An automatic segmentation method for texture of iris freckle

Abstract

The application discloses an automatic iris pigment spot texture segmentation method, and relates to an eye pigment segmentation method, and specifically comprises the following steps: establishing an iris detection data set and preparing an iris pigment spot semantic segmentation data set, including normalizing and preprocessing the data set, histogram equalization, data set amplification, dividing the data set into a training set and a test set, processing the test set, building a DIPSnet semantic segmentation network, training the network to obtain an overall semantic segmentation model, obtaining weight parameters of the network model through a gradient back propagation method, inputting images in the test set into the DIPSnet semantic segmentation network model, outputting the images through a softmax layer to generate a classification probability graph, and obtaining a segmentation result graph of the images according to the category probability in the classification probability graph. The application is applied to an iris recognition system and a human health monitoring system based on the iris.

Description

TECHNICAL FIELD

[0001] The present application relates to an iris pigment segmentation method, in particular to an automatic iris pigment spot texture segmentation method. BACKGROUND

[0002] The iris texture information of human eye is very unique, which has the advantages of stability and safety in a certain period, and even the twins with the same genes also have different iris textures. Different types of characteristic textures such as pigment spots, pits, cracks and ring stripes exist on the iris surface, and the size, position and shape of the characteristic texture elements can be used as auxiliary information for iris identification system and iris health diagnosis. In all biometric systems, the iris is the most promising solution, which can be applied in various fields, such as laboratory, credit card identity verification, bank account security access, airport security inspection and so on, which all need identity authentication. Irisology, also known as iris diagnosis, is a discipline that judges the health status of human organs by checking the iris texture. The pigment spot texture, as a block texture in the iris texture, usually represents several aspects of significance. First, the accumulation of toxins or metabolites in the tissue of the organ tissue attached to the pigment, the sluggishness or energy blockage of the tissue, which causes local blood stasis, tissue ischemia and hypoxia. Second, the attached organ has genetic weakness. Third, the attached organ tissue has disease weakness, such as crystallization tendency of the tissue: kidney stones or gallstones; and when showing a deeper color, it indicates that there is an organic tissue change. Therefore, it is necessary to detect the pigment spot texture.

[0003] Since the theory of iris recognition was proposed, researchers such as Daugman have proposed a series of classic local texture feature extraction methods using multi-scale or multi-resolution ideas, which have achieved good recognition accuracy in the case of good iris image clarity. However, due to the small depth of field of the lens in the iris image acquisition device, the clarity of the image cannot be guaranteed during the acquisition process, resulting in the loss of some texture features, which reduces the feature dimension and cannot realize correct matching.

[0004] In recent years, scholars have taken block textures such as pigment spots on the iris surface as the research object. A series of traditional target detection algorithms have been applied to visible light iris pigment spot texture detection, and good results have been achieved. According to the method of detecting pigment spot texture, the literature is divided into the following two categories.

[0005] The first method can be divided into two types: using template and not using template. Literature [1] uses two different length linear templates to extract all different types of feature texture on the iris, which is used as a feature vector input to the BP neural network for training, and finally realizes the detection of pigment spot texture. Literature [2] uses a set of variable size windows to search the iris image, and then uses clustering method to binarize the obtained region, realizing the detection of pigment spot texture in the iris. This method uses template for initial positioning, although the detection accuracy of pigment spot texture is high, but the size of the detected texture is required, only the pigment spot texture meeting the size of the detection template can be detected.

[0006] The next method is not using template detection, literature [3] uses regional texture energy parameters as iris texture descriptor, uses support vector machine to obtain all texture feature regions with severe gray change, and finally realizes the detection of pigment spot texture through shape factor. Literature [4] uses gray clustering method to realize the initial positioning of the possible existence of pigment spot, and uses support vector machine to realize the detection of pigment spot according to the gray space distribution characteristics of pigment spot. The first method has requirements for the position of pigment spot, only the pigment spot texture in the iris region outside the rolled wheel can be detected, and the detection result is easily affected by light spot and eyelid.

[0007] The second method has the following research: literature [5] divides the image into L representative blocks, calculates its gray probability distribution function, and realizes the contour detection of pigment spot according to the difference between the general representative block and the pigment spot pixel sample and K-S distance. Literature [6] locates the pigment spot texture by calculating the local gray minimum value of the large area center. Literature [7] uses improved Laplace algorithm for iris preprocessing, combines mean shift algorithm and pyramid method, and locates pigment spot texture. This method takes a long time to detect, is easily affected by eyelashes, and is suitable for detecting iris images with only pigment spot texture.

[0008] Unet network is first used for medical image segmentation. Compared with FCN, DeepLab series semantic segmentation network, Unet network model has the characteristics of less training data and high accuracy, which meets the characteristics of iris pigment spot texture detection. The improved model has 3 times of down-sampling coding layer, 3 times of up-sampling decoding layer operation, gray part is jump connection layer, and finally uses classifier to classify each pixel.

[0009] In order to make the neural network model more suitable for iris pigment spot texture detection, the iris pigment spot texture is analyzed: due to the complex and irregular characteristics of iris pigment spots, the following factors need to be considered when studying the segmentation characteristics: (1) The size of the iris pigment spot is uncertain, and the edge is not clear. Since the feature map after convolution is more abstract, the edge blurred part is easy to be missed in detection. (2) The colors of different iris pigment spots are different, some are deep and some are light, and the color is uneven in detail, so it is necessary to more flexibly represent the color of the pigment spot, so that the network focuses on some common features of the pigment spot image. (3) The background of the pigment spot located in the convolution wheel is complex, which is easy to detect other textures in the background as pigment spots, so the region pixel of the pigment spot texture needs to be focused and analyzed.

[0010] In the encoding layer downsampling stage, the space and channel attention module is embedded, which makes the network more focused on certain feature layer channels and spatial regions when extracting features, suppresses the redundant features of certain invalid non-feature regions, and reduces the number of down-sampling layers.

[0011] In the network training stage, this paper adopts a hybrid loss function that combines cross-entropy loss function and Dice loss function. Cross-entropy loss is used when the semantic segmentation platform uses Softmax to classify pixels. Dice loss takes the evaluation index of semantic segmentation as Loss, and Dice coefficient is a set similarity measure function, which is usually used to calculate the similarity of two samples, with a value range of [0, 1].

[0012] The Dice loss function was first proposed to solve the problem of data imbalance in natural language processing, and later used in medical image small target segmentation with good results. Because the iris pigment spot texture is small, the number of pixels occupied by the patch in the image is much smaller than the number of background pixels, and a single cross-entropy loss function will cause the model to be severely biased towards the background, resulting in poor detection effect, so this paper adds the Dice loss function to make up for the shortcomings of the single cross-entropy loss function.

[0013] References:

[0014] [1] Zhu L J, Yuan W Q. Iris pigment spot detection based on bilinear template and block strategy[J]. Chinese Journal of Scientific Instrument, 2015, 36(12):2714-2721.

[0015] [2] Liu X N, Yuan W Q, Zhang B. Iris block texture detection based on combined window search[J]. Chinese Journal of Scientific Instrument, 2014, 35(8):1900-1906.

[0016] [3]Yuan Weiqi, Liu Xiaonan, Sun Xiao. A block texture detection algorithm for iris images [J]. Chinese Journal of Scientific Instrument, 2014, 35(5): 1093-1100.

[0017] [4] Liu Xiaonan, Yuan Weiqi, Zhang Bo. A method for iris pigment block detection and classification [J]. Journal of Shenyang University of Technology, 2014, 36(06): 688-693.

[0018] [5] Zhao Chenxu. Iris image feature extraction and its medical application [D]. Inner Mongolia University, 2018.

[0019] [6] Zhao Libin. Automatic positioning algorithm for iris pigment deposition spots [D]. Shenyang University of Technology, 2012.

[0020] [7] Liu Yukie. Research on iris feature extraction method [D]. Dalian University of Technology, 2014.

[0021] [8] Yuan Weiqi, Lin Zhonghua, Xu Lu. A novel iris positioning algorithm based on human eye structure features [J]. Optoelectronic Engineering, 2007, 34(1): 112-116. SUMMARY

[0022] The purpose of the present application is to provide an eye iris pigment spot texture automatic segmentation method, which comprises a DIPSnet iris pigment spot detection model of encoder and decoder; the model introduces a residual attention module in the encoder stage, which suppresses the redundant features of some invalid non-feature areas and improves the pigment spot detection accuracy. Secondly, to adapt to the simple semantic and fixed structure of the pigment spot image, the number of convolutions is reduced to prevent overfitting. Finally, a hybrid loss is introduced to train the network to solve the class imbalance problem. This model combines cross-entropy and Dice loss to further improve the segmentation accuracy. Without the need to remove the texture inside the convolutional shrinkage wheel, the present application automatically identifies iris pigment spots of various sizes by using deep learning method.

[0023] The purpose of the present application is realized by the following technical solutions:

[0024] An eye iris pigment spot texture automatic segmentation method is a biomedical image automatic segmentation method based on U net network structure. The method is an iris pigment spot texture automatic segmentation method based on DIPSnet semantic segmentation network structure, comprising the following steps:

[0025] S1: Establishing an iris detection data set and making an iris pigment spot semantic segmentation data set, including normalizing the data set, histogram equalization, data set augmentation, dividing the data set into a training set and a test set, and processing the test set and the test;

[0026] S2: label the dataset image using labelme, input the network model, and output an image to generate a classification probability map with 2 channels, wherein the classification probability map has the same resolution as the input image;

[0027] S3: build a DIPSnet semantic segmentation network, train the network to obtain an overall semantic segmentation model, and obtain the weight parameters of the network model by gradient back propagation method;

[0028] S4: input the image in the test set into the DIPSnet semantic segmentation network model, and output the image to generate a classification probability map through the softmax layer;

[0029] S5: obtain the segmentation result image of the image according to the class probability in the classification probability map;

[0030] S6: detect the pigment spot texture edge using the canny operator according to the segmentation result image to obtain an edge detection image;

[0031] The DIPSnet network model comprises an encoder with a residual attention mechanism and a decoder network with a deconvolution double upsampling structure. The encoder comprises, in sequence, an input layer, a first convolutional layer with an output channel number of 64, a second convolutional layer, a first fusion attention mechanism layer, a first residual attention convolution fusion layer, a first skip connection layer, a first max pooling layer, a third convolutional layer, a fourth convolutional layer, a second fusion attention mechanism layer, a second residual attention convolution fusion layer, a second skip connection layer, a second max pooling layer, a fifth convolutional layer, a sixth convolutional layer, a third fusion attention mechanism layer, a third residual attention convolution fusion layer, a third skip connection layer, a third max pooling layer, a seventh convolutional layer, an eighth convolutional layer, and a fourth skip connection layer. The decoder network of the upsampling structure comprises a first upsampling layer, a first connection layer, a first regular convolutional layer, a second upsampling layer, a second connection layer, a second regular convolutional layer, a third upsampling layer, a third connection layer, a fourth regular convolutional layer, a fifth regular convolutional layer, and an output layer. The skip connection layers are connected as follows: the first skip connection layer is connected to the third connection layer, the second skip connection layer is connected to the second connection layer, the third skip connection layer is connected to the first connection layer, and the fourth skip connection layer is connected to the first upsampling layer. Each convolutional layer and regular convolutional layer has an activation function. The operation of reconstructing the upsampling layer comprises:

[0032] S31: Except for the first layer, for a feature map with a resolution of h x w and a channel number of c, first increase the channel number by 1 times by 2 1 x 1 convolution without changing the length and width;

[0033] S32: The feature map output by S31 is subjected to a fusion attention mechanism layer and a residual attention convolution fusion layer to obtain a feature map with a channel number of C;

[0034] S33: The feature map output by S32 is subjected to a max-pooling layer to obtain a feature Figure 1 / 2h x 1 / 2w x 2C, thereby completing the up-sampling process of reducing the resolution by two times and expanding the channel number by one time.

[0035] The automatic iris pigment spot texture segmentation method comprises the following steps S1:

[0036] S11: The collected iris image data 800x600 pixels are positioned and normalized to become 200x720 pixels;

[0037] S12: The image data is subjected to adaptive histogram equalization with a probability of 80%, and then the step S13 is skipped:

[0038] S13: The image data is rotated by 180 degrees, flipped up and down, and flipped left and right, and then the step S14 is skipped:

[0039] S14: The image data is subjected to elastic distortion with a probability of 10%, and then the data augmentation processing is completed;

[0040] The automatic iris pigment spot texture segmentation method adopts a hybrid loss function which combines a cross-entropy loss function and a Dice loss function.

[0041]

[0042] wherein, and are pixel points in a predicted image and a label image respectively, and the hybrid loss function is the sum of the two loss functions.

[0043] The advantages and effects of the present application are:

[0044] A DIPSnet iris pigment spot detection model based on a deep learning algorithm comprises an encoder and a decoder. The model introduces a residual attention module in the encoder stage, pays attention to detailed information while obtaining deep semantic information, suppresses the redundant features of certain invalid non-feature regions, and improves the pigment spot detection precision. Secondly, to adapt to the simple semantic and fixed structure of the pigment spot image, the number of convolutions is reduced to prevent overfitting. Finally, a hybrid loss is introduced to train our network to solve the class imbalance problem. The model combines cross-entropy and Dice loss to further improve the segmentation accuracy.

[0045] Without the need to remove the internal texture of the shrinkage wheel, the deep learning method can automatically identify iris pigment spots of various sizes. This method is applied to an iris recognition system and a human health monitoring system based on the iris. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is the overall flowchart of the present application;

[0047] Figure 2 is a schematic diagram of the network structure used in the present application;

[0048] Figure 3 is a schematic diagram of the attention mechanism of the present application;

[0049] Figure 4 is a schematic diagram of the reconstruction down-sampling structure of the present application;

[0050] Figure 5 is a detailed diagram of the network down-sampling structure in the present application;

[0051] Figure 6 is a detailed flowchart of the algorithm in the test stage of the present application;

[0052] Figure 7 is a schematic diagram of the output results of the algorithm test set of the present application;

[0053] Figure 8 is a segmentation effect diagram obtained by using the method of the present application;

[0054] Figure 9 is a collection device diagram. DETAILED DESCRIPTION

[0055] The present application will be described in detail below in conjunction with the embodiments shown in the accompanying drawings.

[0056] The method of the present application specifically comprises the following steps:

[0057] S1: Establishing an iris detection data set and making an iris pigment spot semantic segmentation data set, including normalizing and pre-processing the data set, histogram equalization, data set augmentation, dividing the data set into a training set and a test set, and processing the test set and the test;

[0058] S2: Labeling the data set image using labelme, inputting the network model, and outputting an image to generate a classification probability map with a channel number of 2, wherein the classification probability map has the same resolution as the input image;

[0059] S3: Building a DIPSnet semantic segmentation network, training the network to obtain an overall semantic segmentation model, and obtaining the weight parameters of the network model through a gradient backpropagation method;

[0060] S4: Inputting the images in the test set into the DIPSnet semantic segmentation network model, and outputting an image to generate a classification probability map through a softmax layer;

[0061] S5: Obtain a segmentation result image of the image according to the class probability in the classification probability map;

[0062] S6: According to the segmentation result image, use the canny operator to detect the pigment spot texture edge to obtain an edge detection image.

[0063] Further, step S1 specifically comprises:

[0064] S11: The collected iris image data 800x600 pixel image is positioned and normalized to become 200x720 pixel,

[0065] S12: The image data is adaptively histogram equalized, and then jump to step S13:

[0066] S13: The image data is rotated 180 degrees, upside down and left-right reversed, and then jump to step S14:

[0067] S14: The image data is elastically twisted with a probability of 10%, and then the data augmentation processing is completed;

[0068] Preferably, the DIPSnet network model comprises an encoder with a residual attention mechanism and a decoder network with an up-sampling structure, and the encoder comprises, in sequence, an input layer, a first convolutional layer with an output channel number of 64, a second convolutional layer, a first fusion attention mechanism layer, a first residual attention convolution fusion layer, a first skip connection layer, a first maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a second fusion attention mechanism layer, a second residual attention convolution fusion layer, a second skip connection layer, a second maximum pooling layer, a fifth convolutional layer, a sixth convolutional layer, a third fusion attention mechanism layer, a third residual attention convolution fusion layer, a third skip connection layer, a third maximum pooling layer, a seventh convolutional layer, an eighth convolutional layer, and a fourth skip connection layer.

[0069] The decoder network with the up-sampling structure comprises a first up-sampling layer, a first connection layer, a first regular convolutional layer, a second up-sampling layer, a second connection layer, a second regular convolutional layer, a third up-sampling layer, a third connection layer, a fourth regular convolutional layer, a fifth regular convolutional layer, and an output layer; the skip connection layers are connected in sequence, i.e., the first skip connection layer is connected to the third connection layer, the second skip connection layer is connected to the second connection layer, the third skip connection layer is connected to the first connection layer, and the fourth skip connection layer is connected to the first up-sampling layer; each convolutional layer and regular convolutional layer has an activation function;

[0070] Preferably, the operation of the convolutional layer comprises:

[0071] S31: Input the feature map with a size of h x w x c into the deformable convolutional layer, and use the convolutional layer with an activation function of relu to convolve the feature map;

[0072] S31: In addition to the first layer, for the feature map with resolution h x w and channel number c, first, 2 1x1 convolution operations are performed to keep the length and width unchanged and increase the channel number by one time;

[0073] S32: The feature map output by S31 is subjected to a fusion attention mechanism layer and a residual attention convolution fusion layer to obtain a feature map with channel number C;

[0074] S33: The feature map output by S32 is subjected to a maximum pooling layer to obtain a feature map with resolution 1 / 2h x 1 / 2w x 2C, thereby completing the up-sampling process of reducing the resolution by two times and expanding the channel number by one time. Figure 1

[0075] Preferably, the operation of the reconstruction up-sampling layer comprises:

[0076] S31: In addition to the first up-sampling, the resolution needs to be adjusted to increase by one time, for the feature map with resolution h x w and channel number c, first, a skip connection layer is used for splicing to become a feature map with resolution h x w x 2c;

[0077] S32: The feature map output by S31 is subjected to 2 1x1 convolution operations, group normalization operations and relu activation functions to obtain a feature map with resolution h x w x c;

[0078] S33: The feature map obtained by S32 is subjected to an up-sampling operation to generate a 2h x 2w x 1 / 2c feature map, thereby completing the up-sampling process of expanding the resolution by two times and reducing the channel number by one time.

[0079] Preferably, a hybrid loss function combining a cross-entropy loss function and a Dice loss function is used.

[0080]

[0081] wherein, and are pixel points in the predicted image and the label image respectively, and the hybrid loss function used herein is the sum of the two loss functions. Embodiment

[0082] The detection process of the pigment spot texture is specifically as follows:

[0083] First step: mobile phone dataset, the present research group uses an HM9918 iris instrument as an iris collection device as shown in Figure 7 ​The collection object is shown in Shenyang Industrial University graduate and undergraduate, Shenyang Fengtian Hospital patients, collection equipment and collection method as shown in Figure 9, including iris samples of different age groups such as old, middle-aged and young, the resolution of the image is 800*600 pixels, 24-bit bitmap. There are 980 iris images. All are natural open iris color images of human eyes collected by visible light source. Visible light iris images containing various iris textures described above are collected. Provide necessary hardware support for the smooth development of this topic.

[0084] For the data set with iris pigment spot texture, the method in document [8] is used for preprocessing, as shown in Figure 2 The iris image is normalized to a rectangle along the diameter direction. The horizontal axis of the normalized iris image is the circumference direction of the iris, and the vertical axis is the diameter direction of the iris. The upper edge is the pupil edge. The size of the normalized picture is 720x200.

[0085] This study adopts histogram equalization for iris images. It can enhance the brightness and contrast of the iris images used for training, so as to simulate the actual application scene. It can also expand the data set. Improve the generalization ability and robustness of the model. To ensure the diversity of the data set, the pictures with modified brightness and contrast are added to the data set, and then the horizontal mirror symmetry, vertical mirror symmetry, and rotation of 180o are used. The expanded data set has 1138 images. 90% of the images are used for model training, and 10% are used for testing.

[0086] Contour labeling is a manual annotation of pigment spot texture using labelme according to VOC data set format. In contour labeling, this paper adopts a group of seven members of the research group who have certain professional cognition of iris pigment spot texture features to evaluate the contour labeling of pigment spots. When there is a dispute in labeling, the minority submits to the majority principle to determine it.

[0087] Second step: first resample the input image to 512 512x3, then introduce the training data set into the model for training,

[0088] The experimental environment of this paper is: in the windows10 environment, the graphics card uses GTX1650, the independent graphics card with 4G video memory, uses python3.6 programming on pytorch1.8 platform. In the training process, the initial freezing stage learning rate is 0.0001, the unfreezing stage learning rate is 0.00001, the batch size is 1, the training times are 100, and the loss factor is 0.96. Through the stochastic gradient descent algorithm, the global optimal point is found, and the best effect is obtained.

[0089] During the training process, the image is processed by 3 times of down-sampling encoding layer, 3 times of up-sampling decoding layer, skip connection layer, and finally classified by the classifier. In the down-sampling stage of the encoding layer, the residual space and channel attention module is embedded, so that the network can pay more attention to certain feature layer channels and spatial regions when extracting features. Take one of the following down-sampling as an example, embed the attention module between the two convolution modules and the max pooling module, and superimpose the feature map processed by the attention module with the original convolution feature map, so that it can better aggregate feature information and play the greatest effect. Each convolution module is composed of a 3x3 convolution and a ReLu activation function.

[0090] In this paper, a hybrid loss function combining cross-entropy loss function and Dice loss function is used. Cross-entropy loss is used when the semantic segmentation platform uses Softmax to classify pixels.

[0091] The Dice loss function was first proposed to solve the problem of data imbalance in natural language processing, and later used in medical image small target segmentation also has good effect. Because the iris pigment spot texture is small, the number of pixels occupied by the spot in the image is much smaller than the number of background pixels, and a single cross-entropy loss function will cause the model to be severely biased towards the background, resulting in poor detection effect, so this paper adds the Dice loss function to make up for the shortcomings of a single cross-entropy loss function.

[0092] The result parameters are put into the test data set for prediction. Evaluation indicators are used to judge the network performance. Finally, the network output result is converted into a binary visualization result image.

[0093] Step 3: Use the Canny operator to detect the edges of the result image to obtain the contour map of the pigment spot target detection, and fuse the contour map with the original image.

[0094] Comparison with different deep learning target detection algorithms

[0095]

[0096] Comparison with traditional detection algorithms

[0097] .

Claims

1. An automatic segmentation method for texture of iris freckle of eye, which is an automatic segmentation method for biomedical image based on U net network structure, characterized in that, The method is an automatic iris pigment spot texture segmentation method based on a DIPSnet semantic segmentation network structure, comprising the following steps: S1: establishing an iris detection data set and making an iris pigment spot semantic segmentation data set, including normalizing preprocessing, histogram equalization, data set augmentation, dividing the data set into a training set and a test set, and processing the test set; S2: labeling the data set image using labelme, inputting the network model, and outputting an image to generate a classification probability map with a channel number of 2, wherein the classification probability map has the same resolution as the input image; S3: building a DIPSnet semantic segmentation network, training the network to obtain an overall semantic segmentation model, and obtaining the weight parameters of the network model through a gradient back propagation method; S4: inputting the image in the test set into the DIPSnet semantic segmentation network model, and outputting an image to generate a classification probability map through a softmax layer; S5: obtaining a segmentation result image according to the class probability in the classification probability map; S6: detecting the pigment spot texture edge using a canny operator according to the segmentation result image to obtain an edge detection image. The DIPSnet network model comprises an encoder with a residual attention mechanism and a decoder network with a deconvolution double upsampling structure, and the encoder comprises, in sequence, an input layer, a first convolution layer with an output channel number of 64, a second convolution layer, a first fusion attention mechanism layer, a first residual attention convolution fusion layer, a first skip connection layer, a first maximum pooling layer, a third convolution layer, a fourth convolution layer, a second fusion attention mechanism layer, a second residual attention convolution fusion layer, a second skip connection layer, a second maximum pooling layer, a seventh convolution layer, an eighth convolution layer, a fourth skip connection layer, and a third maximum pooling layer; the decoder network with the upsampling structure comprises a first upsampling layer, a first connection layer, a first regular convolution layer, a second upsampling layer, a second connection layer, a second regular convolution layer, a third upsampling layer, a third connection layer, a fourth regular convolution layer, a fifth regular convolution layer, and an output layer; the skip connection layers are connected in sequence, i.e., the first skip connection layer is connected with the third connection layer, the second skip connection layer is connected with the second connection layer, the third skip connection layer is connected with the first connection layer, and the fourth skip connection layer is connected with the first upsampling layer; each convolution layer and regular convolution layer has an activation function; and the operation of reconstructing the upsampling layer comprises: S31: except for the first layer, for a feature map with a resolution of h x w and a channel number of c, first, 2 1x1 convolutions are performed to keep the length and width unchanged, and the channel number is doubled; S32: the feature map output by S31 is subjected to a fusion attention mechanism layer and a residual attention convolution fusion layer to obtain a feature map with a channel number of C; S33: the feature map output by S32 is subjected to a maximum pooling layer to obtain a feature map 1 / 2h x 1 / 2w x 2C, thereby completing the upsampling process of reducing the resolution by two times and expanding the channel number by one time.

2. The method of claim 1, wherein the method further comprises: The step S1 specifically comprises: S11: The collected iris image data 800*600 pixel image is positioned and normalized to become 200*720 pixel; S12: The image data is adaptively equalized with a probability of 80%, and then jumps to step S13: S13: The image data is rotated by 180 degrees, upside down and left-right reversed, and then jumps to step S14: S14: The image data is elastically twisted with a probability of 10%, and then the data augmentation processing is completed.

3. The method of claim 1, wherein the method further comprises: The method adopts a hybrid loss function combining a cross-entropy loss function and a Dice loss function; where, and are the pixel points in the predicted image and the label image, respectively, and the hybrid loss function is the sum of the two loss functions.

Images