Weak supervision image segmentation method and system based on superpixels and confidence learning
By generating high-quality pseudo-labels through multi-scale superpixel segmentation and confidence learning, and combining a teacher-student consensus model and a local Transformer module, the problems of high-cost annotation and pseudo-label noise in medical image segmentation are solved, thereby improving the segmentation accuracy and stability of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-10
- Publication Date
- 2026-04-10
AI Technical Summary
Existing deep learning-based medical image segmentation methods rely on large-scale pixel-level annotations, which are costly and difficult to generalize. In weakly supervised methods, pseudo-label noise and sparse annotations lead to insufficient model generalization ability.
A multi-scale superpixel segmentation algorithm is used to generate structure-aware pseudo-labels, which are then trained using a confidence learning and teacher-student consistency model. The pseudo-labels are optimized using a graph attention network, and consistency loss and a local Transformer module are introduced to enhance the segmentation capability.
Generate high-quality pseudo-labels to improve the segmentation accuracy and stability of the model under low-label conditions, enhance the recognition ability of complex image scenes, and improve the overall accuracy of image segmentation.
Smart Images
Figure CN121837619A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning image segmentation, and in particular to a weakly supervised image segmentation method and system based on superpixel and confidence learning. Background Technology
[0002] With the development of deep learning, image segmentation methods based on convolutional neural networks or Transformers have made significant progress in multiple tasks. However, these methods typically rely on large-scale pixel-level labeled data, while pixel-level labeling of medical images, especially optical coherence tomography (OCT) images, requires professional physicians to draw each pixel, which is labor-intensive and costly, making it difficult to widely promote in real-world clinical settings.
[0003] Existing deep learning-based medical image segmentation methods mostly rely on large-scale pixel-level annotations. This annotation work is time-consuming, labor-intensive, requires ophthalmologists, and is extremely costly, severely limiting its large-scale application. Weakly supervised and semi-supervised methods can be trained using a small amount of labeled data combined with unlabeled data, but when relying solely on point annotations, problems arise such as sparse point annotations lacking structural information, making it difficult to generate high-quality pseudo-labels; pseudo-labels contain noise, easily propagating errors during training; and the model's generalization ability is insufficient under low-label conditions. Therefore, there is an urgent need for a medical image segmentation method that can generate high-quality pseudo-labels under limited annotation conditions and effectively utilize both labeled and unlabeled data. Summary of the Invention
[0004] In view of the aforementioned existing problems, the present invention is proposed.
[0005] Therefore, this invention provides a weakly supervised image segmentation method based on superpixel and confidence learning to solve the problems of difficulty in generating high-quality pseudo-labels, the presence of noise in pseudo-labels, and insufficient generalization ability of the model under low-label conditions.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a weakly supervised image segmentation method based on superpixel and confidence learning, which includes using a multi-scale superpixel segmentation algorithm to segment an input image into multiple superpixel blocks; For the superpixel block, point annotation is performed, and based on the point annotation and the superpixel segmentation result, combined with the orientation consistency constraint, a structure-aware pseudo-label and a confidence map are generated. Using the confidence map, confidence learning is employed to identify and correct noisy regions in the structure-aware pseudo-labels. The corrected pseudo-labels were trained using a teacher-student consistency model, and the trained pseudo-labels were then used to perform accurate image segmentation.
[0007] As a preferred embodiment of the weakly supervised image segmentation method based on superpixel and confidence learning described in this invention, the multi-scale superpixel segmentation method includes: using a pyramid pooling module to perform multi-scale pooling on the image to obtain image features at different scales, and inputting the image features for normalization processing to remove noise; based on the SLIC algorithm, setting an appropriate superpixel block size, and performing fine-grained segmentation by adaptively selecting key regions. The superpixel block comprises a set of pixels in the image, ensuring that each block has similar color and texture. The superpixel segmentation result is represented as follows: in, This represents the weights calculated during superpixel segmentation; Superpixels The average color; Superpixels The average color; Represents the gradient of a superpixel boundary; Indicates the degree of inconsistency in direction; , , This represents the weight coefficient of the adjustment term; and the superpixel is labeled and optimized based on the weight.
[0008] As a preferred embodiment of the weakly supervised image segmentation method based on superpixel and confidence learning described in this invention, the point annotation includes marking the precise location of key points in the image and providing a supervision signal, represented as: in, Indicates at point The final pseudo-tag output after propagation; Indicates the structure based on the graph From the marked point The propagation function for the label propagation process to other points; After obtaining the set of label propagation paths, the category labels are propagated from the annotation points along the connecting paths to each pixel on the path, as shown below: in, Superpixels The weights calculated at each location; Represents pixels Its nearest labeled point Euclidean distance; This indicates the color similarity between the superpixel containing the current pixel and the superpixel of the labeled point; Indicates the current student model's position. The prediction confidence level at the location; The hyperparameters representing the distance decay rate are used to obtain the structure-aware pseudo-labels and their credibility maps.
[0009] As a preferred embodiment of the weakly supervised image segmentation method based on superpixel and confidence learning described in this invention, the confidence learning includes: using a graph attention network to construct a label trust graph based on pseudo-labels and a confidence graph; establishing a graph structure with superpixels as nodes and structural similarity between adjacent regions as edges; optimizing the label trust graph through the graph attention network to generate superpixel-level correction results; the optimized node prediction results are mapped back to the pixel space to form refined pseudo-labels that replace the original pseudo-labels; entering the distillation training stage; introducing fully labeled FA images to jointly optimize the teacher-student model; and iteratively obtaining the final segmentation model.
[0010] As a preferred embodiment of the weakly supervised image segmentation method based on superpixel and confidence learning described in this invention, the teacher-student consistency model includes a student model trained using the fully labeled and optimized pseudo-labels, and a teacher model that updates parameters through exponential moving average and provides stable supervision signals. The teacher-student consistency model is expressed as follows: in, express Raw time data; express Data that is constantly updated; express Data that is constantly updated; This represents the decay coefficient of the exponential moving average; This indicates a loss of consistency. Indicates the confidence level of the pseudo-label; Indicates the confidence threshold; Indicates the input image Size; Indicates the Kullback–Leibler divergence; Indicates in the source domain The predicted distribution; Indicates in the target domain R The predicted distribution; During training, a consistency loss function and a PA image are introduced, allowing the student and teacher models to predict the input image respectively, resulting in a prediction distribution. The FA image supervises the student model with full labels and is optimized using standard supervised loss. In each round of training, the teacher model continuously absorbs the latest knowledge from the student model, making the pseudo-labels of the PA image tend to stabilize.
[0011] As a preferred embodiment of the weakly supervised image segmentation method based on superpixel and confidence learning described in this invention, the following is provided: a two-layer feature fusion module is used to fuse the features in the decoder with the features corresponding to the encoder, a fixed-size window is divided on the fused features, and a standard multi-head self-attention mechanism is executed within the window; The multi-head self-attention mechanism includes introducing a local Transformer module and inputting the fused features to capture long-distance dependencies and enhance the ability to model the context of local regions.
[0012] As a preferred embodiment of the weakly supervised image segmentation method based on superpixel and confidence learning described in this invention, the two-layer feature fusion module is represented as follows: in, Indicates the first ; This represents a 1x1 convolution operation; Indicates the first Upsampled feature map of the layer; Indicates the first The feature map after using the two-layer feature fusion module; Indicates a splicing operation; Indicates the first ; Indicates the first Encoder feature map of the layer.
[0013] Secondly, this invention provides a weakly supervised image segmentation system based on superpixel and confidence learning, comprising: The image segmentation unit uses a multi-scale superpixel segmentation algorithm to segment the input image into multiple superpixel blocks; The pseudo-label generation unit performs point annotation on the superpixel block, and generates structure-aware pseudo-labels and confidence maps based on the point annotations and superpixel segmentation results, combined with directional consistency constraints. The correction unit uses confidence learning to identify and correct noise regions in the pseudo-labels based on the confidence map. The training unit uses a teacher-student consistency model to train the corrected pseudo-labels, and uses a decoder to perform accurate image segmentation on the trained pseudo-labels.
[0014] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the weakly supervised image segmentation method based on superpixel and confidence learning as described in the first aspect of the present invention.
[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the weakly supervised image segmentation method based on superpixel and confidence learning as described in the first aspect of the present invention.
[0016] The beneficial effects of this invention are as follows: By using superpixel segmentation and orientation consistency constraints, the generated pseudo-labels can fully consider the structural information and local features of the image. This structure-aware pseudo-label generation method avoids the shortcomings of traditional methods that rely solely on global information, and more accurately reflects the details and boundaries of the image. Through confidence learning, the optimization of pseudo-labels not only considers pixel-level label information, but also dynamically corrects them based on the similarity between superpixels and local contextual information. Pseudo-labels in low-confidence regions are suppressed during the optimization process, avoiding interference from noisy labels on the training process and ensuring the propagation of high-quality labels. Through collaborative training of teacher-student models, combined with high-quality pseudo-labels and fully labeled images, the model can gradually improve its segmentation ability. The introduction of a consistency loss function ensures that the predictions of the teacher and student models remain consistent in high-confidence regions, thereby further improving the model's stability and generalization ability to unknown data. By introducing a local Transformer module, the model's understanding of the local context of the image is enhanced, especially for the segmentation of detailed regions, ensuring accurate identification of target objects even in complex image scenes. By using a two-layer feature fusion module (CFFB), the model can better integrate multi-scale information from the decoder and encoder, thereby improving the overall accuracy of image segmentation, especially in the processing of complex backgrounds and detailed regions. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a flowchart of a weakly supervised image segmentation method based on superpixels and confidence learning.
[0019] Figure 2This diagram illustrates the point-label-guided directional superpixel pseudo-label generation process for a weakly supervised image segmentation method based on superpixel and confidence learning.
[0020] Figure 3 This is a flowchart of pseudo-label optimization for a label trust graph and uncertainty-guided graph attention network in a weakly supervised image segmentation method based on superpixel and confidence learning.
[0021] Figure 4 The diagram shows the network structure of a weakly supervised image segmentation method based on superpixel and confidence learning, which incorporates CFFB and local Transformer modules.
[0022] Figure 5 A framework diagram for generating directional pseudo-labels for multi-point annotation in weakly supervised image segmentation based on superpixel and confidence learning is presented. Detailed Implementation
[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0026] Reference Figures 1-5 This is one embodiment of the present invention, which provides a weakly supervised image segmentation method based on superpixel and confidence learning, including the following steps: S1. Use a multi-scale superpixel segmentation algorithm to segment the input image into multiple superpixel blocks.
[0027] The multi-scale superpixel segmentation method includes: using a pyramid pooling module to perform multi-scale pooling on the image to obtain image features at different scales, and inputting the image features for normalization processing to remove noise; based on the SLIC algorithm, setting an appropriate superpixel block size, and performing fine-grained segmentation by adaptively selecting key regions.
[0028] The superpixel block comprises a set of pixels in the image, ensuring that each block has similar color and texture.
[0029] The superpixel segmentation result is represented as follows: in, This represents the weights calculated during superpixel segmentation; Superpixels The average color; Superpixels The average color; Represents the gradient of a superpixel boundary; Indicates the degree of inconsistency in direction; , , This represents the weight coefficient of the adjustment term; and the superpixel is labeled and optimized based on the weight.
[0030] It's important to know that superpixel segmentation is an image segmentation method that divides an image into a series of relatively consistent regions, making the pixels in each region similar in certain features (such as color and texture).
[0031] The multi-scale features include the ability of images at different scales to reveal different structural information during segmentation. By combining multi-scale features (such as image texture, color, and edges), superpixel segmentation can be performed more accurately; by convolving or filtering images at different scales and then segmenting them on feature maps at multiple scales, image information at different levels can be effectively captured.
[0032] The pyramid pooling module includes pooling the input feature map at different scales to extract multi-level, different-scale image features. These features at different scales can help the network better capture global and detailed information of the image, thereby improving the model's ability to recognize targets of different sizes and providing more useful information for superpixel segmentation.
[0033] S2. For the superpixel block, perform point annotation, and based on the point annotation and superpixel segmentation results, generate structure-aware pseudo-labels and confidence maps in combination with directional consistency constraints.
[0034] The point annotation includes marking the precise locations of key points in the image and providing a supervision signal, represented as follows: in, Indicates at point The final pseudo-tag output after propagation; Indicates the structure based on the graph From the marked point The propagation function for the label propagation process to other points.
[0035] After obtaining the set of label propagation paths, the category labels are propagated from the annotation points along the connecting paths to each pixel on the path, as shown below: in, Superpixels The weights calculated at each location; Represents pixels Its nearest labeled point Euclidean distance; This indicates the color similarity between the superpixel containing the current pixel and the superpixel of the labeled point; Indicates the current student model's position. The prediction confidence level at the location; The hyperparameters representing the distance decay rate are used to obtain the structure-aware pseudo-labels and their credibility maps (e.g., ...). Figure 2 (As shown).
[0036] The directional consistency constraint is a constraint method used in computer vision and image processing to ensure the consistency of different regions or pixels in an image in a specific direction. Typically, this constraint method is used to optimize tasks such as image segmentation, image restoration, and feature matching to enhance the structural or geometric consistency of image processing results. In image segmentation tasks, directional consistency constraints help maintain the consistency of edge information within image regions, thereby avoiding unnecessary segmentation errors at object boundaries. For example, applying directional consistency constraints at the edges of an image ensures that the segmented edges follow the true boundary direction of the image.
[0037] S3. Using the confidence map, confidence learning is used to identify and correct noisy regions in the pseudo-labels.
[0038] The confidence learning recognition includes using a graph attention network to construct a label trust graph based on pseudo-labels and a confidence graph. A graph structure is established with superpixels as nodes and structural similarity between adjacent regions as edges. The label trust graph is then optimized using the graph attention network to generate superpixel-level correction results. The optimized node prediction results are mapped back to the pixel space to form refined pseudo-labels that replace the original pseudo-labels. The process then enters the distillation training phase, where fully labeled FA images are introduced to jointly optimize the teacher-student model. The final segmentation model is obtained iteratively.
[0039] The pseudo-label optimization module is represented as follows: in, Represents a node Feature vectors at layer 0; Represents a node Size; This indicates that for each element The output of the neural network function; Represents a node and nodes Attention weights between them; Indicates an indicator function; This represents the ReLU activation function with leakage; Representing vectors Transpose of; Represents the weight matrix; Represents a node eigenvectors; Represents a node eigenvectors; Represents a node The set of adjacent nodes; Represents a node exist The feature vector of the layer; Indicates the activation function; Attention weights reflect the nodes For nodes Contribution weights to feature updates; Indicates learnable parameters; Represents a node exist The feature vectors of the layer (such as) Figure 3 (As shown).
[0040] The This includes features determined by the similarity and uncertainty of node features, used to weight neighborhood features during feature aggregation, thereby strengthening the influence of high-confidence regions and suppressing noise propagation.
[0041] The FA images include important diagnostic tools used to observe and assess the condition of retinal blood vessels through fluorescence angiography, combined with specific fluorescent dyes and imaging equipment.
[0042] S4. The corrected pseudo-labels are trained using a teacher-student consistency model, and the trained pseudo-labels are then used to perform accurate image segmentation using a decoder.
[0043] The teacher-student consistency model includes a student model trained using the fully labeled and optimized pseudo-labels, and a teacher model that updates parameters through exponential moving average and provides stable supervision signals. The teacher-student consistency model is expressed as follows: in, express Raw time data; express Data that is constantly updated; express Data that is constantly updated; This represents the decay coefficient of the exponential moving average; Represents the consistency loss function; Indicates the confidence level of the pseudo-label; Indicates the confidence threshold; Indicates the input image Size; Indicates the Kullback–Leibler divergence; Indicates in the source domain The predicted distribution; Indicates in the target domain R The predicted distribution (e.g.) Figure 3 (As shown).
[0044] During training, a consistency loss function and a PA image are introduced, allowing the student and teacher models to predict the input image respectively, resulting in a prediction distribution. The FA image supervises the student model with full labels and is optimized using standard supervised loss. In each round of training, the teacher model continuously absorbs the latest knowledge from the student model, making the pseudo-labels of the PA image tend to stabilize.
[0045] The PA image includes an imaging technique that utilizes the photoacoustic effect, which refers to the localized thermal expansion and generation of ultrasonic waves that occur when a material absorbs laser light or other light sources. Image information can be obtained using these ultrasonic signals.
[0046] The two-layer feature fusion module fuses the features in the decoder with the corresponding features in the encoder, divides the fused features into a fixed-size window, and executes a standard multi-head self-attention mechanism within the window.
[0047] The multi-head self-attention mechanism includes introducing a local Transformer module and inputting the fused features to capture long-distance dependencies and enhance the ability to model the context of local regions (e.g., Figure 4 (As shown).
[0048] The dual-layer feature fusion module is represented as follows: in, Indicates the first ; This represents a 1x1 convolution operation; Indicates the first Upsampled feature map of the layer; Indicates the first The feature map after using the two-layer feature fusion module; Indicates a splicing operation; Indicates the first ; Indicates the first The encoder feature map of the layer (e.g.) Figure 5 (As shown).
[0049] Example 2, an embodiment of the present invention, provides a contact erosion assessment system for SF6 circuit breaker breaking tests, comprising: The image segmentation unit uses a multi-scale superpixel segmentation algorithm to segment the input image into multiple superpixel blocks; The pseudo-label generation unit performs point annotation on the superpixel block, and generates structure-aware pseudo-labels and confidence maps based on the point annotations and superpixel segmentation results, combined with directional consistency constraints. The correction unit uses confidence learning to identify and correct noise regions in the pseudo-labels based on the confidence map. The training unit uses a teacher-student consistency model to train the corrected pseudo-labels, and uses a decoder to perform accurate image segmentation on the trained pseudo-labels.
[0050] This embodiment also provides a computer device applicable to the weakly supervised image segmentation method based on superpixel and confidence learning, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the weakly supervised image segmentation method based on superpixel and confidence learning as proposed in the above embodiment.
[0051] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0052] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the weakly supervised image segmentation method based on superpixel and confidence learning as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0053] In summary, this invention effectively addresses the problems of sparse point annotations lacking structural information and making it difficult to generate high-quality pseudo-labels; the presence of noise in pseudo-labels, which easily propagates errors during training; and the insufficient generalization ability of the model under low-annotation conditions by adopting a point annotation-guided image segmentation method based on superpixels and confidence learning, and by introducing a teacher-student collaborative training architecture, a two-layer feature fusion module, and a local Transformer module. This significantly improves the performance of weakly supervised image segmentation methods.
[0054] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A weakly supervised image segmentation method based on superpixel and confidence learning, characterized in that: include, The input image is divided into multiple superpixel blocks using a multi-scale superpixel segmentation algorithm. For the superpixel block, point annotation is performed, and based on the point annotation and the superpixel segmentation result, combined with the orientation consistency constraint, a structure-aware pseudo-label and a confidence map are generated. Using the confidence map, confidence learning is employed to identify and correct noisy regions in the structure-aware pseudo-labels. The corrected pseudo-labels were trained using a teacher-student consistency model, and the trained pseudo-labels were then used to perform accurate image segmentation.
2. The weakly supervised image segmentation method based on superpixel and confidence learning as described in claim 1, characterized in that: The multi-scale superpixel segmentation method includes using a pyramid pooling module to perform multi-scale pooling on the image to obtain image features at different scales, and inputting the image features for normalization processing to remove noise. Based on the SLIC algorithm, an appropriate superpixel block size is set, and fine-grained segmentation is performed by adaptively selecting key regions; The superpixel block comprises a set of pixels in the image, ensuring that each block has similar color and texture. The superpixel segmentation result is represented as follows: in, This represents the weights calculated during superpixel segmentation; Superpixels The average color; Superpixels The average color; Represents the gradient of a superpixel boundary; Indicates the degree of inconsistency in direction; , , This represents the weight coefficient of the adjustment term; and the superpixel is labeled and optimized based on the weight.
3. The weakly supervised image segmentation method based on superpixel and confidence learning as described in claim 2, characterized in that: The point annotation includes marking the precise locations of key points in the image and providing a supervision signal, represented as follows: in, Indicates at point The final pseudo-tag output after propagation; Indicates the structure based on the graph From the marked point The propagation function for the label propagation process to other points; After obtaining the set of label propagation paths, the category labels are propagated from the annotation points along the connecting paths to each pixel on the path, as shown below: in, Superpixels The weights calculated at each location; Represents pixels Its nearest labeled point Euclidean distance; This indicates the color similarity between the superpixel containing the current pixel and the superpixel of the labeled point; Indicates the current student model's position. The prediction confidence level at the location; The hyperparameters representing the distance decay rate are used to obtain the structure-aware pseudo-labels and their credibility maps.
4. The weakly supervised image segmentation method based on superpixel and confidence learning as described in claim 3, characterized in that: The confidence learning process includes using a graph attention network to construct a label confidence graph based on pseudo-labels and a confidence graph. A graph structure is established with superpixels as nodes and structural similarity between adjacent regions as edges. The label confidence graph is then optimized using the graph attention network to generate superpixel-level correction results. The optimized node prediction results are mapped back to the pixel space to form refined pseudo-labels that replace the original pseudo-labels. The process then enters the distillation training phase, where fully labeled FA images are introduced to jointly optimize the teacher-student model. The final segmentation model is obtained iteratively.
5. The weakly supervised image segmentation method based on superpixel and confidence learning as described in claim 4, characterized in that: The teacher-student consistency model includes a student model trained using the fully labeled and optimized pseudo-labels, and a teacher model that updates parameters through exponential moving average and provides stable supervision signals. The teacher-student consistency model is expressed as follows: in, express Raw time data; express Data that is constantly updated; express Data that is constantly updated; This represents the decay coefficient of the exponential moving average; This indicates a loss of consistency. Indicates the confidence level of the pseudo-label; Indicates the confidence threshold; Indicates the input image Size; Indicates the Kullback–Leibler divergence; Indicates in the source domain The predicted distribution; Indicates in the target domain R The predicted distribution; During training, a consistency loss function and a PA image are introduced, allowing the student and teacher models to predict the input image respectively, resulting in a prediction distribution. The FA image supervises the student model with full labels and is optimized using standard supervised loss. In each round of training, the teacher model continuously absorbs the latest knowledge from the student model, making the pseudo-labels of the PA image tend to stabilize.
6. The weakly supervised image segmentation method based on superpixel and confidence learning as described in claim 5, characterized in that: The two-layer feature fusion module fuses the features in the decoder with the features corresponding to the encoder, divides the fused features into a fixed-size window, and executes a standard multi-head self-attention mechanism within the window. The multi-head self-attention mechanism includes introducing a local Transformer module and inputting the fused features to capture long-distance dependencies and enhance the ability to model the context of local regions.
7. The weakly supervised image segmentation method based on superpixel and confidence learning as described in claim 6, characterized in that: The two-layer feature fusion module is represented as follows: in, Indicates the first ; This represents a 1x1 convolution operation; Indicates the first Upsampled feature map of the layer; Indicates the first The feature map after using the two-layer feature fusion module; Indicates a splicing operation; Indicates the first ; Indicates the first Encoder feature map of the layer.
8. A weakly supervised image segmentation system based on superpixel and confidence learning, based on the method of any one of claims 1 to 7, characterized in that: include, The image segmentation unit uses a multi-scale superpixel segmentation algorithm to segment the input image into multiple superpixel blocks; The pseudo-label generation unit performs point annotation on the superpixel block, and generates structure-aware pseudo-labels and confidence maps based on the point annotations and superpixel segmentation results, combined with directional consistency constraints. The correction unit uses confidence learning to identify and correct noise regions in the pseudo-labels based on the confidence map. The training unit uses a teacher-student consistency model to train the corrected pseudo-labels, and uses a decoder to perform accurate image segmentation on the trained pseudo-labels.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the weakly supervised image segmentation method based on superpixel and confidence learning as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the weakly supervised image segmentation method based on superpixel and confidence learning as described in any one of claims 1 to 7.