Input multiple generated images and real images into the discriminator respectively and calculate the adversarial loss;
[0011] The objective function is calculated based on the contrast loss and adversarial loss to optimize the parameters of the generative adversarial network.
[0012] In a second aspect, an iris image generation system includes:
[0013] Obtain a unit for the latent space Z, which includes multiple vector groups, each of which includes vector Z i , vector Z o and vector Z t , vector Z i , vector Z o and vector Z t are all random arrays sampled from Gaussian distribution;
[0014] An execution unit is configured to execute multiple GAN training operations until the objective function is within a set range. Each GAN training operation includes:
[0015] Multiple vector groups are input into the generator to obtain multiple generated images; the multiple generated images are divided into intra-class images and inter-class images. The intra-class images are: vector Z i and vector Z t The images generated by multiple vector groups are all the same, and the inter-class images are: vector Z i and vector Z t Images generated by multiple vector groups that are all different;
[0016] Input every two intra-class images and every two inter-class images into the recognition network respectively, obtain the matching probability of every two intra-class images and every two inter-class images, and calculate the contrast loss based on the matching probability of every two intra-class images and every two inter-class images;
[0017] Input multiple generated images and real images into the discriminator respectively and calculate the adversarial loss;
[0018] The objective function is calculated based on the contrast loss and adversarial loss to optimize the parameters of the generative adversarial network.
[0019] The working principle and beneficial effects of the present invention are:
[0020] The present invention obtains multiple vector groups from the latent space Z and inputs the multiple vector groups into the generator respectively, and the generator learns how to synthesize realistic images. i and Z t The images generated by the same vector groups are called intra-class images, Z i and Z t Images generated by multiple sets of vectors that are all different are called inter-class images. Every two intra-class images and every two inter-class images are fed into the recognition network. A contrastive loss is calculated based on the matching probability between the two intra-class images or the two inter-class images. Multiple generated images and real images are fed into the discriminator to calculate the adversarial loss. The objective function is calculated based on the contrastive and adversarial losses to optimize the parameters of the generative adversarial network. Training stops when the generator can generate intra-class images with high matching probabilities and inter-class images with low matching probabilities.
[0021] For ease of understanding, we define vector Z i , vector Z o and vector Z t They control the attributes related to identity (such as left and right eyes), attributes that vary with conditions (such as pupil size, iris exposure ratio) and texture, respectively, with the same vector Z i and vector Z t Multiple images belong to the same identity, with different vectors Z i and vector Z t Multiple images belong to different identities.
[0022] The present invention can generate multiple iris images based on multiple vector groups in the latent space Z to help establish a large-scale synthetic iris database. At the same time, when training the generator, the identity vector is decoupled from the latent space Z through comparative learning of intra-class images and inter-class images, which facilitates identity control of the synthesized iris images. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0024] Figure 1 This is a flow chart of the iris image generation method of the present invention;
[0025] Figure 2 This is a schematic diagram of the training block diagram of the generative adversarial network in the present invention;
[0026] Figure 3 Schematic diagram of the generator structure in the present invention;
[0027] Figure 4 This is a schematic diagram of an iris image generated using the method of this embodiment;
[0028] Figure 5 A comparison of iris images synthesized by the method of this embodiment and the existing method;
[0029] Figure 6 The comparison chart of the separation results of iris topology and texture channels;
[0030] Figure 7 Schematic diagram of the comparison loss between a single-channel input generator and a dual-channel input generator of this embodiment;
[0031] Figure 8 This is a structural diagram of the iris image generation system of the present invention. DETAILED DESCRIPTION
[0032] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.
[0033] Example 1
[0034] like Figure 1 FIG. 1 is a flow chart of the iris image generation method according to this embodiment, including:
[0035] S100: Obtain a latent space Z, which includes multiple vector groups, each of which includes vector Z i , vector Z o and vector Z t , vector Z i , vector Z o and vector Z t are all random arrays sampled from Gaussian distribution;
[0036] Specifically, by calling the random number generation function random() and specifying the corresponding dimension, the random vector Z can be obtained. i , vector Z o and vector Z t .
[0037] S200: Execute multiple generative adversarial network training operations until the objective function is within a set range; each generative adversarial network training operation includes:
[0038] S210: Figure 2 As shown, multiple vector groups are input into the generator to obtain multiple generated images; multiple generated images are divided into intra-class images and inter-class images. The intra-class image is: vector Z i and vector Z tThe images generated by multiple vector groups are all the same, and the inter-class images are: vector Z i and vector Z t Images generated by multiple vector groups that are all different;
[0039] The generator structure is as follows Figure 3 As shown in the figure, it includes multiple convolutional layers connected in sequence. From input to output, the size of the convolutional layers increases continuously, starting from 4×4 and continuously increasing to 256×256. Correspondingly, it is possible to gradually generate a high-resolution 256×256 image from a low-resolution 4×4 image.
[0040] S220: Inputting every two intra-class images and every two inter-class images into the recognition network respectively, obtaining the matching probability of every two intra-class images and every two inter-class images, and calculating the contrast loss based on the matching probability of every two intra-class images and every two inter-class images;
[0041] like Figure 2 As shown, in this embodiment, the recognition network first adopts the ResNet50 network structure to extract the 128-dimensional features of any two intra-class images or any two inter-class images respectively, and then merges the two features to obtain the matching probability through multi-layer convolution.
[0042] S230: Input multiple generated images and real images into the discriminator respectively, and calculate the adversarial loss;
[0043] S240: Calculate the objective function based on the contrast loss and the adversarial loss, and optimize the parameters of the generative adversarial network.
[0044] In this invention, by inputting multiple vector groups into the generator, the generator learns how to synthesize images to maximize the deception of the discriminator. In the generative adversarial network training, the parameters of the generator and the discriminator are updated alternately until the generated image is difficult to distinguish from the real image. At the same time, the vector Z i and Z t The images generated by the same vector groups are called intra-class images, Z i and Z t Images generated by multiple different vector groups are called inter-class images. Every two intra-class images and every two inter-class images are input into the recognition network respectively, and the matching probability is calculated. The contrast loss is calculated based on the matching probability. If the loss function is within the set range, it indicates that the training is completed and the generator is able to generate intra-class images with high matching probability and inter-class images with low matching probability.
[0045] For ease of understanding, we define vector Z i , vector Z o and vector Z tThey control the attributes related to identity (such as left and right eyes), attributes that vary with conditions (such as pupil size, iris exposure ratio) and texture, respectively, with the same vector Z i and vector Z t Multiple images belong to the same identity, with different vectors Z i and vector Z t Multiple images belong to different identities.
[0046] Existing iris generation methods lack constraints on identity information and cannot synthesize intra-class images with specific identities. Therefore, these synthesized images are mainly used to present attacks and defenses rather than for iris recognition. This invention generates multiple iris images through multiple vector groups in the latent space Z, helping to build a large-scale synthetic iris database. Furthermore, during generator training, the identity vector is decoupled from the latent space Z through comparative learning between intra-class and inter-class images, facilitating identity control of the synthesized iris images and improving iris recognition performance.
[0047] Furthermore, for any vector group, the steps of generating an image include:
[0048] S211: vector Z in any vector group i and vector Z o Merge and input to each convolution layer of the generator; the output of each convolution layer and vector Z t Add them together as the input feature map of the next convolutional layer;
[0049] Taking the 8×8 convolution layer as an example, the weights of the 8×8 convolution layer are Z i 、Z o After adjustment, it is convolved with the input feature map (the output of the 4×4 convolution layer); the output of the convolution and Z t Add them together as the input feature map of the 16×16 convolutional layer.
[0050] S212: The output of the last convolutional layer is used as the generated image corresponding to any vector group.
[0051] The size of the last convolutional layer is determined according to the actual required image resolution. In this embodiment, an image with an output resolution of 256×256 is required, and the last convolutional layer is set to 256×256.
[0052] Vector Z i and vector Z o As a channel, through each layer of the mapping network input generator, they affect the overall composition of the generated image; vector Z tIt is added as a channel after each layer of the generator to generate fine texture details. With the dual-channel input of structure and texture, the generator can infer iris images under multiple conditions while maintaining unique texture details.
[0053] Furthermore, in step S220, the contrast loss is calculated based on the matching probability between every two intra-class images and every two inter-class images, specifically including:
[0054] Contrastive loss
[0055] Among them, I n =G(z n ) represents the input vector group The generated image; P+ and P- represent the set of all intra-class vector pairs and inter-class vector pairs, respectively. The intra-class vector pairs have the same vector Z i and vector Z t For any two vector groups, the inter-class vector pair is vector Z i and vector Z t Any two vector groups that are different; E[] is the expected value, V(I m ,I n ) represents image I m ,I n The matching probability.
[0056] Furthermore, in step S230, the generated images and the real images are input into the discriminator respectively to calculate the adversarial loss, which specifically includes:
[0057] Fighting Losses
[0058] Among them, I r represents the real iris image, G(z) represents the generated image, and D represents the discriminator.
[0059] Furthermore, in step S240, the objective function is calculated based on the contrast loss and the adversarial loss, specifically including:
[0060] Objective function
[0061] Where G represents the generator and λ is the weight of the contrast loss.
[0062] like Figure 2 As shown in , we use contrastive learning to decouple the identity vector. Specifically, in this embodiment, multiple batches of training are performed, and pairs of intra-class and inter-class latent vectors are sampled in each training batch. The intra-class latent vectors share the same z i and z t Let I n =G(z n ) represents the input latent vector The generated images, P+, P- respectively represent the set of all intra-class and inter-class latent vector pairs in a batch. The contrast loss of the training batch is defined as:
[0063]
[0064] like Figure 2 As shown, z i 1=z i 2, z t 1=z t 2, images I1 and I2 form an intra-class image pair, and z i 2≠z i 3, z t 2≠z t 3, images I2 and I3 form an inter-class image pair. The contrastive loss will penalize the difference between the identities of I1 and I2, and the similarity between the identities of I2 and I3.
[0065] In order to improve the realism of generated images, we adopt adversarial loss, which is formulated as:
[0066]
[0067] Among them I r represents the real iris image, and D represents the discriminator.
[0068] Our overall objective function can be summarized as follows:
[0069]
[0070] The discriminator is updated to increase the loss, while the generator is updated to decrease the loss. λ is the weight of the contrastive loss. In this example, a total of 1.4 million generative adversarial network training operations were performed. In each training run, the generator produced an equal number of within-class and between-class images. At the beginning of training, the contrastive loss was disabled, meaning the weight λ was set to 0. After 400,000 training runs, λ was set to 1.
[0071] The following describes the experimental process:
[0072] 1. Dataset
[0073] In this embodiment, the iris dataset we collected contains a total of 30,726 high-quality iris images from 5,000 subjects. Each subject has at least three images of the left eye and the right eye respectively. The resolution of these images is 640×480, and they are reduced to 256×256 after positioning and center cropping.
[0074] 2. Quality of the synthesized image
[0075] The iris image generated by the method of this embodiment is as follows Figure 4 As shown in Figure 2, it can be seen that these synthetic images are visually realistic with clear textures. Multiple images in the same row are intra-class images and have the same identity information. Figure 5 The following is a comparison of iris images obtained by the existing method and this embodiment. Among them, images [4] and
[20] are obtained using traditional algorithms, and images
[14] ,
[16] ,
[25] , and
[26] are obtained using a method based on generative adversarial networks. It can be seen that images [4] and
[20] are visually unrealistic, and images
[14] ,
[16] ,
[25] , and
[26] are blurry. The method of this embodiment surpasses the existing method.
[0076] 3. Iris recognition
[0077] Through contrastive learning, the method of this embodiment can decompose identity vectors and generate intra-class images and inter-class images. To verify whether the generated images can be used for iris recognition, we conducted a validation experiment. First, we selected high-quality generated images and established a synthetic database consisting of 187,717 images, 10,000 classes (identities), and 10-20 images per class. We then trained the ResNet50 network using three types of data: a) real data, b) synthetic data, and c) real and synthetic data.
[0078] The experimental results are shown in Table 1. Because the scale of the synthetic data is over ten times that of the real data, the error metrics for training the recognition network using synthetic data are significantly lower than those using real data. Furthermore, training with both real and synthetic data further improves the training effect. These experimental results demonstrate that using the method in this embodiment to generate large-scale synthetic datasets can significantly improve iris recognition performance.
[0079] Table 1 Comparison of iris recognition errors
[0080] Data Type Training set size FRR@FAR=0.1 EER a) Real data 15,595 images, 5,000 categories 1.96％ 3.52％ b) Synthetic Data 187,717 images, 10,000 categories 0.29％ 1.92％ c) Real and synthetic data 203,312 images, 15,000 categories 0.14％ 1.24％
[0081] 4. Ablation Experiment
[0082] Because the iris has a multi-degree-of-freedom topological structure and high-entropy texture, traditional generators, with only a single input channel, struggle to infer multi-conditional iris images while preserving unique texture details. This embodiment employs two channels to separate the iris' topology and texture. Part of the latent vector is input to each layer of the generator to influence the overall composition of the synthesized image, while the other part is added after each layer to specifically generate texture. This mechanism allows the generator to generate multi-conditional iris images without destroying texture information.
[0083] Figure 6The following figure shows a comparison of the separation results of the iris topology and texture channels. Figures (a), (b), (c), and (d) show four sets of images, with the texture channel disabled on the left and enabled on the right. As can be seen, the texture channel can generate fine texture details in specific areas without affecting the overall composition. With the combined control of these two channels, the method in this embodiment has excellent identity representation capabilities.
[0084] To verify the superiority of the dual-channel mechanism of this embodiment, we deleted the texture channel added after each layer of the generator, used the same experimental settings, and retrained the generator. The comparison loss between this generator and the generator of this embodiment is as follows: Figure 7 As shown in , it can be seen that the contrast loss of the dual-channel decreases faster than that of the single-channel input, which indicates that the dual-channel input mechanism of this embodiment is beneficial to identity representation.
[0085] 5. Summary
[0086] The method in this example can generate a large number of realistic and clearly textured iris images. The comparative learning of intra-class and inter-class images, as well as the dual-channel input mechanism, ensure that our model has excellent identity representation capabilities. Experimental results show that large-scale synthetic databases are beneficial for improving iris recognition performance.
[0087] Example 2
[0088] Based on the same inventive concept as the first embodiment above, this example proposes an iris image generation system, including:
[0089] Obtain a unit for the latent space Z, which includes multiple vector groups, each of which includes vector Z i , vector Z o and vector Z t , vector Z i , vector Z o and vector Z t are all random arrays sampled from Gaussian distribution;
[0090] An execution unit is configured to execute multiple GAN training operations until the objective function is within a set range. Each GAN training operation includes:
[0091] Multiple vector groups are input into the generator to obtain multiple generated images; the multiple generated images are divided into intra-class images and inter-class images. The intra-class images are: vector Z i and vector Z t The images generated by multiple vector groups are all the same, and the inter-class images are: vector Z i and vector Z t Images generated by multiple vector groups that are all different;
[0092] Input every two intra-class images and every two inter-class images into the recognition network respectively, obtain the matching probability of every two intra-class images and every two inter-class images, and calculate the contrast loss based on the matching probability of every two intra-class images and every two inter-class images;
[0093] Input multiple generated images and real images into the discriminator respectively and calculate the adversarial loss;
[0094] The objective function is calculated based on the contrast loss and adversarial loss to optimize the parameters of the generative adversarial network.
[0095] Furthermore, it also includes:
[0096] The first processing unit is used to process any vector group, the vector Z in the vector group i and vector Z o Merge and input to each convolution layer of the generator; the output of each convolution layer and vector Z t Add them together as the input feature map of the next convolutional layer;
[0097] The output of the last convolutional layer is used as the generated image corresponding to any vector group.
[0098] Furthermore, it also includes:
[0099] The second processing unit extracts features of any two intra-class images or any two inter-class images to obtain feature pairs;
[0100] The feature pairs are merged and calculated through multiple layers of convolution to obtain the matching probability.
[0101] Furthermore, it also includes:
[0102] The first calculation unit is used to calculate the contrast loss based on the matching probability of every two intra-class images and every two inter-class images, specifically including:
[0103] Contrastive loss
[0104] Among them, I n =G(z n ) represents the input vector group The generated image; P+ and P- represent the set of all intra-class vector pairs and inter-class vector pairs, respectively. The intra-class vector pairs have the same vector Z i and vector Z t For any two vector groups, the inter-class vector pair is vector Z i and vector Z t Any two vector groups that are different; E[] is the expected value, V(I m ,I n ) represents image I m ,In The matching probability.
[0105] Furthermore, it also includes:
[0106] The second calculation unit is used to input multiple generated images and real images into the discriminator respectively and calculate the adversarial loss, which specifically includes:
[0107] Fighting Losses
[0108] Among them, I r represents the real iris image, G(z) represents the generated image, and D represents the discriminator.
[0109] Furthermore, it also includes:
[0110] The third calculation unit is used to calculate the objective function based on the contrast loss and the adversarial loss, specifically including:
[0111] Objective function
[0112] Where G represents the generator and λ is the weight of the contrast loss.
[0113] The various variations and specific examples of the iris image generation method in the aforementioned embodiment 1 are also applicable to the iris image generation system of this embodiment. Based on the detailed description of the aforementioned iris image generation method, those skilled in the art can clearly understand the implementation method of the iris image generation system of this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.
[0114] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. Iris image generation method based on generative adversarial network, which includes a generator and a discriminator, characterized by: include: Obtain the latent space Z, which includes multiple vector groups, each of which includes vector Z i , vector Z o and vector Z t , vector Z i , vector Z o and vector Z t are all random arrays sampled from Gaussian distribution; Perform multiple GAN training operations until the objective function is within the set range; Any generative adversarial network training operation includes: Multiple vector groups are input into the generator to obtain multiple generated images; the multiple generated images are divided into intra-class images and inter-class images. The intra-class images are: vector Z i and vector Z t The images generated by multiple vector groups are all the same, and the inter-class images are: vector Z i and vector Z t Images generated by multiple vector groups that are all different; Input every two intra-class images and every two inter-class images into the recognition network respectively, obtain the matching probability of every two intra-class images and every two inter-class images, and calculate the contrast loss based on the matching probability of every two intra-class images and every two inter-class images; Input multiple generated images and real images into the discriminator respectively and calculate the adversarial loss; Calculate the objective function based on contrast loss and adversarial loss to optimize the parameters of the generative adversarial network; The generator consists of multiple convolutional layers connected in sequence. The size of the convolutional layers increases from input to output. Multiple vector groups are input into the generator to obtain multiple generated images, including: For any vector group, the vector Z in the vector group i and vector Z o Merge and input to each convolution layer of the generator; the output of each convolution layer and vector Z t Add them together as the input feature map of the next convolutional layer; The output of the last convolutional layer is used as the generated image corresponding to any vector group; Among them, vector Zi, vector Zo and vector Zt respectively control identity-related attributes, condition-varying attributes and texture. Multiple images with the same vector Zi and vector Zt belong to the same identity, and multiple images with different vector Zi and vector Zt belong to different identities.
2. The iris image generation method according to claim 1, wherein: The recognition network is based on the ResNet50 network structure and calculates the matching probability between every two images within a class and every two images between classes. Specifically, it includes: For any two intra-class images or any two inter-class images, extract the features of the two intra-class images or the two inter-class images to obtain a feature pair; The feature pairs are merged and calculated through multiple layers of convolution to obtain the matching probability.
3. The iris image generation method according to claim 1, wherein: The contrast loss is calculated based on the matching probability of every two intra-class images and every two inter-class images, specifically including: Contrastive loss ; Among them, I n =G(z n ) represents the input vector group z n =(z i n , z o n , z t n ) generated image; P+, P- respectively represent the set of all intra-class vector pairs and inter-class vector pairs, and the intra-class vector pairs are those with the same vector Z i and vector Z t For any two vector groups, the inter-class vector pair is vector Z i and vector Z t Any two vector groups that are different; E[] is the expected value, V(I m ,I n ) represents image I m ,I n The matching probability.
4. The iris image generation method according to claim 3, wherein: The process of inputting multiple generated images and real images into the discriminator and calculating the adversarial loss specifically includes: Fighting Losses ; Among them, I r represents the real iris image, G(z) represents the generated image, and D represents the discriminator.
5. The iris image generation method according to claim 4, characterized in that: The objective function is calculated based on the contrast loss and the adversarial loss, specifically including: Objective function ; Where G represents the generator and λ is the weight of the contrast loss.
6. An iris image generation system, implementing the iris image generation method according to any one of claims 1 to 5, characterized in that: include: Obtain a unit for the latent space Z, which includes multiple vector groups, each of which includes vector Z i , vector Z o and vector Z t , vector Z i , vector Z o and vector Z t are all random arrays sampled from Gaussian distribution; An execution unit, used to perform multiple generative adversarial network training operations until the objective function is within a set range; Any generative adversarial network training operation includes: Multiple vector groups are input into the generator to obtain multiple generated images; the multiple generated images are divided into intra-class images and inter-class images. The intra-class images are: vector Z i and vector Z t The images generated by multiple vector groups are all the same, and the inter-class images are: vector Z i and vector Z t Images generated by multiple vector groups that are all different; Input every two intra-class images and every two inter-class images into the recognition network respectively, obtain the matching probability of every two intra-class images and every two inter-class images, and calculate the contrast loss based on the matching probability of every two intra-class images and every two inter-class images; Input multiple generated images and real images into the discriminator respectively and calculate the adversarial loss; Calculate the objective function based on contrast loss and adversarial loss to optimize the parameters of the generative adversarial network; The first processing unit is used to process any vector group, the vector Z in the vector group i and vector Z o Merge and input to each convolution layer of the generator; the output of each convolution layer and vector Z t Add them together as the input feature map of the next convolutional layer; The output of the last convolutional layer is used as the generated image corresponding to any vector group; Among them, vector Zi, vector Zo and vector Zt respectively control identity-related attributes, condition-varying attributes and texture. Multiple images with the same vector Zi and vector Zt belong to the same identity, and multiple images with different vector Zi and vector Zt belong to different identities.
7. The iris image generation system according to claim 6, characterized in that: Also includes: The second processing unit extracts features of any two intra-class images or any two inter-class images to obtain feature pairs; The feature pairs are merged and calculated through multiple layers of convolution to obtain the matching probability.
8. The iris image generation system according to claim 6, wherein: Also includes: The first calculation unit is used to calculate the contrast loss based on the matching probability of every two intra-class images and every two inter-class images, specifically including: Contrastive loss ; Among them, I n =G(z n ) represents the input vector group z n =(z i n , z o n , z t n ) generated image; P+, P- respectively represent the set of all intra-class vector pairs and inter-class vector pairs, and the intra-class vector pairs are those with the same vector Z i and vector Z t For any two vector groups, the inter-class vector pair is vector Z i and vector Z t Any two vector groups that are different; E[] is the expected value, V(I m ,I n ) represents image I m ,I n The matching probability.