Pupil image segmentation method and system based on deep learning model

Abstract

The invention belongs to the field of medical image processing, and discloses a pupil image segmentation method and system based on a deep learning model, and the method comprises the steps: dividing a plurality of preprocessed historical pupil images into a training image set and a verification image set, so as to carry out the training of the deep learning model, the deep learning model is trained through the diversified historical pupil images, so that the deep learning model obtains good generalization ability and adapts to various complex scenes; secondly, a target pupil image can be subjected to sliding dicing processing to obtain a plurality of sub-pupil images, the sub-pupil images are input into the trained deep learning model one by one for prediction, local errors or failures possibly caused during image overall segmentation are avoided through the mode, and the processing precision of detail parts is improved; and then, all sub-segmentation results predicted by the model can be combined to ensure that the overall segmentation effect of the deep learning model in a complex environment is not influenced by local interference, so that the overall segmentation precision is ensured.

Description

Technical Field

[0001] The present application relates to the field of medical image processing, and in particular to a pupil image segmentation method and system based on a deep learning model. Background Art

[0002] Pupil images generally refer to images that contain the pupil part of the eye. Such images are usually used in medicine, biometrics, and other application scenarios that require eye analysis. The segmentation of pupil images is a key technology in the fields of biometrics, ophthalmic disease diagnosis, and human-computer interaction.

[0003] Traditional pupil image segmentation methods are mainly based on image processing techniques including edge detection, threshold segmentation and morphological operations. These methods perform well when processing images with simple backgrounds and good lighting conditions. However, the segmentation accuracy decreases significantly when the background is complex, the lighting changes and the noise interference is large. In addition, due to the poor adaptability to the diversity of pupil shapes and position changes, it is easy to cause problems such as over-segmentation or under-segmentation, which in turn affects the overall segmentation effect. Summary of the invention

[0004] In order to solve the technical problems that the segmentation accuracy is significantly reduced under complex background, illumination change and noise interference, and the adaptability to the diversity of pupil shapes and position changes is poor, it is easy to produce over-segmentation or under-segmentation, thereby affecting the overall segmentation effect, a pupil image segmentation method and system based on a deep learning model are proposed, and the technical solution is as follows:

[0005] In a first aspect, an embodiment of the present application provides a pupil image segmentation method based on a deep learning model, comprising:

[0006] Acquire at least two historical pupil images, and preprocess each historical pupil image;

[0007] Dividing all preprocessed historical pupil images into a training image set and a verification image set, and training a deep learning model based on the training image set and the verification image set;

[0008] Perform sliding block processing on the target pupil image to obtain at least two sub-pupil images, and input each sub-pupil image into the trained deep learning model to obtain corresponding sub-segmentation results;

[0009] All sub-segmentation results are combined to obtain the segmentation result of the target pupil image.

[0010] In an optional solution of the first aspect, each historical pupil image is preprocessed, including:

[0011] Perform background cropping processing on each historical pupil image to obtain a corresponding background processed image;

[0012] Perform pixel normalization processing on each background processing image to obtain a corresponding pixel processing image;

[0013] Perform flip processing on each pixel processed image to obtain a corresponding flip processed image;

[0014] At least two sub-processed images are cropped from each flipped processed image so that the size of each sub-processed image is consistent with a preset model input size.

[0015] In another optional solution of the first aspect, five groups of training image sets and a verification image set corresponding to each group of training image sets are determined according to all sub-image sets; wherein each group of sub-image sets includes at least two sub-processed images;

[0016] Determine the label image corresponding to each sub-processed image, and add the label image corresponding to each sub-processed image in each set of training images to the corresponding training image set;

[0017] The label image corresponding to each sub-processed image in each verification image set is added to the corresponding verification image set.

[0018] In yet another optional solution of the first aspect, the number of deep learning models is five;

[0019] The deep learning model is trained based on the training image set and the validation image set, including:

[0020] Based on each set of training images and the corresponding set of verification images, any deep learning model is trained to obtain a corresponding trained deep learning model;

[0021] Selecting a trained deep learning model from five trained deep learning models based on a preset performance indicator; or

[0022] Combine the five trained deep learning models into one trained deep learning model.

[0023] In another optional scheme of the first aspect, the deep learning model includes an encoder, a channel attention module, a spatial attention module, a multi-layer perceptron module, and a decoder arranged in sequence, wherein:

[0024] The encoder is used to extract a multi-scale feature sequence from each preprocessed historical pupil image;

[0025] The channel attention module is used to perform channel fusion processing on multi-scale feature sequences;

[0026] The spatial attention module is used to perform spatial fusion processing on the multi-scale feature sequence after channel fusion processing;

[0027] The multi-layer perceptron module is used to perform nonlinear transformation on the multi-scale feature sequence after spatial fusion processing;

[0028] The decoder is used to perform deconvolution processing on the multi-scale feature sequence after nonlinear transformation to obtain the corresponding segmentation result.

[0029] In yet another optional solution of the first aspect, performing sliding slicing processing on the target pupil image to obtain at least two sub-pupil images includes:

[0030] Sliding and slicing the target pupil image according to a preset image axis to obtain at least two sub-pupil images; wherein the size of each sub-pupil image is consistent with a preset model input size;

[0031] According to the cutting order of each sub-pupil image, the corresponding sub-pupil image is labeled.

[0032] In yet another optional solution of the first aspect, all sub-segmentation results are combined to obtain a segmentation result of the target pupil image, including:

[0033] Classify each pixel on each sub-segmentation result, and perform pixel processing on the corresponding sub-segmentation result according to the classification result of each pixel;

[0034] According to the labeling result of each sub-pupil image, all sub-segmentation results after pixel processing are combined to obtain the segmentation result of the target pupil image.

[0035] In yet another alternative of the first aspect, the method further comprises:

[0036] Based on the preset area, filtering is performed on the segmentation result of the target pupil image;

[0037] The segmentation result of the filtered target pupil image is subjected to edge smoothing.

[0038] In a second aspect, an embodiment of the present application provides a pupil image segmentation system based on a deep learning model, comprising:

[0039] An image preprocessing module, used for acquiring at least two historical pupil images and preprocessing each historical pupil image;

[0040] A model training module, used for dividing all pre-processed historical pupil images into a training image set and a verification image set, and training a deep learning model based on the training image set and the verification image set;

[0041] A sub-segmentation result prediction module is used to perform sliding block processing on the target pupil image to obtain at least two sub-pupil images, and input each sub-pupil image into the trained deep learning model to obtain a corresponding sub-segmentation result;

[0042] The post-processing module is used to combine all sub-segmentation results to obtain the segmentation result of the target pupil image.

[0043] In a third aspect, an embodiment of the present application further provides a pupil image segmentation system based on a deep learning model, including a processor and a memory;

[0044] The processor is connected to the memory;

[0045] A memory for storing executable program codes;

[0046] The processor runs the program corresponding to the executable program code by reading the executable program code stored in the memory, so as to implement the pupil image segmentation method based on the deep learning model provided by the first aspect of the embodiment of the present application or any one of the implementation methods of the first aspect.

[0047] In a fourth aspect, an embodiment of the present application provides a computer storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, the pupil image segmentation method based on the deep learning model provided by the first aspect of the embodiment of the present application or any one of the implementation methods of the first aspect can be implemented.

[0048] Beneficial effects of this application:

[0049] When performing pupil image segmentation, multiple pre-processed historical pupil images can be divided into a training image set and a verification image set to train the deep learning model. By using diverse historical pupil images to train the deep learning model, the deep learning model can obtain good generalization ability and adapt to various complex scenes; secondly, the target pupil image can be processed by sliding block to obtain multiple sub-pupil images, which are input into the trained deep learning model one by one for prediction. In this way, local errors or failures that may be caused by the overall segmentation of the image can be avoided, and the processing accuracy of the details is improved; then, all sub-segmentation results predicted by the deep learning model can be combined to ensure that the overall segmentation effect of the deep learning model in a complex environment is not affected by local interference, thereby ensuring the overall segmentation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 An overall flow chart of a pupil image segmentation method based on a deep learning model provided in an embodiment of the present application;

[0052] Figure 2 A schematic diagram of the effect of a deep learning model provided in an embodiment of the present application;

[0053] Figure 3 A schematic diagram of a pupil image and segmentation result provided in an embodiment of the present application;

[0054] Figure 4 A schematic diagram of the structure of a pupil image segmentation system based on a deep learning model provided in an embodiment of the present application;

[0055] Figure 5 A schematic diagram of the structure of another pupil image segmentation system based on a deep learning model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0057] In the following introduction, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. The following introduction provides multiple embodiments of the present application, and different embodiments can be replaced or combined, so the present application can also be considered to include all possible combinations of the same and / or different embodiments recorded. Therefore, if one embodiment includes features A, B, and C, and another embodiment includes features B and D, then the present application should also be considered to include embodiments containing one or more of all other possible combinations of A, B, C, and D, although the embodiment may not be clearly recorded in the following text.

[0058] The following description provides examples and does not limit the scope, applicability or examples set forth in the claims. Changes may be made to the functions and arrangements of the elements described without departing from the scope of the present application. Various processes or components may be appropriately omitted, substituted or added to each example. For example, the described method may be performed in an order different from the order described, and various steps may be added, omitted or combined. In addition, features described in some examples may be combined in other examples.

[0059] See also Figure 1 , Figure 1 An overall flow chart of a pupil image segmentation method based on a deep learning model provided in an embodiment of the present application is shown.

[0060] like Figure 1 As shown, the pupil image segmentation method based on the deep learning model may at least include the following steps:

[0061] Step 102: Obtain at least two historical pupil images, and preprocess each historical pupil image.

[0062] In an embodiment of the present application, the pupil image segmentation method based on the deep learning model can be but is not limited to being applied to a control terminal. The control terminal can establish a connection with a user terminal or a medical equipment terminal to receive multiple pupil images collected by the user terminal or the medical equipment terminal, and can perform segmentation processing on one or more pupil images to obtain corresponding segmentation results.

[0063] It can be understood that when the control terminal performs segmentation processing on the pupil image, it can divide multiple pre-processed historical pupil images into a training image set and a verification image set to train the deep learning model. Through the diversity of historical pupil images and the model prediction method, it can avoid the influence of various environments such as complex background, lighting changes or large noise interference on the segmentation accuracy, and improve the image segmentation efficiency; secondly, the target pupil image can also be subjected to sliding block processing to obtain multiple sub-pupil images, and each sub-pupil image can be input into the trained deep learning model to ensure the segmentation accuracy of each sub-pupil image, and by combining all sub-segmentation results predicted by the model, the overall segmentation accuracy of the pupil image in various complex environments can be effectively guaranteed.

[0064] Specifically, when performing segmentation processing on the pupil image, the control terminal can, but is not limited to, extract multiple historical pupil images collected by the user terminal or the medical equipment terminal in a folder in a specified storage path (for example, an img folder), and pre-process each historical pupil image to effectively ensure the image quality and diversity, and to ensure that the pre-processed image meets the input requirements of the deep learning model, thereby ensuring the generalization ability and preventing overfitting during subsequent deep learning model training.

[0065] As an option of the embodiment of the present application, each historical pupil image is preprocessed, including:

[0066] Perform background cropping processing on each historical pupil image to obtain a corresponding background processed image;

[0067] Perform pixel normalization processing on each background processing image to obtain a corresponding pixel processing image;

[0068] Perform flip processing on each pixel processed image to obtain a corresponding flip processed image;

[0069] At least two sub-processed images are cropped from each flipped processed image so that the size of each sub-processed image is consistent with a preset model input size.

[0070] Specifically, when preprocessing each historical pupil image, the control terminal can specifically perform background cropping processing on each historical pupil image through a preset image processing algorithm to remove the background area to reduce the amount of invalid calculation of the historical pupil image. Here, the preset image processing algorithm can be, but is not limited to, first converting each historical pupil image into a grayscale image, then separating the foreground image and the background image in each historical pupil image by binarization, and then cropping the foreground image contour in each historical pupil image by contour recognition, and is not limited to this.

[0071] Next, after background cropping each historical pupil image to obtain a corresponding background processed image, the control terminal can also perform pixel normalization on each background processed image to normalize each pixel value in each background processed image within the range of [0, 1] or [-1, 1] to reduce the pixel intensity difference between different background processed images (i.e., standardize the pixel intensity value of each background processed image), thereby ensuring the training effect of the deep learning model. Here, the pixel normalization processing method can be, but is not limited to, linear normalization, normalization based on mean and standard deviation, and is not limited to this.

[0072] Next, after performing pixel normalization processing on each background processed image to obtain the corresponding pixel processed image, the control terminal may also perform flipping and rotation processing on each pixel processed image to ensure the diversity of the image, to ensure the generalization ability during the subsequent deep learning model training, and to prevent overfitting. Here, the flipping processing method may be, but is not limited to, processing each pixel processed image in a preset flipping direction (for example, horizontal flipping and vertical flipping), and the rotation processing method may be, but is not limited to, processing each pixel processed image in a preset rotation angle (for example, 90 degrees, 180 degrees, and 270 degrees), and each pixel processed image may be subjected to flipping processing and rotation processing to obtain a corresponding plurality of flipped processed images, but is not limited thereto.

[0073] Next, after flipping and rotating each pixel processed image to obtain the corresponding flipped processed image, the control terminal can also crop at least two sub-processed images from each flipped processed image according to the preset model input size, so that the size of each sub-processed image is consistent with the preset model input size, thereby ensuring that each sub-processed image can meet the input requirements of the deep learning model. Here, the size of each pixel processed image can be but is not limited to 1920×1080 pixels, the preset model input size can be but is not limited to 224×224 pixels, and one or more sub-processed images can be cropped out in each flipped processed image based on random directions to further enhance the diversity of the data.

[0074] Of course, in addition to the above-mentioned preprocessing methods of background cropping, pixel normalization, flipping and cropping, the present application can also process historical pupil images in other ways, which will not be elaborated here.

[0075] Step 104: Divide all pre-processed historical pupil images into a training image set and a verification image set, and train the deep learning model based on the training image set and the verification image set.

[0076] Specifically, after preprocessing each historical pupil image, the control terminal may, but is not limited to, divide all preprocessed historical pupil images into a training image set and a verification image set according to a preset ratio, for example, randomly selecting 50% of all preprocessed historical pupil images as a training image set, and using the remaining part of the historical pupil images as a verification image set. Here, the training image set is mainly used to train the deep learning model, which may, but is not limited to, include multiple preprocessed historical pupil images corresponding to at least one flipping direction, and multiple preprocessed historical pupil images corresponding to at least one rotation angle, and the number of preprocessed historical pupil images in the training image set must be guaranteed to be a multiple of 2, so as to facilitate the parallel computing efficiency of image batch processing; the verification image set is mainly used to obtain the performance indicators of the trained deep learning model, which may, but is not limited to, include multiple preprocessed historical pupil images corresponding to at least one flipping direction, and multiple preprocessed historical pupil images corresponding to at least one rotation angle.

[0077] It is understandable that in order to ensure the training effect of the deep learning model and ensure the generalization ability and robustness of the model, the control terminal can also, but is not limited to, extract the sample image corresponding to each historical pupil image in the folder of the specified storage path (for example, the labelocl folder), and add all the sample images corresponding to the training image set to the training image set, and add all the sample images corresponding to the verification image set to the verification image set. Here, the sample image can be understood as a historical pupil image (and the historical pupil image can be gray-scale processed) with the category of each pixel point marked (which can be manually marked or automatically marked by a program), and the category of each pixel point can be divided into a pupil area or a background area, and the segmentation effect for the pupil area can be, but is not limited to, represented by white, and the segmentation effect for the background area can be, but is not limited to, represented by black, and is not limited to this.

[0078] As another option of the embodiment of the present application, all pre-processed historical pupil images are divided into a training image set and a verification image set, including:

[0079] Dividing all sub-processed images into five groups of sub-image sets, and determining five groups of training image sets and a verification image set corresponding to each group of training image sets based on all the sub-image sets; wherein each group of sub-image sets includes at least two sub-processed images;

[0080] Determine the label image corresponding to each sub-processed image, and add the label image corresponding to each sub-processed image in each set of training images to the corresponding training image set;

[0081] The label image corresponding to each sub-processed image in each verification image set is added to the corresponding verification image set.

[0082] Specifically, in order to more comprehensively evaluate the generalization ability of the deep learning model and reduce performance fluctuations caused by different image divisions, the control terminal can also divide all sub-processed images (that is, all historical pupil images after preprocessing) into five groups of sub-processed image sets, and the number of all sub-processed images in each group of sub-processed image sets can be a multiple of 2. By selecting any four groups of sub-processed image sets as training image sets and using the remaining one group of sub-processed image sets as the corresponding verification image set, five training image sets and a verification image set corresponding to each training image set are obtained.

[0083] Next, in order to ensure the training effect of the deep learning model, the control terminal can first determine the historical pupil image corresponding to each sub-processed image, as well as the cropping position on the corresponding historical pupil image; then, the sample image corresponding to the corresponding historical pupil image can be extracted from the folder of the specified storage path (for example, the labelocl folder), and the area image on the sample image consistent with the corresponding cropping position can be used as the label image corresponding to each sub-processed image.

[0084] Then, after determining the label image corresponding to each sub-processed image, the control terminal can filter out the label image corresponding to each sub-processed image in each group of training image sets, and add all the label images to the corresponding training image set; in addition, the control terminal can also filter out the label image corresponding to each sub-processed image in each group of verification image sets, and add all the label images to the corresponding verification image set.

[0085] Further, after obtaining the training image set and the verification image set, the control terminal can be, but is not limited to, setting the network hyperparameters and loss function of the deep learning model, and training the deep learning model with the network hyperparameters and loss function settings based on the training image set until convergence, so as to obtain the performance index of the trained deep learning model in combination with the verification image set, and then determine that the deep learning model has been trained when the performance index meets the conditions. Here, the network hyperparameter setting may specifically include the setting of the initial learning rate (for example, it may be set to an initial value, 0.001), the setting of the optimizer (for example, the Adam optimizer may be selected), the setting of the number of training rounds (for example, 500 rounds), and the setting of the learning rate adjustment strategy (for example, the CosineAnnealingWarmRestarts strategy may be used, if the loss function does not decrease within the specified patience round, the learning rate is reduced to the original factor times); the setting of the loss function may be to use cross entropy loss (CE) and Dice loss as the loss function, and assign different weights to different categories. Of course, in the embodiment of the present application, other types of parameter settings may also be included, but they are not described in detail here.

[0086] It can be understood that the performance indicators mentioned above can be but are not limited to the Dice coefficient, Jaccard index (IoU), precision or recall rate calculated based on the segmentation results predicted by the deep learning model and the corresponding sample images (which can also be understood as the label images mentioned above), and when a certain indicator exceeds the preset indicator threshold, it indicates that the deep learning model has been trained and has good segmentation accuracy.

[0087] It should be noted that traditional pupil image segmentation methods mainly rely on image processing techniques such as edge detection, threshold segmentation, and morphological operations. These methods may perform well when processing simple scenes, but the segmentation accuracy drops significantly when encountering complex backgrounds, lighting changes, or scenes with large noise interference, and it is difficult to cope with the diversity of pupil shapes and position changes.

[0088] In recent years, with the development of deep learning technology, especially the successful application of convolutional neural networks in the field of image processing, pupil image segmentation methods based on convolutional neural networks have gradually become a research hotspot. Common pupil image segmentation methods based on convolutional neural networks can achieve high-precision pupil segmentation by constructing a convolutional neural network model to extract the features of the pupil image using the convolutional neural network model. However, this pupil image segmentation method based on convolutional neural networks still has some problems. First, due to the diversity and complexity of pupil images, the convolution operation that simply relies on convolutional neural networks lacks the ability to adapt to complex environments, and is prone to over-segmentation (i.e., the pupil edge is incorrectly enlarged) or under-segmentation (the pupil area is not fully recognized), thereby affecting the overall segmentation effect; secondly, the existing convolutional neural networks have a slow processing speed for large-scale image processing tasks, especially when facing high-resolution images and complex backgrounds, and the processing efficiency is difficult to meet the needs of real-time applications.

[0089] Based on this, in order to solve the technical problems encountered by the above-mentioned common convolutional neural networks, such as the deficiencies in extracting pupil detail features and adapting to the diversity of pupil shapes, the large amount of calculation when processing large-scale data, and the difficulty in pupil images meeting the requirements of real-time processing, the deep learning model mentioned in one or more embodiments of the present application may specifically include an encoder, a channel attention module, a spatial attention module, a multi-layer perceptron module and a decoder arranged in sequence. By combining the superior performance of the encoder and the decoder in processing long-distance dependencies and capturing global features, and the advantages of the channel attention module, the spatial attention module and the multi-layer perceptron module in finely processing the local features and boundary information of the image, it not only effectively improves the accuracy and efficiency of pupil position and size detection and reduces the amount of calculation, but also can be applied to pupil image segmentation tasks in various complex scenarios.

[0090] This can also be combined with Figure 2 The effect diagram of a deep learning model provided by an embodiment of the present application is shown in FIG. Figure 2 As shown, the encoder may include a plurality of sequentially connected encoding layers (i.e. Figure 2The invention relates to five encoder modules in the present invention, each encoding layer may include a convolution layer and a pooling layer. The convolution layer may be used to extract the features of the pupil image through a convolution operation, and the pooling layer may be used to downsample the features of the pupil image, thereby reducing the spatial dimension while retaining important feature information. Furthermore, a multi-scale feature sequence fused by multiple feature sequences may be extracted from each historical pupil image of the training image set in a layer-by-layer manner.

[0091] Channel attention module (i.e. Figure 2 The CFA in (CFA) can be used to perform attention calculation on the feature sequence extracted by each coding layer, and the feature sequence extracted by each coding layer can be linearly changed to generate a corresponding query matrix (for matching and weight calculation in the attention mechanism), and the feature sequence extracted by each coding layer can be linearly changed to generate a corresponding key matrix (for calculating the weights in the attention mechanism in combination with the query matrix), and the feature sequence extracted by each coding layer can be linearly changed to generate a corresponding value matrix (for calculating the final output according to the attention weights), and the number of heads in the multi-head attention mechanism can be used to calculate attention in parallel in different subspaces (to improve the expressive power of the deep learning model), and by combining the attention calculation results corresponding to each coding layer, the query matrix, the key matrix and the value matrix, a multi-scale feature sequence after channel fusion processing is obtained. Here, the channel attention module may also be provided with a specific operation function for executing the attention mechanism, so as to calculate the correlation between the query matrix and the key matrix and generate the attention weight, and may also be provided with a function for applying the attention mechanism, so as to perform weighted processing on the value matrix according to the calculated attention weight. Both functions may be understood as conventional technical means of the attention mechanism, and will not be elaborated here.

[0092] The spatial attention module (i.e. Figure 2The SSA in the multi-scale feature sequence (SSA) can be used to perform attention calculations on different spatial positions in the multi-scale feature sequence after channel fusion processing, and linear changes can be performed on different spatial positions in the multi-scale feature sequence after channel fusion processing to generate corresponding query matrices (for matching and weight calculation in the attention mechanism), and linear changes can be performed on different spatial positions in the multi-scale feature sequence after channel fusion processing to generate corresponding key matrices (for calculating the weights in the attention mechanism in combination with the query matrix), and linear changes can be performed on different spatial positions in the multi-scale feature sequence after channel fusion processing to generate corresponding value matrices (for calculating the final output according to the attention weights), and the number of heads in the multi-head attention mechanism can be used to calculate attention in parallel in different subspaces (to improve the expressive power of the deep learning model), and the multi-scale feature sequence after spatial fusion processing is obtained by combining the attention calculation results, query matrix, key matrix and value matrix corresponding to different spatial positions in the multi-scale feature sequence after channel fusion processing. Here, the spatial attention module may also be provided with a specific operation function for executing the attention mechanism, which is used to calculate the correlation between the query matrix and the key matrix and generate attention weights, and may also be provided with a function for applying the attention mechanism to perform weighted processing on the value matrix according to the calculated attention weights. Both functions can be understood as conventional technical means of the attention mechanism, and will not be elaborated here.

[0093] Multilayer Perceptron Module (i.e. Figure 2 The MLPS in the deep learning algorithm can perform nonlinear changes on the multi-scale feature sequence after spatial fusion processing by setting multiple linear transformation layers and multiple activation functions, so that the deep learning model can learn more complex representations.

[0094] The decoder may include a plurality of sequentially connected decoding layers (i.e. Figure 2 Each decoding layer may include a deconvolution layer and an upsampling layer for the last encoding (i.e. Figure 2 The multi-scale feature sequence output by the encoder module 5) is upsampled and deconvolved layer by layer, and the decoder guides the attention module (also understood as DRA) to perform dynamic attention processing on the feature map after upsampling at each decoding layer (that is, enhancing the detail information of the feature map) by linearly projecting the multi-scale feature sequence after nonlinear change processing, thereby obtaining a segmentation result with restored image spatial resolution.

[0095] The deep learning model proposed in the embodiment of the present application can extract and understand pupil features by learning a large number of historical pupil images, so as to better adapt to complex environments, such as lighting changes and background interference; secondly, by introducing channel attention and spatial attention mechanisms, the deep learning model can more effectively capture the feature relationships in different channels and spaces, thereby improving segmentation accuracy. It should be noted that this channel attention and spatial attention mechanism can help the deep learning model to be more robust when dealing with complex visual scenes.

[0096] Secondly, by using a large number of diverse historical pupil images during the training process, covering different lighting, background and noise conditions, the deep learning model can better generalize to pupil images in different scenarios and improve the ability to process complex images.

[0097] Secondly, by dividing the target image into multiple sub-images (sliding block processing), the deep learning model can perform accurate segmentation independently on each sub-image, thereby improving the accuracy of the overall image segmentation, especially in high-resolution images. This method can avoid the accuracy loss and computational burden caused by global processing.

[0098] Secondly, the multi-level perception mechanism of the deep learning model (such as channel attention and spatial attention) can be used to segment the pupil edge more accurately, thereby effectively reducing the phenomenon of over-segmentation and under-segmentation. Compared with traditional threshold or edge detection-based methods, this deep learning model has stronger adaptability to pupils of different shapes and sizes.

[0099] Secondly, after training, the deep learning model can greatly improve the segmentation processing speed, especially when facing large-scale data sets. Compared with the traditional step-by-step processing method, the deep learning model can complete complex pupil image segmentation tasks in a shorter time, thereby improving the overall efficiency of the system.

[0100] As another option of the embodiment of the present application, the number of deep learning models is five;

[0101] The deep learning model is trained based on the training image set and the validation image set, including:

[0102] Based on each set of training images and the corresponding set of verification images, any deep learning model is trained to obtain a corresponding trained deep learning model;

[0103] Selecting a trained deep learning model from five trained deep learning models based on a preset performance indicator; or

[0104] Combine the five trained deep learning models into one trained deep learning model.

[0105] Specifically, in order to further and more comprehensively evaluate the generalization ability of the deep learning model and reduce the performance fluctuations caused by different pupil image divisions, the control terminal can set five deep learning models with the same structure, set the network hyperparameters and loss function for each deep learning model respectively, and train any deep learning model with the network hyperparameters and loss function settings based on each set of training images until convergence, so as to obtain the performance indicators of the corresponding deep learning model in combination with the corresponding verification image set.

[0106] It can be understood that after obtaining the performance indicators of five trained deep learning models, the control terminal can use the trained deep learning model with the best performance indicator as the trained deep learning model for the final application. For example, taking the Jaccard index as the performance indicator, the control terminal can use the trained deep learning model with the highest Jaccard index as the trained deep learning model for the final application, and is not limited to this.

[0107] Of course, in the embodiments of the present application, it is also possible but not limited to that when the performance indicators of each trained deep learning model meet the requirements, the five trained deep learning models are used together as the trained deep learning model for the final application, and the final decomposition result can be obtained by performing mean processing or other integration processing on the decomposition results of the five trained deep learning models.

[0108] It should be noted that the final decomposition result can be a pupil image marked with the classification result of each pixel (and the pupil image can be processed by grayscale), and the classification result can include but is not limited to the probability of the pixel belonging to the pupil area and the probability of belonging to the background area.

[0109] Step 106: Perform sliding block processing on the target pupil image to obtain at least two sub-pupil images, and input each sub-pupil image into the trained deep learning model to obtain corresponding sub-segmentation results.

[0110] Specifically, after obtaining the trained deep learning model, the control terminal may, but is not limited to, perform sliding block processing on the currently acquired target pupil image to divide the target pupil image into multiple sub-pupil images, ensuring that each sub-pupil image can meet the input requirements of the deep learning model, and input each sub-pupil image into the trained deep learning model respectively to obtain a sub-segmentation result corresponding to each sub-pupil image. Here, each sub-segmentation result may be a sub-pupil image marked with the classification result of each pixel (and the sub-pupil image may be grayscale processed).

[0111] As another option of the embodiment of the present application, the target pupil image is subjected to sliding block processing to obtain at least two sub-pupil images, including:

[0112] Sliding and slicing the target pupil image according to a preset image axis to obtain at least two sub-pupil images; wherein the size of each sub-pupil image is consistent with a preset model input size;

[0113] According to the cutting order of each sub-pupil image, the corresponding sub-pupil image is labeled.

[0114] Specifically, when performing sliding slicing processing on the target pupil image, the control terminal can perform sliding slicing processing on the target pupil image according to a preset image axis and a preset step size, so that the size of each sub-pupil image can be but is not limited to 224×224 pixels, and in order to ensure the efficiency of subsequent combination processing, the corresponding sub-pupil images can also be labeled according to the slicing order of each sub-pupil image. For example, when a sub-pupil image is the first sub-pupil image obtained after slicing processing, the first sub-pupil image can be labeled 1, and is not limited to this.

[0115] It should be noted that, after obtaining the target pupil image, the control terminal in the embodiment of the present application may also, but is not limited to, preprocess the target pupil image to ensure the quality of the target pupil image. The preprocessing method may refer to one or more of the above embodiments, which will not be elaborated here.

[0116] Step 108: Combine all sub-segmentation results to obtain a segmentation result of the target pupil image.

[0117] Specifically, after obtaining all sub-segmentation results, the control terminal can, but is not limited to, use the argmax function to classify each pixel on each sub-segmentation result. For example, but not limited to, when the probability of a certain pixel belonging to the pupil area is greater than the probability of belonging to the background area, the pixel can be classified as the pupil area; when the probability of a certain pixel belonging to the pupil area is less than the probability of belonging to the background area, the pixel can be classified as the background area, and corresponding pixel processing can be performed according to the classification results of all pixels on each sub-segmentation result. Here, in order to highlight the pupil segmentation effect, when performing pixel processing according to the classification results of all pixels on each sub-segmentation result, the color of all pixels belonging to the pupil area can be processed as white (for example, by adjusting the RGB three channels), and the color of all pixels belonging to the background area can be processed as black, and it is not limited to this.

[0118] Furthermore, after performing pixel processing on each sub-segmentation result, the control terminal may combine and process each processed sub-segmentation result in order from small to large according to the corresponding label sequence, so as to obtain the segmentation result of the target image.

[0119] Also see here Figure 3 FIG. 1 is a schematic diagram showing the effect of a pupil image and a segmentation result provided by an embodiment of the present application. Figure 3 As shown, 3A shows a pupil image, 3a is the segmentation result corresponding to 3A, wherein the color of the pupil area is processed as white, and the color of the background area is processed as black; 3B shows a pupil image, 3b is the segmentation result corresponding to 3B, wherein the color of the pupil area is processed as white, and the color of the background area is processed as black; 3C shows a pupil image, 3c is the segmentation result corresponding to 3C, wherein the color of the pupil area is processed as white, and the color of the background area is processed as black.

[0120] As another option of the embodiment of the present application, the method further includes:

[0121] Based on the preset area, filtering is performed on the segmentation result of the target pupil image;

[0122] The segmentation result of the filtered target pupil image is subjected to edge smoothing.

[0123] Specifically, after obtaining the segmentation result of the target pupil image, in order to further improve the image quality of the segmentation result, the control terminal can also filter the segmentation result based on the preset area to filter out all areas smaller than the preset area (i.e., misclassified pixel groups with smaller areas, which can be small noise areas or misclassified areas), thereby achieving the purpose of removing small isolated areas. Of course, the embodiment of the present application can also filter the segmentation result through morphological operations, but it will not be described in detail here.

[0124] Next, in order to make the edges of the segmentation results smoother and more natural, the control terminal can also perform edge smoothing on the filtered segmentation results, such as Gaussian filtering or edge shrinkage, which not only makes the edges of the segmentation results smoother and more natural, but also improves the visual quality of the segmentation results.

[0125] Of course, after obtaining the segmentation result of the target pupil image, the control terminal can also combine the segmentation result and the sample image corresponding to the target pupil image to obtain a segmentation evaluation index to verify the performance of the trained deep learning model in practical applications and ensure that the trained deep learning model can accurately segment the pupil area. Here, the standard image corresponding to the target pupil image can be, but is not limited to, obtained by manual analysis or automatically generated by a program, and the calculation method of the corresponding segmentation evaluation index can be any of the above-mentioned Dice coefficient, Jaccard index (IoU), precision or recall rate, etc., but no further elaboration is given here.

[0126] See also Figure 4 , Figure 4 A structural schematic diagram of a pupil image segmentation system based on a deep learning model provided in an embodiment of the present application is shown.

[0127] like Figure 4 As shown, the pupil image segmentation system based on the deep learning model may include at least an image preprocessing module 401, a model training module 402, a sub-segmentation result prediction module 403 and a post-processing module 404, wherein:

[0128] An image preprocessing module 401 is used to obtain at least two historical pupil images and preprocess each historical pupil image;

[0129] A model training module 402, for dividing all pre-processed historical pupil images into a training image set and a verification image set, and training a deep learning model based on the training image set and the verification image set;

[0130] The sub-segmentation result prediction module 403 is used to perform sliding block processing on the target pupil image to obtain at least two sub-pupil images, and input each sub-pupil image into the trained deep learning model to obtain a corresponding sub-segmentation result;

[0131] The post-processing module 404 is used to combine all the sub-segmentation results to obtain the segmentation result of the target pupil image.

[0132] In some possible embodiments, each historical pupil image is preprocessed, including:

[0133] Perform background cropping processing on each historical pupil image to obtain a corresponding background processed image;

[0134] Perform pixel normalization processing on each background processing image to obtain a corresponding pixel processing image;

[0135] Perform flip processing on each pixel processed image to obtain a corresponding flip processed image;

[0136] At least two sub-processed images are cropped from each flipped processed image so that the size of each sub-processed image is consistent with a preset model input size.

[0137] In some possible embodiments, five groups of training image sets and a verification image set corresponding to each group of training image sets are determined according to all sub-image sets; wherein each group of sub-image sets includes at least two sub-processing images;

[0138] Determine the label image corresponding to each sub-processed image, and add the label image corresponding to each sub-processed image in each set of training images to the corresponding training image set;

[0139] The label image corresponding to each sub-processed image in each verification image set is added to the corresponding verification image set.

[0140] In some possible embodiments, the number of deep learning models is five;

[0141] The deep learning model is trained based on the training image set and the validation image set, including:

[0142] Based on each set of training images and the corresponding set of verification images, any deep learning model is trained to obtain a corresponding trained deep learning model;

[0143] Selecting a trained deep learning model from five trained deep learning models based on a preset performance indicator; or

[0144] Combine the five trained deep learning models into one trained deep learning model.

[0145] In some possible embodiments, the deep learning model includes an encoder, a channel attention module, a spatial attention module, a multi-layer perceptron module, and a decoder arranged in sequence, wherein:

[0146] The encoder is used to extract a multi-scale feature sequence from each preprocessed historical pupil image;

[0147] The channel attention module is used to perform channel fusion processing on multi-scale feature sequences;

[0148] The spatial attention module is used to perform spatial fusion processing on the multi-scale feature sequence after channel fusion processing;

[0149] The multi-layer perceptron module is used to perform nonlinear transformation on the multi-scale feature sequence after spatial fusion processing;

[0150] The decoder is used to perform deconvolution processing on the multi-scale feature sequence after nonlinear transformation to obtain the corresponding segmentation result.

[0151] In some possible embodiments, performing sliding slicing processing on the target pupil image to obtain at least two sub-pupil images includes:

[0152] Sliding and slicing the target pupil image according to a preset image axis to obtain at least two sub-pupil images; wherein the size of each sub-pupil image is consistent with a preset model input size;

[0153] According to the cutting order of each sub-pupil image, the corresponding sub-pupil image is labeled.

[0154] In some possible embodiments, all sub-segmentation results are combined to obtain a segmentation result of a target pupil image, including:

[0155] Classify each pixel on each sub-segmentation result, and perform pixel processing on the corresponding sub-segmentation result according to the classification result of each pixel;

[0156] According to the labeling result of each sub-pupil image, all sub-segmentation results after pixel processing are combined to obtain the segmentation result of the target pupil image.

[0157] In some possible embodiments, the pupil image segmentation system based on the deep learning model further includes:

[0158] Based on the preset area, filtering is performed on the segmentation result of the target pupil image;

[0159] The segmentation result of the filtered target pupil image is subjected to edge smoothing.

[0160] Those skilled in the art can clearly understand that the technical solutions of the embodiments of the present application can be implemented with the help of software and / or hardware. The "unit" and "module" in this specification refer to software and / or hardware that can independently complete or cooperate with other components to complete specific functions, where the hardware can be, for example, a field programmable gate array (Field-Programmable Gate Array, FPGA), an integrated circuit (Integrated Circuit, IC), etc.

[0161] See also Figure 5 , Figure 5 A structural schematic diagram of another pupil image segmentation system based on a deep learning model provided in an embodiment of the present application is shown.

[0162] like Figure 5 As shown, the pupil image segmentation system 500 based on the deep learning model may include at least one processor 501, at least one network interface 504, a user interface 503, a memory 505 and at least one communication bus 502.

[0163] The communication bus 502 may be used to realize the connection and communication among the above-mentioned components.

[0164] The user interface 503 may include buttons, and the optional user interface may also include a standard wired interface or a wireless interface.

[0165] The network interface 504 may include, but is not limited to, a Bluetooth module, an NFC module, a Wi-Fi module, etc.

[0166] Among them, the processor 501 may include one or more processing cores. The processor 501 uses various interfaces and lines to connect the various parts of the pupil image segmentation system 500 based on the deep learning model, and executes various functions and processes data of the pupil image segmentation system 500 based on the deep learning model by running or executing instructions, programs, code sets or instruction sets stored in the memory 505, and calling data stored in the memory 505. Optionally, the processor 501 can be implemented in at least one hardware form of DSP, FPGA, and PLA. The processor 501 can integrate one or a combination of CPU, GPU, modem, etc. Among them, the CPU mainly processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor 501, and it can be implemented by a single chip.

[0167] The memory 505 may include a RAM or a ROM. Optionally, the memory 505 includes a non-transitory computer-readable medium. The memory 505 may be used to store instructions, programs, codes, code sets or instruction sets. The memory 505 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned method embodiments, etc.; the data storage area may store data involved in the above-mentioned method embodiments, etc. The memory 505 may optionally be at least one storage device located away from the aforementioned processor 501. As Figure 5 As shown, the memory 505 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a pupil image segmentation application based on a deep learning model.

[0168] Specifically, the processor 501 may be used to call the pupil image segmentation application based on the deep learning model stored in the memory 505, and specifically perform the following operations:

[0169] Acquire at least two historical pupil images, and preprocess each historical pupil image;

[0170] Dividing all preprocessed historical pupil images into a training image set and a verification image set, and training a deep learning model based on the training image set and the verification image set;

[0171] Perform sliding block processing on the target pupil image to obtain at least two sub-pupil images, and input each sub-pupil image into the trained deep learning model to obtain corresponding sub-segmentation results;

[0172] All sub-segmentation results are combined to obtain the segmentation result of the target pupil image.

[0173] In some possible embodiments, each historical pupil image is preprocessed, including:

[0174] Perform background cropping processing on each historical pupil image to obtain a corresponding background processed image;

[0175] Perform pixel normalization processing on each background processing image to obtain a corresponding pixel processing image;

[0176] Perform flip processing on each pixel processed image to obtain a corresponding flip processed image;

[0177] At least two sub-processed images are cropped from each flipped processed image so that the size of each sub-processed image is consistent with a preset model input size.

[0178] In some possible embodiments, five groups of training image sets and a verification image set corresponding to each group of training image sets are determined according to all sub-image sets; wherein each group of sub-image sets includes at least two sub-processing images;

[0179] Determine the label image corresponding to each sub-processed image, and add the label image corresponding to each sub-processed image in each set of training images to the corresponding training image set;

[0180] The label image corresponding to each sub-processed image in each verification image set is added to the corresponding verification image set.

[0181] In some possible embodiments, the number of deep learning models is five;

[0182] The deep learning model is trained based on the training image set and the validation image set, including:

[0183] Based on each set of training images and the corresponding set of verification images, any deep learning model is trained to obtain a corresponding trained deep learning model;

[0184] Selecting a trained deep learning model from five trained deep learning models based on a preset performance indicator; or

[0185] Combine the five trained deep learning models into one trained deep learning model.

[0186] In some possible embodiments, the deep learning model includes an encoder, a channel attention module, a spatial attention module, a multi-layer perceptron module, and a decoder arranged in sequence, wherein:

[0187] The encoder is used to extract a multi-scale feature sequence from each preprocessed historical pupil image;

[0188] The channel attention module is used to perform channel fusion processing on multi-scale feature sequences;

[0189] The spatial attention module is used to perform spatial fusion processing on the multi-scale feature sequence after channel fusion processing;

[0190] The multi-layer perceptron module is used to perform nonlinear transformation on the multi-scale feature sequence after spatial fusion processing;

[0191] The decoder is used to perform deconvolution processing on the multi-scale feature sequence after nonlinear transformation to obtain the corresponding segmentation result.

[0192] In some possible embodiments, performing sliding slicing processing on the target pupil image to obtain at least two sub-pupil images includes:

[0193] Sliding and slicing the target pupil image according to a preset image axis to obtain at least two sub-pupil images; wherein the size of each sub-pupil image is consistent with a preset model input size;

[0194] According to the cutting order of each sub-pupil image, the corresponding sub-pupil image is labeled.

[0195] In some possible embodiments, all sub-segmentation results are combined to obtain a segmentation result of a target pupil image, including:

[0196] Classify each pixel on each sub-segmentation result, and perform pixel processing on the corresponding sub-segmentation result according to the classification result of each pixel;

[0197] According to the labeling result of each sub-pupil image, all sub-segmentation results after pixel processing are combined to obtain the segmentation result of the target pupil image.

[0198] In some possible embodiments, the pupil image segmentation system based on the deep learning model further includes:

[0199] Based on the preset area, filtering is performed on the segmentation result of the target pupil image;

[0200] The segmentation result of the filtered target pupil image is subjected to edge smoothing.

[0201] The present application also provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the steps of the above method are implemented. The computer-readable storage medium may include, but is not limited to, any type of disk, including a floppy disk, an optical disk, a DVD, a CD-ROM, a micro drive, and a magneto-optical disk, a ROM, a RAM, an EPROM, an EEPROM, a DRAM, a VRAM, a flash memory device, a magnetic card or an optical card, a nanosystem (including a molecular memory IC), or any type of medium or device suitable for storing instructions and / or data.

[0202] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0203] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0204] In the several embodiments provided in the present application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are only schematic, such as the division of units, which is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interfaces, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0205] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0206] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0207] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a memory and includes several instructions for a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of each embodiment method of the present application. The aforementioned memory includes: U disk, read-only memory (ROM), random access memory (RAM), mobile hard disk, disk or optical disk, etc., and other media that can store program codes.

Claims

1. A pupil image segmentation method based on a deep learning model, characterized in that: include: Acquire at least two historical pupil images, and preprocess each of the historical pupil images; Dividing all the preprocessed historical pupil images into a training image set and a verification image set, and training a deep learning model based on the training image set and the verification image set; Performing sliding block processing on the target pupil image to obtain at least two sub-pupil images, and inputting each of the sub-pupil images into a trained deep learning model to obtain corresponding sub-segmentation results; All the sub-segmentation results are combined to obtain a segmentation result of the target pupil image.

2. The method according to claim 1, characterized in that The preprocessing of each of the historical pupil images comprises: Performing background cropping processing on each of the historical pupil images to obtain a corresponding background processed image; Performing pixel normalization processing on each of the background processed images to obtain a corresponding pixel processed image; Performing flip processing on each of the pixel processed images to obtain a corresponding flip processed image; At least two sub-processed images are cropped from each of the flipped processed images so that the size of each of the sub-processed images is consistent with a preset model input size.

3. The method according to claim 2, characterized in that Dividing all the pre-processed historical pupil images into a training image set and a verification image set, including: Dividing all the sub-processed images into five groups of sub-image sets, and determining five groups of training image sets and a verification image set corresponding to each group of the training image sets according to all the sub-image sets; wherein each group of the sub-image sets includes at least two of the sub-processed images; Determine a label image corresponding to each of the sub-processed images, and add the label image corresponding to each of the sub-processed images in each group of the training image set to the corresponding training image set; The label image corresponding to each of the sub-processed images in each group of the verification image set is added to the corresponding verification image set.

4. The method according to claim 3, characterized in that The number of deep learning models is five; The training of the deep learning model based on the training image set and the verification image set includes: Based on each group of the training image sets and the corresponding verification image sets, training any one of the deep learning models to obtain the corresponding trained deep learning model; Selecting one of the trained deep learning models from the five trained deep learning models based on a preset performance indicator; or The five trained deep learning models are combined into one trained deep learning model.

5. The method according to claim 1, characterized in that The deep learning model includes an encoder, a channel attention module, a spatial attention module, a multi-layer perceptron module and a decoder arranged in sequence, wherein: The encoder is used to extract a multi-scale feature sequence from each of the pre-processed historical pupil images; The channel attention module is used to perform channel fusion processing on the multi-scale feature sequence; The spatial attention module is used to perform spatial fusion processing on the multi-scale feature sequence after channel fusion processing; The multi-layer perceptron module is used to perform nonlinear transformation processing on the multi-scale feature sequence after spatial fusion processing; The decoder is used to perform deconvolution processing on the multi-scale feature sequence after nonlinear transformation processing to obtain corresponding segmentation results.

6. The method according to claim 1, characterized in that The step of performing sliding block processing on the target pupil image to obtain at least two sub-pupil images includes: Performing sliding block processing on the target pupil image according to a preset image axis to obtain at least two sub-pupil images; wherein the size of each of the sub-pupil images is consistent with a preset model input size; According to the order in which each sub-pupil image is cut into blocks, the corresponding sub-pupil image is numbered.

7. The method according to claim 6, characterized in that The combining and processing of all the sub-segmentation results to obtain the segmentation result of the target pupil image includes: Performing classification processing on each pixel point on each of the sub-segmentation results, and performing pixel processing on the corresponding sub-segmentation results according to the classification result of each pixel point; According to the labeling result of each sub-pupil image, all the sub-segmentation results after pixel processing are combined to obtain the segmentation result of the target pupil image.

8. The method according to claim 1, characterized in that The method further comprises: Based on a preset area, filtering the segmentation result of the target pupil image; The segmentation result of the target pupil image after filtering is subjected to edge smoothing.

9. A pupil image segmentation system based on a deep learning model, characterized in that: include: An image preprocessing module, used for acquiring at least two historical pupil images and preprocessing each of the historical pupil images; A model training module, used for dividing all the pre-processed historical pupil images into a training image set and a verification image set, and training a deep learning model based on the training image set and the verification image set; A sub-segmentation result prediction module is used to perform sliding block processing on the target pupil image to obtain at least two sub-pupil images, and input each of the sub-pupil images into a trained deep learning model to obtain a corresponding sub-segmentation result; The post-processing module is used to combine all the sub-segmentation results to obtain the segmentation result of the target pupil image.

10. A pupil image segmentation system based on a deep learning model, characterized in that: including a processor and a memory; The processor is connected to the memory; The memory is used to store executable program code; The processor runs a program corresponding to the executable program code by reading the executable program code stored in the memory, so as to execute the steps of the method according to any one of claims 1 to 7.

Images