Choroid segmentation method based on multi-scale feature extraction and adaptive fusion

Abstract

The invention discloses a choroid segmentation method based on multi-scale feature extraction and adaptive fusion. The method comprises the following steps: acquiring a choroid OCT image; the OCT image is sent to a trained choroid segmentation model to obtain a choroid segmentation graph, and a choroid segmentation model construction method comprises the following steps that the choroid OCT image is collected to construct a data set, and the data set is divided into a training set and a test set; a choroid segmentation model framework is constructed, the choroid segmentation model framework comprises a multi-scale feature extraction module and a self-adaptive feature fusion module, boundary segmentation of the choroid segmentation model framework is optimized through a mixed loss function, and the mixed loss function is composed of a cross entropy loss function and a Dice loss item; and inputting the data of the training set and the test set into a network model for training to obtain a trained choroid segmentation model. According to the method, the segmentation precision of the choroid of the OCT image can be effectively improved.

Description

Technical Field

[0001] The present invention relates to a choroid segmentation method based on multi-scale feature extraction and adaptive fusion, belonging to the technical field of medical image processing. Background Art

[0002] As the vascular structure supplying the outer retina, the choroid can exhibit pathological changes of various retinal diseases, such as high myopia, diabetic retinopathy, age-related macular degeneration, glaucoma, etc. The use of high-resolution optical coherence tomography images provides the possibility for clinicians to analyze 3D retinal and choroid tissues. However, this requires clinicians to repeatedly observe and diagnose, which is time-consuming and highly subjective.

[0003] The existing literature (CN117315721A) discloses an OCT image choroid segmentation network model based on a feature integration method. The network model includes a preprocessing module, a feature extraction module, a feature fusion module, and a secondary feature fusion module. First, in the preprocessing module, an anisotropic diffusion filtering algorithm is used for filtering, and an exponential and linear enhancement method is used for enhancement. Then, in the feature extraction module, a residual Inception structure is used, and receptive fields of different sizes are fused to extract rich feature expressions of the image. In the feature fusion process, pyramid pooling is used to further obtain context information, and the channel attention module pays more attention to the relationship between channels, extracts the channel weighted vector of the feature map, and obtains the importance of each channel. In the secondary feature fusion process, the last layer of the feature encoder is fused with the result of the last layer of the feature encoder, and the loss is calculated. This solution can accurately segment the choroid layer by developing an automatic segmentation and layering algorithm, providing a reliable basis for clinical diagnosis and pathological research.

[0004] However, the applicant has at least found the following problems in the actual application of the existing technology:

[0005] First, most current deep learning methods fail to fully utilize the feature extraction capabilities of different convolutional blocks in feature extraction. The network uses convolutional blocks of the same size for feature extraction in the encoder stage, reducing the accuracy and sensitivity of small area and similar category classification.

[0006] Second, the operation of simply fusing low-level and high-level image feature information by the existing segmentation method through skip connections is an unselective combination of local information, which will introduce interference factors such as noise, thus affecting the segmentation effect. Summary of the Invention

[0007] Based on the above, the present invention provides a choroid segmentation method based on multi-scale feature extraction and adaptive fusion to improve the accuracy of choroid segmentation and overcome the deficiencies of the existing technology.

[0008] The technical solution of the present invention is as follows:

[0009] In a first aspect, the present invention provides a choroid segmentation method based on multi-scale feature extraction and adaptive fusion, comprising the following steps:

[0010] Obtain choroidal OCT images;

[0011] Send the OCT image into a trained choroid segmentation model to obtain a choroid segmentation map, where

[0012] The construction method of the choroid segmentation model is as follows:

[0013] Collect choroidal OCT images to construct a dataset, and divide the dataset into a training set and a test set;

[0014] Construct a choroid segmentation model framework, the choroid segmentation model framework includes a multi-scale feature extraction module and an adaptive feature fusion module, and optimizes the boundary segmentation of the choroid segmentation model framework through a hybrid loss function. The hybrid loss function is composed of a cross-entropy loss function and a Dice loss term;

[0015] Input the training set and test set data into the network model for training to obtain a trained choroid segmentation model.

[0016] Preferably, the model framework has an encoder-decoder architecture;

[0017] The multi-scale feature extraction module is composed of a multi-branch structure and a heterogeneous convolution strategy, and replaces the standard convolution block in the encoder part of the encoder-decoder architecture. The multi-branch structure interacts between different channels of the same feature map to obtain features of different scales. The heterogeneous convolution strategy dynamically selects a suitable convolution kernel according to the size of the feature map and the size of the receptive field to achieve a comprehensive extraction of multi-level features of the image;

[0018] The adaptive feature fusion module reconstructs the skip connection in the encoder-decoder architecture based on the spatial attention mechanism, assigns different weights to the feature maps at different stages, and enables the model to focus on the key spatial regions in the feature map;

[0019] The hybrid loss function fuses the output of each layer of the decoder to obtain the final result.

[0020] Preferably, the feature extraction method of the multi-scale feature extraction module includes:

[0021] S1: Convert the input feature map into F(X), a common convolution layer for ontology feature extraction;

[0022] S2: Uniformly divide F(X) into n feature sub - graphs by the split operation, denoted by Xi, where i = {1, 2, 3…, n}; i

[0023] S3: Except for the first feature sub - graph, the remaining groups all go through a convolutional layer, denoted as F( ), where Y represents the convolutional kernel size, and Y can be written as: k () i i

[0024]

[0025] According to the formula, X1 will be directly transmitted to the next stage to retain the information from the input layer;

[0026] S4: Concatenate all the results, denoted as

[0027] S5: Fuse the local features and multi - scale features by addition:

[0028] S6: Gradually increase the size of the convolutional kernel to keep it consistent with the reduction of the feature map resolution and the expansion of the receptive field.

[0029] Preferably, the feature fusion method of the adaptive feature fusion module includes:

[0030] S1: Map the feature maps of all stages to the same channel space as the current stage through a conventional 1×1 convolution;

[0031] S2: Obtain n feature maps by performing a convolution operation on the concatenated feature maps, where n is equal to the number of concatenated feature maps;

[0032] S3: Generate a pixel - direction attention map in the spatial direction through the softmax operator:

[0033]

[0034] In the formula, H×W represents the input image size;

[0035] S4: Obtain the fused feature map: F = A⊙F A +B⊙F B where ⊙ represents element - wise multiplication, and F is the fused feature map.

[0036] Preferably, the hybrid loss function is expressed as:

[0037] Loss = λ1Loss CE +λ2Loss Dcie

[0038] ​​​Among them, λ1 and λ2 are the weights of the loss terms.

[0039] Preferably, the cross-entropy loss function is expressed as:

[0040]

[0041] Among them, N is the number of samples (the number of pixels in the input image); p i is the probability of the i-th pixel target category predicted by the model; y i is the true label value of the i-th pixel, that is, 0 or 1.

[0042] Preferably, the Dice loss term is expressed as:

[0043]

[0044] Among them, N is the total number of pixels; p i is the probability of the i-th pixel target category predicted by the model; y i is the true label value of the i-th pixel, that is, 0 or 1.

[0045] In a second aspect, the present invention provides an electronic device, including: a memory on which program code is stored; a processor coupled to the memory, and when the program code is executed by the processor, a choroid segmentation method based on multi-scale feature extraction and adaptive fusion is implemented.

[0046] In a third aspect, the present invention provides a computer-readable storage medium on which program instructions are stored, and when the program instructions are executed, a choroid segmentation method based on multi-scale feature extraction and adaptive fusion is implemented.

[0047] In a fourth aspect, the present invention provides a computer program product, including computer programs / instructions, and when the computer programs / instructions are executed by a processor, a choroid segmentation method based on multi-scale feature extraction and adaptive fusion is implemented.

[0048] Advantages of the present invention: The present invention can effectively improve the segmentation accuracy of the choroid layer. The multi-scale feature extraction module constructed by the multi-branch structure and the heterogeneous convolution strategy can dynamically adjust the convolution kernel size according to the feature map size and receptive field, adapting to the size and receptive field requirements of different-level feature maps. Small convolution kernels are used in the shallow layer to capture local details, and large convolution kernels are used in the deep layer to extract global information, effectively balancing the detailed and abstract features, significantly enhancing the model's recognition ability for small choroid regions and complex structures, and improving the segmentation accuracy. The adaptive feature fusion module constructed based on the spatial attention mechanism assigns dynamic weights to the feature maps at different stages, enabling the model to focus on the key regions of the input image or feature map, enhancing the model's processing ability for important information. Compared with the simple splicing of traditional skip connections, this module realizes selective information fusion, enhancing the extraction ability of important features, thereby improving the accuracy and robustness of the segmentation results. By combining the cross-entropy loss and the Dice loss in a hybrid loss function, the problem of background pixel dominance is effectively alleviated, improving the robustness of the segmentation model. Especially in the case of low contrast between the choroid and surrounding tissues, the phenomenon of low boundary segmentation accuracy caused by edge blurring or breaking is significantly reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic diagram of the choroid segmentation model framework;

[0050] Figure 2 It is a schematic diagram of the multi-branch structure;

[0051] Figure 3 It is a schematic diagram of the heterogeneous convolution strategy;

[0052] Figure 4 It is a schematic diagram of the adaptive fusion module;

[0053] Figure 5 It is a schematic diagram of the OCT image and the choroid segmentation result;

[0054] Figure 6 It is a graph of the choroid layer segmentation boundary results of each algorithm. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0055] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be given with reference to the accompanying drawings. Many specific details are set forth in the following description to facilitate a thorough understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0056] Example 1

[0057] A choroid segmentation method based on multi-scale feature extraction and adaptive fusion, as Figures 1 to 5 shown, the method includes the following steps:

[0058] S1 Obtain choroid OCT images;

[0059] S2 Feed the OCT image into the trained choroid segmentation model to obtain a choroid segmentation map.

[0060] Among them, the construction method of the above choroid segmentation model is as follows:

[0061] S21 Collect a certain number of choroid OCT images to construct a dataset, and divide the dataset into a training set and a test set;

[0062] S22 Construct a choroid segmentation model framework, which includes a multi-scale feature extraction module and an adaptive feature fusion module, and optimize the boundary segmentation of the choroid segmentation model framework through a mixed loss function.

[0063] First, refer to Figure 1 , and build a choroid segmentation model framework with an encoder-decoder architecture.

[0064] Secondly, a multi-scale feature extraction module is composed of a multi-branch structure and a heterogeneous convolution strategy to replace the standard convolution block in the encoder part of the encoder-decoder architecture. The multi-branch structure groups the feature maps by channel, connects different groups as cross-stage connections, and retains the information from the previous layer. Finally, all groups are connected together to interact between all groups, so that the multi-branch structure interacts between different channels of the same feature map to obtain features of different scales; the heterogeneous convolution strategy refers to dynamically adjusting the size of the convolution to adapt to the size and receptive field requirements of feature maps at different levels. The size of the convolution kernel gradually increases to keep it consistent with the reduction of the feature map resolution and the expansion of the receptive field, which can effectively balance the extraction of local details and global information, so as to realize the comprehensive extraction of multi-level features of the image.

[0065] Considering the computing resources and computing costs, the spilt operation in the multi-branch structure uniformly divides the feature map into 4 groups; the heterogeneous convolution strategy dynamically selects a suitable convolution kernel according to the size of the feature map and the receptive field. In the encoder part of the network, as the number of network layers increases, the size of the feature map gradually decreases, and the degree of abstraction of the features gradually increases. The size of the shallow feature map is large, corresponding to a small receptive field, so a small convolution kernel is used to extract specific information; the size of the deep feature map is small, corresponding to a large receptive field, so a large convolution kernel is used to extract abstract information.

[0066] The feature extraction method of the multi-scale feature extraction module is as follows:

[0067] A1: Convert the input feature map into F(X), which is a common convolutional layer for ontology feature extraction;

[0068] A2: Divide F(X) evenly into n feature sub - maps by channels through the split operation, denoted as X i where i = {1, 2, 3…, n};

[0069] A3: Except for the first feature sub - map, the remaining groups all go through a convolutional layer, denoted as F k (), where Y i represents the convolutional kernel size, and Y i can be written as:

[0070]

[0071] According to the formula, X1 will be directly transmitted to the next stage to retain the information from the input layer;

[0072] A4: Concatenate all the results, denoted as

[0073] A5: Fuse local features and multi - scale features by addition:

[0074] Specifically, refer to Figure 2 , the multi - branch structure groups the same feature map by channels, each branch independently processes the features of different groups, the different groups are connected as cross - stage connections, and the information from the previous layer is retained. Finally, all the groups are connected together and a 1×1 convolution is applied to interact between all the groups. Each group encodes features of different scales, reducing cross - channel redundant calculations and enhancing local feature interaction.

[0075] A6: The size of the convolutional kernel gradually increases (from 3x3 to 5x5 and then to 7x7) to be consistent with the reduction of the feature map resolution and the expansion of the receptive field.

[0076] Specifically, refer to Figure 3 , the heterogeneous convolution strategy dynamically selects a suitable convolutional kernel according to the size of the feature map and the size of the receptive field. In the encoder part of the network, as the number of network layers increases, the size of the feature map gradually decreases and the degree of feature abstraction gradually increases. The size of the shallow - layer feature map is larger and the corresponding receptive field is smaller, so a small convolutional kernel is used to extract specific information; the size of the deep - layer feature map is smaller and the corresponding receptive field is larger, so a large convolutional kernel is used to extract abstract information.

[0077] Thirdly, the adaptive feature fusion module reconstructs the skip connections in the encoder-decoder architecture based on the spatial attention mechanism. The adaptive feature fusion module uses the spatial attention mechanism to generate an attention map, representing the importance of each spatial position. The adaptive fusion module uses the spatial attention mechanism to fuse the feature maps of the current stage with the feature maps of all higher stages instead of the skip connections, assigns different weights to the feature maps of different stages, enables the model to focus on the key spatial regions in the feature maps, and enhances the module's processing ability for important information.

[0078] See Figure 4 Figure, the feature fusion method of the adaptive feature fusion module is as follows:

[0079] B1: Map the feature maps of all stages to the same channel space as the current stage through a conventional 1×1 convolution;

[0080] B2: Obtain n feature maps after convolving the concatenated feature maps once, where n is equal to the number of concatenated feature maps;

[0081] B3: Generate a pixel-wise attention map in the spatial direction through the softmax operator:

[0082]

[0083] In the formula, H×W represents the size of the input image;

[0084] B4: Obtain the fused feature map: F = A⊙F A +B⊙F B , where ⊙ represents element-wise multiplication, and F is the fused feature map.

[0085] Finally, the outputs of each layer of the decoder are fused by the hybrid loss function to obtain the final result. The hybrid loss function consists of the cross-entropy loss function and the Dice loss term, to optimize the boundary segmentation, improve the boundary segmentation accuracy in the choroid segmentation process, so that the hybrid loss function effectively alleviates the problem dominated by background pixels, and directly optimizes the overlapping ratio of the target region rather than the per-pixel error, improving the robustness of the segmentation model.

[0086] The above hybrid loss function is expressed as:

[0087] Loss = λ1Loss CE +λ2Loss Dcie

[0088] where λ1 and λ2 are the weights of the loss terms.

[0089] The above cross-entropy loss function is expressed as:

[0090]

[0091] Among them, N is the number of samples (the number of pixels in the input image); p i is the probability of the i-th pixel target category predicted by the model; y i is the true label value of the i-th pixel, that is, 0 or 1.

[0092] The above Dice loss term is expressed as:

[0093]

[0094] Among them, N is the total number of pixels; p i is the probability of the i-th pixel target category predicted by the model; y i is the true label value of the i-th pixel, that is, 0 or 1.

[0095] S23. Input the training set and test set data into the network model for training to obtain a trained choroid segmentation model.

[0096] Refer to Figure 5 For the choroid layer segmentation result diagrams, from left to right are: the original OCT image, the choroid segmentation result diagram.

[0097] The following verifies the effect of the method of the present invention through specific experiments:

[0098] I. Dataset

[0099] The dataset used for experiments and verification comes from the Baidu PaddlePaddle Learning Competition GOALS dataset. The GOALS dataset is provided by the Zhongshan Ophthalmic Center of Sun Yat-sen University in Guangzhou, China. The dataset contains 100 peripapillary circular scan OCT data. To increase the diversity of the data, data augmentation is performed on the dataset. In this experiment, the images are rotated to 16 different angles, and the rotated images are randomly cropped; the images are also flipped at each angle. Finally, an augmented training set 32 times larger than the non-augmented training set is obtained. Among them, 70% of the data is used for training, 10% for verification, and 20% for testing. To objectively evaluate the performance of the method of the present invention, a total of 4 evaluation indicators are used, namely accuracy (Accuracy, Acc), sensitivity (Sensitivity, Sen), Dice coefficient, and Mean Intersection over Union (MIoU).

[0100] II. Experimental settings

[0101] All experiments were carried out under the PyTorch framework and ran on an NVIDIA RTX 4070 (12GB) GPU. In this experiment, fundus OCT B-Scan images were divided into two categories: choroid category and background category. At the beginning of the training of the algorithm of the present invention, Stochastic Gradient Descent (SGD) was used to update the parameters of the network, with the momentum set to 0.9, the batch size set to 4, and the weight decay set to 0.0001. Compared with batch gradient descent, SGD has less computing time. Mini-batches can effectively balance the convergence speed and stability, providing better generalization performance. The number of iterations was set to 100. Due to the existence of overfitting, the model trained to 100 epochs was not directly saved. Instead, after each round of training and validation, it was judged: if the performance of this round of the model exceeded the previous best performance, the model parameter variables were updated and saved. The methods of SegNet, DeepLabV3+, U-Net, UNetV2, UNetV2, and SA_DCACN were used as controls.

[0102] III. Experimental Results and Analysis

[0103] The segmentation results of each treatment group are as Figure 6 shown, Figure 6 The choroid layer segmentation boundary result diagrams in are respectively: the original OCT image, the segmentation result diagram of SegNet, the segmentation result diagram of DeepLabV3+, the segmentation result diagram of U-Net, the segmentation result diagram of SA_DCACN, and the segmentation result diagram of the algorithm of this article. Table 1 shows the results and comparisons of the comparative experiments on choroid segmentation. As shown in Table 1, the choroid segmentation method of the present invention performs optimally in all evaluation indicators. It is worth noting that compared with DeepLabV3+, the method of the present invention achieves maximum improvements of approximately 6% and 7% in sensitivity (Sen) and Dice coefficient respectively. The experimental results show that the present invention can clearly segment the boundary of the choroid layer, providing support for the early diagnosis and quantitative analysis of diseases such as diabetic retinopathy and glaucoma.

[0104] Table 1 Evaluation Indicators of the Segmentation Results of the Present Invention and Existing Neural Network Models

[0105]

[0106] In summary, the present invention can achieve accurate segmentation of the choroid layer in OCT images at one time. Through the multi-scale feature extraction module with a multi-branch structure and heterogeneous convolution strategy, the model's multi-scale feature extraction ability for the choroid layer is strengthened; the adaptive fusion module based on the spatial attention mechanism enables the model to focus on the key regions of the input image or feature map, thereby enhancing the model's processing ability for important information. The use of the hybrid loss function effectively alleviates the problem of background pixel dominance and improves the robustness of the segmentation model.

[0107] Embodiment 2

[0108] An electronic device includes, but is not limited to: a memory having program code stored thereon; a processor coupled to the memory, and when the program code is executed by the processor, a choroid segmentation method based on multi-scale feature extraction and adaptive fusion is implemented. For a detailed description of the method, reference may be made to the corresponding description in the above method embodiments, and details are not repeated here.

[0109] Embodiment 3

[0110] A computer-readable storage medium has program instructions stored thereon, and when the program instructions are executed, a choroid segmentation method based on multi-scale feature extraction and adaptive fusion is implemented. For a detailed description of the method, reference may be made to the corresponding description in the above method embodiments, and details are not repeated here.

[0111] Embodiment 4

[0112] A computer program product includes a computer program / instructions, and when the computer program / instructions are executed by a processor, a choroid segmentation method based on multi-scale feature extraction and adaptive fusion is implemented. For a detailed description of the method, reference may be made to the corresponding description in the above method embodiments, and details are not repeated here.

[0113] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but should not be construed as a limitation on the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the appended claims.

Claims

1. A choroid segmentation method based on multi-scale feature extraction and adaptive fusion, characterized in that It includes the following steps: Obtain choroidal OCT images; Send the OCT images into a trained choroidal segmentation model to obtain a choroidal segmentation map. Among them, the construction method of the choroidal segmentation model is as follows: Collect choroidal OCT images to construct a dataset, and divide the dataset into a training set and a test set; Construct a choroidal segmentation model framework. The choroidal segmentation model framework includes a multi-scale feature extraction module and an adaptive feature fusion module, and optimizes the boundary segmentation of the choroidal segmentation model framework through a hybrid loss function. The hybrid loss function consists of a cross-entropy loss function and a Dice loss term; Input the training set and test set data into the network model for training to obtain a trained choroidal segmentation model.

2. The choroidal segmentation method based on multi-scale feature extraction and adaptive fusion according to claim 1, wherein The model framework is based on an encoder-decoder architecture; The multi-scale feature extraction module is composed of a multi-branch structure and a heterogeneous convolution strategy, which replaces the standard convolution block in the encoder part of the encoder-decoder architecture. The multi-branch structure interacts between different channels of the same feature map to obtain features of different scales. The heterogeneous convolution strategy dynamically selects a suitable convolution kernel according to the size of the feature map and the receptive field to achieve comprehensive extraction of multi-level features of the image; The adaptive feature fusion module reconstructs the skip connection in the encoder-decoder architecture based on the spatial attention mechanism, assigns different weights to the feature maps at different stages, and enables the model to focus on the key spatial regions in the feature map; The hybrid loss function fuses the output of each layer of the decoder to obtain the final result.

3. The choroid segmentation method based on multi-scale feature extraction and adaptive fusion according to claim 1, wherein The feature extraction method of the multi-scale feature extraction module includes: S1: Convert the input feature map into F(X), which is a common convolution layer for ontology feature extraction; S2: Uniformly divide F(X) into n feature sub - graphs by channel through the split operation, denoted by Xi, where i = {1, 2, 3…, n}; i ​ S3: Except for the first feature sub - graph, the remaining groups all pass through a convolutional layer, denoted as F k (), where, Y i represents the convolutional kernel size, Y i can be written as: According to the formula, X1 will be directly transmitted to the next stage to retain the information from the input layer; S4: Concatenate all the results and denote them as S5: Fuse the local features and multi-scale features by addition: S6: The size of the convolution kernel gradually increases to keep it consistent with the reduction of the feature map resolution and the expansion of the receptive field.

4. The choroid segmentation method based on multi-scale feature extraction and adaptive fusion according to claim 1, wherein The feature fusion method of the adaptive feature fusion module includes: S1: Map the feature maps of all stages to the same channel space as the current stage through a conventional 1×1 convolution; S2: Pass the concatenated feature maps through a convolution operation to obtain n feature maps, where n is equal to the number of concatenated feature maps; S3: Generate a pixel-wise attention map in the spatial direction through the softmax operator: In the formula, H×W represents the size of the input image; S4: Obtain the fused feature map: F = A ⊙ F A + B ⊙ F B , where ⊙ represents element-wise multiplication, and F is the fused feature map.

5. The choroid segmentation method based on multi-scale feature extraction and adaptive fusion according to claim 1, wherein The hybrid loss function is expressed as: Loss = λ1Loss CE + λ2Loss Dcie Where λ1 and λ2 are loss term weights.

6. The choroid segmentation method based on multi-scale feature extraction and adaptive fusion according to claim 5, wherein The cross-entropy loss function is expressed as: where N is the number of samples (number of pixels in the input image); p i is the probability of the i-th pixel target class predicted by the model; y i is the true label value of the i-th pixel, i.e., 0 or 1.

7. The choroid segmentation method based on multi-scale feature extraction and adaptive fusion according to claim 5, wherein The Dice loss term is expressed as: where N is the total number of pixels; p i is the probability of the i-th pixel's target class predicted by the model; y i is the true label value of the i-th pixel, i.e., 0 or 1.

8. An electronic device, characterized in that, It includes: A memory, on which program code is stored; A processor, which is connected to the memory, and when the program code is executed by the processor, the method described in any one of claims 1 to 7 is implemented.

9. A computer-readable storage medium, characterized in that, Program instructions are stored thereon, and when the program instructions are executed, the method described in any one of claims 1 to 7 is implemented.

10. A computer program product, comprising computer programs / instructions, characterized in that, When the computer program / instructions are executed by the processor, the method described in any one of claims 1 to 7 is implemented.

Images