High-precision image binary segmentation method and system based on model lightweight

By replacing Swin-Large with Swin-Tiny and adopting knowledge distillation strategy, the problem of excessive parameters of high-precision image segmentation model is solved, and the model parameters are reduced and performance improvement is achieved, which is suitable for mobile devices and embedded systems.

CN119131394APending Publication Date: 2024-12-13NANKAI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411257808.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-09
Publication Date
2024-12-13

AI Technical Summary

Technical Problem

The existing high-precision image segmentation model has limited its deployment and image processing speed on mobile devices or embedded systems due to its complex structure and excessive parameters, which has affected the widespread application of technology.

Method used

Replace the Swin-Large skeleton network with Swin-Tiny, and adopt a knowledge distillation strategy to guide students' network training by supervising the feature map of residual connection blocks, reducing model parameters, and improving image segmentation performance.

Benefits of technology

A significant reduction in the amount of model parameters is achieved while minimizing performance losses, improving the efficiency and performance of image segmentation, especially on mobile devices.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119131394A_ABST
    Figure CN119131394A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of image processing, and provides a high-precision image binary segmentation method and system based on model lightweight, and the method comprises the steps: obtaining a to-be-segmented image; segmenting the to-be-segmented image through a student network to obtain a binary image; wherein the training method of the student network comprises the following steps: taking an original image in a training set as the input of the teacher network and the student network at the same time, and enabling the teacher network to guide the student network to train through a feature map obtained by supervising a residual error connection block by adopting a knowledge distillation strategy; the teacher network and the student network both adopt a binary segmentation model based on a bilateral reference framework, the teacher network takes Swindow-Large as a skeleton network, and the student network takes Swindow-Tiny as a skeleton network. The model parameter quantity is reduced, and the image segmentation performance loss is reduced to the minimum.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a high-precision image binary segmentation method and system based on lightweight model. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] In the field of high-precision image segmentation, although existing computer vision segmentation models can achieve foreground object extraction and background removal of complex images with hair-level accuracy, this is at the cost of an extremely complex network structure. Model network structures with more parameters will not only greatly limit the deployment of high-precision segmentation models on mobile devices or embedded systems, but will also seriously slow down image processing speed, thus affecting the widespread application of this technology. Summary of the invention

[0004] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a high-precision image binary segmentation method and system based on model lightweight, replaces the Swin-Large skeleton used in the binary segmentation model based on the bilateral reference frame with Swin-Tiny to obtain a student network, and adopts a knowledge distillation strategy to use Swin-Large as the teacher network of the skeleton network, and guides the student network to be trained by supervising the feature map of the residual connection block, so that the model parameters are reduced while the image segmentation performance loss is minimized.

[0005] In order to achieve the above object, the present invention adopts the following technical solution: A first aspect of the present invention provides a high-precision image binary segmentation method based on lightweight model, comprising: Get the image to be segmented.

[0006] For the image to be segmented, segment it through the student network to obtain a binary image; Among them, the training method of the student network is: the original image in the training set is used as the input of the teacher network and the student network at the same time, and the knowledge distillation strategy is adopted. The feature map obtained by supervising the residual connection block is used to enable the teacher network to guide the student network for training; the teacher network and the student network both adopt a binary segmentation model based on a bilateral reference frame, the teacher network uses Swin-Large as the skeleton network, and the student network uses Swin-Tiny as the skeleton network.

[0007] Furthermore, the feature map obtained by the residual connection block of the teacher network is aligned with the number of channels of the feature map obtained by the residual connection block of the student network through the adaptation layer.

[0008] Furthermore, the binary segmentation model based on the bilateral reference frame includes four transformer encoders connected in sequence and a dilated spatial convolution pooling pyramid; the output features of the first transformer encoder are simultaneously used as inputs of the second transformer encoder and the fourth transformer encoder; the output features of the second transformer encoder are simultaneously used as inputs of the third transformer encoder and the fourth transformer encoder; the output features of the third transformer encoder are only used as inputs of the fourth transformer encoder; and the output features of the fourth transformer encoder are used as inputs of the dilated spatial convolution pooling pyramid.

[0009] Furthermore, the binary segmentation model based on the bilateral reference frame includes three bidirectional reference modules connected in sequence, and each bidirectional reference module is connected to a residual connection block; the output features of the third transformer encoder are passed through a convolution layer and combined with the output features of the first bidirectional reference module to serve as the input of the first residual connection block; the output features of the second transformer encoder are passed through a convolution layer and combined with the output features of the second bidirectional reference module to serve as the input of the second residual connection block; the output features of the first transformer encoder are passed through a convolution layer and combined with the output features of the third bidirectional reference module to serve as the input of the third residual connection block.

[0010] Furthermore, the bidirectional reference module includes an internal reference module, a reconstruction module and an external reference module connected in sequence; after the internal reference module processes the input features of the bidirectional reference module, the input features are input into the reconstruction module together with the input features of the bidirectional reference module; after the output features of the reconstruction module are processed by the external reference module and the activation function, the output features of the reconstruction module are multiplied at the pixel level with the output features of the reconstruction module to obtain the output features of the bidirectional reference module.

[0011] Furthermore, the output features of the third residual connection block pass through a predictor to obtain the binary image.

[0012] Furthermore, the internal reference module of the student network adopts an attention internal reference, which crops the original image into image blocks of the same size as the input features, stacks the image blocks with the input features, and processes them through a convolutional attention module to obtain output features.

[0013] A second aspect of the present invention provides a high-precision image binary segmentation system based on lightweight model, comprising: The training module is configured as follows: the original image in the training set is used as the input of the teacher network and the student network at the same time, and the knowledge distillation strategy is adopted to enable the teacher network to guide the student network to train by supervising the feature map obtained by the residual connection block; the teacher network and the student network both adopt a binary segmentation model based on a bilateral reference frame, the teacher network uses Swin-Large as the skeleton network, and the student network uses Swin-Tiny as the skeleton network; An image acquisition module is configured to: acquire an image to be segmented; The image segmentation module is configured to: segment the image to be segmented through the student network to obtain a binary image.

[0014] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the high-precision image binary segmentation method based on lightweight model as described above.

[0015] The fourth aspect of the present invention provides a computer device, comprising a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, wherein when the processor executes the program, the steps in the high-precision image binary segmentation method based on lightweight model as described above are implemented.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The present invention replaces the Swin-Large skeleton used in the binary segmentation model based on the bilateral reference frame with Swin-Tiny to obtain a student network, and adopts a knowledge distillation strategy to use Swin-Large as the teacher network of the skeleton network. By supervising the feature map of the residual connection block, the student network is guided to train, and a lightweight and high-precision segmentation model is obtained, which reduces the number of model parameters while minimizing the performance loss.

[0017] The present invention improves the inward reference module and adds a CBAM attention module after feature stacking to enhance the model's perception of the importance of high-definition image blocks at different positions. For image blocks containing the main part of the image, the attention mechanism will increase the weight of the block, and reduce the weight of image blocks that do not contain the main part of the image or have a small proportion of the main part, so as to help the reconstruction module better utilize the high-definition information of the original image. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0019] Figure 1is a schematic diagram of a teacher network and a student network according to the first embodiment of the present invention; Figure 2 is a schematic diagram of an attention inward reference according to the first embodiment of the present invention; Figure 3 is a structural diagram of the BiRefNet model of the first embodiment of the present invention; Figure 4 1 is a comparison diagram of BiRefNet and other high-precision segmentation models according to the first embodiment of the present invention; Figure 5 It is a structural diagram of a computer device according to a fourth embodiment of the present invention. DETAILED DESCRIPTION

[0020] To make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0021] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0022] Embodiment 1 This embodiment provides a high-precision image binary segmentation method based on lightweight model.

[0023] This embodiment provides a high-precision image binary segmentation method based on lightweight model, which is committed to solving the problem of overly complex structure and excessive parameters of existing high-precision image segmentation network models, and ensures that the number of parameters is reduced while the model performance is not greatly affected by improving the network module.

[0024] BiRefNet (binary segmentation model based on bilateral reference frame) is currently the best high-precision binary segmentation model. Its model structure is as follows: Figure 3 As shown in Figure 2, its effect is compared with other mainstream high-precision segmentation models. Figure 4 shown.

[0025] Therefore, this embodiment provides a high-precision image binary segmentation method based on lightweight model, which makes the following improvements on the basis of the baseline model proposed by BiRefNet: (1) Adding the CBAM attention mechanism to the inward reference part of the bidirectional reference module makes it easier for the input features to obtain high-definition detail information of the main target of the image during the inward reference process; (2) The Swin-Large skeleton network used in the object localization module was replaced with the Swin-Tiny model, reducing the number of model parameters by 75%. A knowledge distillation strategy was adopted, with the BiRefNet baseline model as the teacher network. The feature maps of the three residual connection blocks in the supervised reconstruction module were used to guide the model training.

[0026] This embodiment provides a high-precision image binary segmentation method based on lightweight model, and its complete framework is as follows: Figure 1 shown.

[0027] (1) Add an inward reference module to the CBAM attention mechanism.

[0028] In the original BiRefNet's internal reference module, the original image is first cropped into image blocks of the same size as the input features. In the original BiRefNet, these image blocks are stacked with the input features and directly passed to the reconstruction module for feature discovery. However, the deformable convolutional layer used in the reconstruction module is actually difficult to discover which image blocks are more helpful for the input features to obtain high-definition detail information. For example, some image blocks do not contain the main part of the image, that is, the image blocks cannot provide useful high-definition information for the model, but the reconstruction module cannot distinguish the importance of image blocks.

[0029] This embodiment provides a high-precision image binary segmentation method based on lightweight model, such as Figure 2 As shown in the figure, the inward reference module is improved, and the CBAM attention module is added after the feature stacking to enhance the model's ability to perceive the importance of high-definition image blocks at different positions. For image blocks containing the main part of the image, the attention mechanism will increase the weight of the block, and reduce the weight of image blocks that do not contain the main part of the image or the main part accounts for a small proportion, so as to help the reconstruction module better utilize the high-definition information of the original image. It is worth noting that the number of parameters required to build the CBAM module is very small, which allows the model performance to be improved without increasing the computing power requirements.

[0030] (2) Perform knowledge distillation on BiRefNet.

[0031] This embodiment provides a high-precision image binary segmentation method based on lightweight model, which adjusts the backbone network of BiRefNet from Swin-Large to Swin-Tiny.

[0032] After adjusting the backbone network of BiRefNet from Swin-Large to Swin-Tiny, although the complexity of the model is significantly reduced and the consumption of computing resources is reduced, the reduction in the number of network parameters will also lead to a decrease in the generalization ability of the model. In order to enhance the performance of the small model, this embodiment adopts a knowledge distillation strategy to guide the learning of the small model by learning the features of the large model.

[0033] Therefore, this embodiment provides a high-precision image binary segmentation method based on model lightweighting, which uses the BiRefNet model with Swin-Large as the skeleton network as the teacher network, thereby guiding the model with Swin-Tiny as the skeleton network to learn the knowledge and feature representation of the baseline model during the training process. The teacher network supervises the student network through the feature maps of the three residual connection blocks in the reconstruction module.

[0034] Specifically, if Figure 1 As shown in the figure, the corresponding feature map of the student network will first pass through an adaptation layer with a convolution kernel size of 1×1, so that the number of channels of the student network and the teacher network feature map are aligned; then, the model will calculate the loss value and back-propagate through the L2 loss function, which is used to measure the difference between the feature map of the student network and the feature map of the teacher network, thereby guiding the student network to learn the knowledge and feature representation of the teacher network. It is worth noting that only three adaptation layers are introduced for knowledge distillation, and in the reasoning process of the model, the three adaptation layers have no effect on the output prediction results of the network, so the introduction of the knowledge distillation strategy will not increase the number of network parameters, that is, it will not increase the computational overhead of the network.

[0035] This embodiment provides a high-precision image binary segmentation method based on lightweight model, including the following steps: Step 1: Get the training set.

[0036] The training set includes original images and labeled binary images.

[0037] Step 2: Use the original images in the training set as the input of both the teacher network and the student network. Use the knowledge distillation strategy to supervise the feature maps obtained by the three residual connection blocks in the segmentation map reconstruction module, so that the teacher network guides the student network for training and obtains the student network (lightweight and high-precision segmentation model).

[0038] Among them, the teacher network is a pre-trained network.

[0039] It should be noted that both the teacher network and the student network include an object localization module, a segmentation map reconstruction module and a predictor which are connected in sequence.

[0040] The object localization module includes four sequentially connected Transformer blocks (also known as Transformer Encoders) and an ASPP (Dilated Spatial Convolution Pooling Pyramid). The output of the first Transformer block is used as the input of both the second and fourth Transformer blocks; the output of the second Transformer block is used as the input of both the third and fourth Transformer blocks; the output of the third Transformer block is only used as the input of the fourth Transformer block.

[0041] The segmentation map reconstruction module includes three bidirectional reference modules connected in sequence, and each bidirectional reference module is connected to a residual connection block. The output of the third Transformer structure block passes through a 1×1 convolution layer and is combined with the output of the first bidirectional reference module as the input of the first residual connection block; the output of the second Transformer structure block passes through a 1×1 convolution layer and is combined with the output of the second bidirectional reference module as the input of the second residual connection block; the output of the first Transformer structure block passes through a 1×1 convolution layer and is combined with the output of the third bidirectional reference module as the input of the third residual connection block.

[0042] Among them, the predictor is a 1×1 convolution layer, and the output features of the third residual connection block pass through the predictor to obtain the segmentation result (binary image).

[0043] Among them, the bidirectional reference module includes an internal reference module, a reconstruction module and an external reference module connected in sequence; after the internal reference module processes the input features of the bidirectional reference module, it is input into the reconstruction module together with the input features of the bidirectional reference module; after the output features of the reconstruction module are processed by the external reference module and the activation function, they are multiplied at the pixel level with the output features of the reconstruction module to obtain the output features of the bidirectional reference module.

[0044] The difference is that the teacher network is a BiRefNet model with Swin-Large as the skeleton network (object positioning module), and the student network is a BiRefNet model with Swin-Tiny as the skeleton network. And the feature map obtained by the residual connection block of the teacher network is aligned with the number of channels of the feature map obtained by the residual connection block of the student network through the adaptation layer. The difference between Swin-Large and Swin-Tiny is that Swin-Large has more model parameters.

[0045] In addition, the inward reference modules in the bidirectional reference modules of the student network and the teacher network both use the inward reference, e.g. Figure 2As shown in the figure, the intra-attention reference crops the original image into image blocks of the same size as the input features, stacks the image blocks with the input features, and processes them through the CBAM attention module (also known as the convolutional attention module) before passing them to the reconstruction module for feature discovery.

[0046] Step 3: Get the image to be segmented.

[0047] Step 4: For the image to be segmented, the student network trained in step 2 is used to obtain the segmentation result (ie, binary image).

[0048] In order to verify the effectiveness of the BiRefNet model optimization strategy, on the DIS5K dataset, this example uses the BiRefNet model with Swin-Tiny as the backbone network without any optimization strategy as the baseline, and adds the CBAM attention mechanism module and knowledge distillation strategy to the baseline model. After training for 180 rounds using exactly the same training parameters, the optimal round weights of the three methods are compared, and the S-measure ( ), maximum / weighted F-measure ( / ), mean absolute error MAE (M), E-measure ( ) and human correction amount (HCE) to evaluate model performance from different perspectives. For example, pixel-level metrics such as M focus on emphasizing the accuracy of the model, while average-based metrics Can better reflect the requirements of practical applications, structural metrics The results are shown in Table 1.

[0049] Table 1. Comparison results

[0050] In Table 1, ”(“ "") indicates that the higher (lower) the better the indicator is. The BASE model refers to the unoptimized BiRefNet model with Swin-Tiny as the backbone network; CBAM refers to the optimized model with the CBAM attention mechanism added on the basis of the BASE model; DISTILL refers to the optimized model with the knowledge distillation strategy added on the basis of the BASE model; WHOLE refers to the model that uses all optimization mechanisms.

[0051] As can be seen from Table 1, as the image complexity of the DIS-TE1 to DIS-TE4 evaluation data sets gradually increases, the two optimization strategies proposed in this embodiment gradually widen the gap with the baseline in various indicators, especially in the indicator of human correction (HCE). On the DIS-TE4 evaluation data set, the optimization model that deploys the knowledge distillation strategy improves the HCE indicator by more than 3% compared with the baseline model. Such data shows that both optimization methods proposed in this article can effectively improve the performance of lightweight BiRefNet, and especially when the model is put into actual use, a lower HCE indicator means that the user experience will be effectively improved.

[0052] In summary, this embodiment proposes a high-precision image segmentation lightweight method, which reduces the number of parameters of the best model in this field by 75% while minimizing the performance loss.

[0053] Embodiment 2 This embodiment provides a high-precision image binary segmentation system based on lightweight model, which specifically includes: The training module is configured as follows: the original image in the training set is used as the input of the teacher network and the student network at the same time, and the knowledge distillation strategy is adopted to enable the teacher network to guide the student network to train by supervising the feature map obtained by the residual connection block; the teacher network and the student network both adopt a binary segmentation model based on a bilateral reference frame, the teacher network uses Swin-Large as the skeleton network, and the student network uses Swin-Tiny as the skeleton network; An image acquisition module is configured to: acquire an image to be segmented; The image segmentation module is configured to: segment the image to be segmented through the student network to obtain a binary image.

[0054] It should be noted here that each module in this embodiment corresponds to each step in Example 1 one by one, and the specific implementation process is the same, which will not be repeated here.

[0055] Embodiment 3 This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the high-precision image binary segmentation method based on model lightweighting as described in the above-mentioned embodiment 1 are implemented.

[0056] Embodiment 4 This embodiment provides a computer device, such as Figure 5As shown, it includes a computer-readable storage medium 1003, a processor 1001, a communication interface 1002, and a computer program stored on the computer-readable storage medium 1003 and executable on the processor 1001, wherein the processor 1001, the communication interface 1002, and the computer-readable storage medium 1003 can be connected via a bus or other means. The communication interface 1002 is used to receive and send data, and when the processor 1001 executes the program, the steps in the high-precision image binary segmentation method based on model lightweight as described in the first embodiment are implemented.

[0057] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0058] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0059] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A high-precision image binary segmentation method based on lightweight model, characterized in that: include: Get the image to be segmented. For the image to be segmented, segment it through the student network to obtain a binary image; Among them, the training method of the student network is: the original image in the training set is used as the input of the teacher network and the student network at the same time, and the knowledge distillation strategy is adopted. The feature map obtained by supervising the residual connection block is used to enable the teacher network to guide the student network for training; the teacher network and the student network both adopt a binary segmentation model based on a bilateral reference frame, the teacher network uses Swin-Large as the skeleton network, and the student network uses Swin-Tiny as the skeleton network.

2. A high-precision image binary segmentation method based on lightweight model as claimed in claim 1, characterized in that: The feature map obtained by the residual connection block of the teacher network is aligned with the number of channels of the feature map obtained by the residual connection block of the student network through the adaptation layer.

3. The high-precision image binary segmentation method based on lightweight model according to claim 1, characterized in that: The binary segmentation model based on the bilateral reference frame includes four transformer encoders connected in sequence and a dilated spatial convolution pooling pyramid; the output features of the first transformer encoder are simultaneously used as inputs of the second transformer encoder and the fourth transformer encoder; the output features of the second transformer encoder are simultaneously used as inputs of the third transformer encoder and the fourth transformer encoder; the output features of the third transformer encoder are only used as inputs of the fourth transformer encoder; and the output features of the fourth transformer encoder are used as inputs of the dilated spatial convolution pooling pyramid.

4. The high-precision image binary segmentation method based on lightweight model as claimed in claim 3, characterized in that: The binary segmentation model based on the bilateral reference frame includes three bidirectional reference modules connected in sequence, and each bidirectional reference module is connected to a residual connection block; the output features of the third converter encoder are passed through a convolution layer and combined with the output features of the first bidirectional reference module, and used as the input of the first residual connection block; the output features of the second converter encoder are passed through a convolution layer and combined with the output features of the second bidirectional reference module, and used as the input of the second residual connection block; the output features of the first converter encoder are passed through a convolution layer and combined with the output features of the third bidirectional reference module, and used as the input of the third residual connection block.

5. The high-precision image binary segmentation method based on lightweight model according to claim 4, characterized in that: The bidirectional reference module includes an internal reference module, a reconstruction module and an external reference module connected in sequence; after the internal reference module processes the input features of the bidirectional reference module, the input features are input into the reconstruction module together with the input features of the bidirectional reference module; after the output features of the reconstruction module are processed by the external reference module and the activation function, the output features of the reconstruction module are multiplied at the pixel level with the output features of the reconstruction module to obtain the output features of the bidirectional reference module.

6. A high-precision image binary segmentation method based on lightweight model as claimed in claim 4, characterized in that: After the output features of the third residual connection block pass through the predictor, the binary image is obtained.

7. The high-precision image binary segmentation method based on lightweight model according to claim 1, characterized in that: The internal reference module of the student network adopts the attention internal reference, which crops the original image into image blocks with the same size as the input features, stacks the image blocks with the input features, and processes them through the convolutional attention module to obtain the output features.

8. A high-precision image binary segmentation system based on lightweight model, characterized in that: include: The training module is configured as follows: the original image in the training set is used as the input of the teacher network and the student network at the same time, and the knowledge distillation strategy is adopted to enable the teacher network to guide the student network to train by supervising the feature map obtained by the residual connection block; the teacher network and the student network both adopt a binary segmentation model based on a bilateral reference frame, the teacher network uses Swin-Large as the skeleton network, and the student network uses Swin-Tiny as the skeleton network; An image acquisition module is configured to: acquire an image to be segmented; The image segmentation module is configured to: segment the image to be segmented through the student network to obtain a binary image.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in a high-precision image binary segmentation method based on model lightweighting as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a computer-readable storage medium, a processor, and a computer program stored in the computer-readable storage medium and executable on the processor, characterized in that: When the processor executes the program, the steps in the high-precision image binary segmentation method based on model lightweight as described in any one of claims 1 to 7 are implemented.