A natural image segmentation method and system based on double evidence modeling

Abstract

The application discloses a natural image segmentation method and system based on double-evidence modeling, and relates to the technical field of natural image processing. The method comprises the following steps: acquiring a natural image without a label; constructing a double-evidence segmentation network model based on a Mean Teacher, including a parallel Fisher-EDL branch and an R-EDL branch, and both branches comprising a student model and a teacher model; inputting an image into each model to obtain a prediction and uncertainty, and generating a pseudo label according to the teacher prediction; obtaining a consistency mask and a prediction difference according to the double-teacher prediction, constructing a region risk map in combination with the uncertainty, mapping a cross-branch consistency loss weight, constructing a cross-branch consistency loss function to train the double-evidence segmentation network model; and inputting an image to be segmented into the trained model to obtain a segmentation result. The method can inhibit the propagation of false pseudo labels and improve the segmentation accuracy.

Description

Technical Field

[0001] This application relates to the field of natural image processing technology, and in particular to a natural image segmentation method and system based on dual-evidence modeling. Background Technology

[0002] Natural image segmentation is a fundamental task in computer vision, aiming to divide target objects, background regions, and fine-grained structures in scene images according to pixel-level semantics or instance boundaries. This task is widely used in natural image analysis scenarios such as scene understanding, object recognition, image editing, robot perception, and autonomous driving environment understanding.

[0003] Existing Mean Teacher methods typically involve a teacher model generating pseudo-labels for unlabeled natural images using a softmax probability function, and then constraining the student model to maintain consistency with the teacher model's predictions. However, natural images often exhibit issues such as blurred object boundaries, similar textures, occlusion, varying lighting, and significant differences in category appearance. When the Mean Teacher method's teacher model generates pseudo-labels for unlabeled natural images, the softmax probability function, which only reflects the relative strength of predicted categories, struggles to identify high-confidence but erroneous pseudo-labels in these complex situations. This results in unreliable pseudo-labels generated by the Mean Teacher model in these regions. These erroneous pseudo-labels directly participate in model training, allowing the student model in the Mean Teacher to continuously learn and reinforce incorrect predictions. Consequently, the trained model suffers from segmentation boundary shifts and category confusion during image segmentation, ultimately leading to insufficient accuracy in natural image segmentation. Summary of the Invention

[0004] Therefore, it is necessary to provide a natural image segmentation method and system based on dual-evidence modeling to address the aforementioned technical problems.

[0005] The present invention adopts the following technical solution: This invention provides a natural image segmentation method based on dual-evidence modeling, comprising: Acquire unlabeled natural image data; A dual-evidence segmentation network model based on the Mean Teacher model is constructed. The dual-evidence segmentation network model includes: a parallel Fisher-EDL branch and an R-EDL branch. Both the Fisher-EDL branch and the R-EDL branch include a parallel student model branch and a teacher model branch. Natural image data is input into the two-branch teacher model and student model respectively, and the prediction results and pixel-level uncertainties of each pixel in the unlabeled natural image are output; and pseudo-labels of each pixel in the unlabeled natural image are generated based on the prediction results of the two-branch teacher model. Based on the prediction results output by the two-branch teacher model, a prediction consistency mask and prediction divergence between the two-branch teacher models are determined; based on the prediction divergence and the pixel-level uncertainty output by the two-branch teacher model, a regional risk map is determined to characterize the reliability of the pseudo-label, and the regional risk map is mapped to cross-branch consistency loss weights using the prediction consistency mask; based on the prediction results output by the two-branch teacher model and the student model and the cross-branch consistency weights, a cross-branch consistency loss function is constructed to train the dual-evidence segmentation network model. The natural image to be segmented is input into a trained dual-evidence segmentation network model, which outputs the image segmentation result of the natural image to be segmented.

[0006] Preferably, both the student and teacher models in the Fisher-EDL and R-EDL branches are constructed based on evidence-based deep learning models; The Fisher-EDL branch is obtained by introducing the Fisher information matrix into the evidence deep learning model; the R-EDL branch is obtained by introducing adjustable relaxed prior parameters into the fixed prior Dirichlet distribution parameters of the evidence deep learning model.

[0007] Preferably, based on the prediction results output by the two-branch teacher models, the prediction consistency mask and prediction divergence between the two-branch teacher models are determined, specifically including: Based on the prediction results of the Fisher-EDL branch teacher model and the R-EDL branch teacher model for each pixel region of the unlabeled natural image, the prediction category of each pixel region of the natural image is determined. Based on whether the prediction categories of each pixel region are consistent between the two branch teacher models, a prediction consistency mask is generated between the two branch teacher models. The prediction divergence is determined based on the difference in prediction probabilities between the prediction results of the two-branch teacher models for each pixel region.

[0008] Preferably, based on the predicted divergence and the pixel-level uncertainty of the two-branch teacher model output, a regional risk map is determined to characterize the reliability of the pseudo-label, specifically including: The pixel-level uncertainty output by the Fisher-EDL branch teacher model and the R-EDL branch teacher model is transformed to a unified scale to obtain the unified scale uncertainty of natural image data in the two branches. Based on the uniform scale uncertainty and prediction discrepancy corresponding to the R-EDL branch teacher model, a regional risk map of the Fisher-EDL branch teacher model is determined; the regional risk map of the Fisher-EDL branch teacher model is used to characterize the reliability of the Fisher-EDL branch student model in learning the pseudo-labels corresponding to the R-EDL branch teacher model. Based on the uniform scale uncertainty and prediction discrepancy corresponding to the Fisher-EDL branch teacher model, a regional risk map of the R-EDL branch teacher model is determined; the regional risk map of the R-EDL branch teacher model is used to characterize the reliability of the R-EDL branch student model in learning the pseudo-labels corresponding to the Fisher-EDL branch teacher model.

[0009] Preferably, a cross-branch consistency loss function is constructed based on the prediction results output by the two-branch teacher model and the student model, and the cross-branch consistency weights. Specifically, this includes: Based on the cross-branch consistency weights corresponding to the Fisher-EDL branch, the prediction results output by the student model in the Fisher-EDL branch, and the prediction results output by the teacher model in the R-EDL branch, the cross-branch consistency loss function corresponding to the Fisher-EDL branch is constructed, and its expression is: ; In the formula, The cross-branch consistency loss corresponding to the Fisher-EDL branch. This represents the prediction results of the Fisher-EDL branch student model on natural image data. This represents the prediction results of the R-EDL branch teacher model for natural image data. The cross-branch consistency weights corresponding to the Fisher-EDL branches. To prevent extremely small positive numbers with a denominator of zero; Based on the cross-branch consistency weights corresponding to the R-EDL branches, the prediction results output by the student model in the R-EDL branch, and the prediction results output by the teacher model in the Fisher-EDL branch, the cross-branch consistency loss function corresponding to the R-EDL branches is constructed, and its expression is: ; In the formula, This represents the cross-branch consistency loss corresponding to the R-EDL branch. The cross-branch consistency weights corresponding to the R-EDL branches. This represents the prediction results of the R-EDL branch student model on natural image data. This represents the prediction results of the Fisher-EDL branch teacher model for natural image data output.

[0010] Preferably, when training the dual-evidence segmentation network model, the method further includes: Acquire labeled natural image data; Labeled natural image data is input into a two-branch student model, and the class evidence corresponding to the two branches of the student model is output respectively. Based on the category evidence, construct the Dirichlet distribution parameters for the two branches; and based on the Dirichlet distribution parameters, determine the prediction results of the student model for the labeled natural image for the two branches. Based on the prediction results of the two-branch student model on labeled natural image data and their corresponding real labels, Dice loss function and evidence loss are constructed respectively to conduct supervised training of the two-branch student model.

[0011] Preferably, the evidence loss construction process of the Fisher-EDL branch student model specifically includes: Construct the Fisher information matrix based on the Dirichlet distribution parameters of the Fisher-EDL branch student model; Based on the difference between the prediction results of the two-branch student model on labeled natural image data and their corresponding true labels, as well as the Fisher information matrix, the Fisher weighted error term is determined. A KL regularization term is introduced, and the evidence loss of the Fisher-EDL branch student model is constructed based on the Fisher weighted error term, the determinant of the Fisher information matrix, and the KL regularization term.

[0012] Preferably, the evidence loss construction process of the R-EDL branch student model specifically includes: An adjustable relaxation prior parameter is introduced, and the Dirichlet distribution parameters of the R-EDL branch are constructed based on the adjustable relaxation prior parameter and the class evidence output by the R-EDL branch student model. The prediction results of the R-EDL branch student model for the labeled natural image are determined based on the Dirichlet distribution parameters. Based on the difference between the prediction result and the true label of the labeled natural image data, a projection probability error term is determined; Based on the Dirichlet distribution parameters and the relaxed prior parameters, construct the relaxed form Dirichlet parameters corresponding to the true label; determine the KL regularization term of the R-EDL branch based on the KL divergence between the Dirichlet distribution parameters and the relaxed form Dirichlet parameters. Based on the projection probability error term and the KL regularization term of the R-EDL branch, construct the evidence loss of the R-EDL branch student model.

[0013] Preferably, when training the dual-evidence segmentation network model, the method further includes: For the teacher and student models within any branch, based on the prediction results and pixel-level uncertainties output by the teacher and student models, an intra-branch consistency loss is constructed, specifically including: For the teacher model and student model within any branch, the prediction difference between the teacher model and the student model is determined based on the prediction results of the student model on natural image data and the prediction results of the teacher model on natural image data. Based on the pixel-level uncertainty of the teacher model output, construct the consistency weights within this branch; Based on the prediction differences within the branch and the consistency weight within the branch, an intra-branch consistency loss is constructed.

[0014] This invention provides a natural image segmentation system based on dual-evidence modeling, comprising: The data acquisition module is used to acquire unlabeled natural image data; The model building module is used to build a dual-evidence segmentation network model based on the Mean Teacher model. The dual-evidence segmentation network model includes: parallel Fisher-EDL branches and R-EDL branches. Both the Fisher-EDL branches and the R-EDL branches include parallel student model branches and teacher model branches. The model training module is used to input natural image data into the two-branch teacher model and student model respectively, output the prediction results and pixel-level uncertainty of each pixel in the unlabeled natural image; and generate pseudo-labels for each pixel in the unlabeled natural image based on the prediction results of the two-branch teacher model; determine the prediction consistency mask and prediction divergence between the two-branch teacher model according to the prediction results output by the two-branch teacher model; determine the region risk map to characterize the reliability of the pseudo-labels based on the prediction divergence and the pixel-level uncertainty output by the two-branch teacher model, and map the region risk map to cross-branch consistency loss weights using the prediction consistency mask; and construct a cross-branch consistency loss function based on the prediction results and cross-branch consistency weights output by the two-branch teacher model and student model to train the dual-evidence segmentation network model. The image segmentation module is used to acquire the natural image to be segmented and input it into the trained dual-evidence segmentation network model, and output the image segmentation result of the natural image to be segmented.

[0015] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described natural image segmentation method based on dual-evidence modeling.

[0016] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described natural image segmentation method based on dual evidence modeling.

[0017] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects: This invention employs a cross-branch consistency loss function constructed based on cross-branch consistency loss weights obtained from region risk map mapping to train a dual-evidence segmentation network model. The region risk map reflects the consistency of prediction results from the two branch teacher models for the same natural image data and the degree of pseudo-label risk in each pixel region through the prediction discrepancy and pixel-level uncertainty output by the two branch teacher models. This effectively characterizes the reliability of pseudo-labels generated by the teacher models. During the training of the dual-branch evidence network model, the cross-branch consistency loss weights obtained from the aforementioned region risk map mapping can adaptively differentiate the supervision intensity of pseudo-labels in different pixel regions of the natural image while maintaining consistency between the predictions of the student model in one branch and the teacher model in the other. Specifically, reliable pseudo-labels in low-risk pixel regions with small prediction discrepancies, low consistency, and low uncertainty are given high weights to allow the cross-branch student model to learn fully. Simultaneously, unreliable pseudo-labels in high-risk pixel regions with large prediction discrepancies or high uncertainty are given low weights to suppress the negative impact of unreliable pseudo-labels on the student model's learning. This effectively solves the problems of segmentation boundary shift and category confusion caused by erroneous pseudo-labels participating in model training, and effectively improves the segmentation accuracy of natural images.

[0018] Furthermore, this invention employs the Fisher-EDL branch to adaptively weight the prediction error based on the Fisher information matrix and integrates KL regularization terms to construct evidence loss, and the R-EDL branch to construct evidence loss based on relaxed prior parameters and KL regularization terms. This allows the model to adjust the learning process according to the information content of the evidence distribution during supervised training with labeled data. In reliable regions, the Fisher-EDL branch can form a more stable and clear discriminative evidence expression, while the R-EDL branch can avoid the excessively sharp Dirichlet distribution caused by variance minimization constraints in regions with insufficient evidence, thus maintaining a more conservative evidence distribution in uncertain regions. As a result, the two branches establish differentiated reliable evidence modeling capabilities from the early stages of training, providing pixel-level uncertainty estimation with complementary advantages for training on unlabeled data. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0020] Figure 1 A flowchart illustrating a natural image segmentation method based on dual-evidence modeling provided by the present invention; Figure 2 A schematic diagram of the dual-evidence segmentation network architecture for a natural image segmentation method based on dual-evidence modeling provided by the present invention; Figure 3 A schematic diagram of a natural image segmentation system based on dual-evidence modeling provided by the present invention; Figure 4 A schematic diagram of a computer device for implementing a natural image segmentation method based on dual-evidence modeling, provided by the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in the specification without creative effort are within the scope of protection of this application.

[0022] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0023] Figure 1 This is a schematic diagram of a natural image segmentation method based on dual-evidence modeling in this invention, which specifically includes the following steps: S101: Obtain unlabeled natural image data.

[0024] S102: Construct a dual-evidence segmentation network model based on the Mean Teacher model; the dual-evidence segmentation network model includes: parallel Fisher-EDL branches and R-EDL branches; both the Fisher-EDL branches and the R-EDL branches include parallel student model branches and teacher model branches.

[0025] Optionally, the student and teacher models of both the Fisher-EDL and R-EDL branches are constructed based on the evidence deep learning model; wherein, the Fisher-EDL branch is obtained by introducing the Fisher information matrix into the evidence deep learning model; and the R-EDL branch is obtained by introducing adjustable relaxed prior parameters into the fixed prior Dirichlet distribution parameters of the evidence deep learning model.

[0026] S103: Input the natural image data into the two-branch teacher model and student model respectively, output the prediction results and pixel-level uncertainty of each pixel in the unlabeled natural image; and generate pseudo-labels for each pixel in the unlabeled natural image based on the prediction results of the two-branch teacher model.

[0027] S104: Based on the prediction results output by the two-branch teacher model, determine the prediction consistency mask and prediction divergence between the two-branch teacher models; based on the prediction divergence and the pixel-level uncertainty output by the two-branch teacher model, determine the regional risk map used to characterize the reliability of the pseudo-label, and use the prediction consistency mask to map the regional risk map into cross-branch consistency loss weights; based on the prediction results output by the two-branch teacher model and the student model and the cross-branch consistency weights, construct a cross-branch consistency loss function to train the dual-evidence segmentation network model.

[0028] Optionally, based on the prediction results output by the two-branch teacher models, a prediction consistency mask and prediction divergence between the two-branch teacher models are determined. Specifically, this includes: determining the prediction category of each pixel region of the natural image based on the prediction results output by the Fisher-EDL branch teacher model and the R-EDL branch teacher model for each pixel region of the unlabeled natural image; generating a prediction consistency mask between the two-branch teacher models based on whether the prediction categories of each pixel region are consistent; and determining the prediction divergence based on the difference in prediction probabilities between the prediction results output by the two-branch teacher models for each pixel region.

[0029] Optionally, based on the prediction divergence and the pixel-level uncertainty output by the two-branch teacher models, a regional risk map is determined to characterize the reliability of the pseudo-labels. Specifically, this includes: transforming the pixel-level uncertainties output by the Fisher-EDL branch teacher model and the R-EDL branch teacher model to a unified scale to obtain the unified scale uncertainty of the natural image data for the two branches; determining the regional risk map of the Fisher-EDL branch teacher model based on the unified scale uncertainty corresponding to the R-EDL branch teacher model and the prediction divergence; the regional risk map of the Fisher-EDL branch teacher model is used to characterize the reliability of the Fisher-EDL branch student model when learning the pseudo-labels corresponding to the R-EDL branch teacher model; and determining the regional risk map of the R-EDL branch teacher model based on the unified scale uncertainty corresponding to the Fisher-EDL branch teacher model and the prediction divergence; the regional risk map of the R-EDL branch teacher model is used to characterize the reliability of the R-EDL branch student model when learning the pseudo-labels corresponding to the Fisher-EDL branch teacher model.

[0030] Optionally, a cross-branch consistency loss function is constructed based on the prediction results output by the two-branch teacher model and the student model, and the cross-branch consistency weights. Specifically, this includes: Based on the cross-branch consistency weights corresponding to the Fisher-EDL branch, the prediction results output by the student model in the Fisher-EDL branch, and the prediction results output by the teacher model in the R-EDL branch, the cross-branch consistency loss function corresponding to the Fisher-EDL branch is constructed, and its expression is: ; In the formula, The cross-branch consistency loss corresponding to the Fisher-EDL branch. This represents the prediction results of the Fisher-EDL branch student model on natural image data. This represents the prediction results of the R-EDL branch teacher model for natural image data. The cross-branch consistency weights corresponding to the Fisher-EDL branches. for,; Based on the cross-branch consistency weights corresponding to the R-EDL branches, the prediction results output by the student model in the R-EDL branch, and the prediction results output by the teacher model in the Fisher-EDL branch, the cross-branch consistency loss function corresponding to the R-EDL branches is constructed, and its expression is: ; In the formula, This represents the cross-branch consistency loss corresponding to the R-EDL branch. The cross-branch consistency weights corresponding to the R-EDL branches. This represents the prediction results of the R-EDL branch student model on natural image data. This represents the prediction results of the Fisher-EDL branch teacher model for natural image data output.

[0031] Additionally, training the dual-evidence segmentation network model also includes: Acquire labeled natural image data; input the labeled natural image data into a two-branch student model, and output the category evidence corresponding to the two branches of the student model respectively; construct the Dirichlet distribution parameters of the two branches based on the category evidence; determine the prediction results of the two branches of the student model for the labeled natural image based on the Dirichlet distribution parameters; construct the Dice loss function and evidence loss respectively based on the prediction results of the two branches of the student model for the labeled natural image data and their corresponding real labels, so as to conduct supervised training of the two branches of the student model.

[0032] The evidence loss construction process of the Fisher-EDL branch student model specifically includes: constructing the Fisher information matrix based on the Dirichlet distribution parameters of the Fisher-EDL branch student model; determining the Fisher weighted error term based on the difference between the prediction results of the two branch student models on labeled natural image data and their corresponding true labels, as well as the Fisher information matrix; introducing the KL regularization term, and constructing the evidence loss of the Fisher-EDL branch student model based on the Fisher weighted error term, the determinant of the Fisher information matrix, and the KL regularization term.

[0033] The evidence loss construction process of the R-EDL branch student model specifically includes: introducing adjustable relaxation prior parameters; constructing Dirichlet distribution parameters of the R-EDL branch based on the adjustable relaxation prior parameters and the category evidence output by the R-EDL branch student model; determining the prediction result of the R-EDL branch student model for the labeled natural image based on the Dirichlet distribution parameters; determining the projection probability error term based on the difference between the prediction result and the true label of the labeled natural image data; constructing relaxed form Dirichlet parameters corresponding to the true label based on the Dirichlet distribution parameters and the relaxation prior parameters; determining the KL regularization term of the R-EDL branch based on the KL divergence between the Dirichlet distribution parameters and the relaxed form Dirichlet parameters; and constructing the evidence loss of the R-EDL branch student model based on the projection probability error term and the KL regularization term of the R-EDL branch.

[0034] In addition, when training the dual-evidence segmentation network model, the following steps are also included: for the teacher model and student model within any branch, based on the prediction results and pixel-level uncertainty output by the teacher model and student model, constructing an intra-branch consistency loss, specifically including: for the teacher model and student model within any branch, determining the prediction difference between the teacher model and student model based on the prediction results of the student model on natural image data and the prediction results of the teacher model on natural image data; constructing the intra-branch consistency weight based on the pixel-level uncertainty output by the teacher model; and constructing the intra-branch consistency loss based on the prediction difference within the branch and the intra-branch consistency weight.

[0035] The dual-evidence segmentation network model is trained using intra-branch consistency loss, cross-branch consistency loss, Dice loss, and evidence loss to obtain a well-trained dual-evidence segmentation network model.

[0036] S105: Input the natural image to be segmented into the trained dual-evidence segmentation network model, and output the image segmentation result of the natural image to be segmented.

[0037] The following embodiment provides the complete technical solution of the present invention to further verify the effectiveness of the present invention.

[0038] S101, Obtain the natural image segmentation model. Specifically, obtain a natural image segmentation dataset, construct a dual-evidence segmentation network including Fisher-EDL and R-EDL branches, perform supervised training on labeled natural images, perform intra-module consistency and cross-branch consistency training on unlabeled natural images, and adjust the pseudo-label supervision weights according to the region risk map to obtain the natural image segmentation model.

[0039] In one embodiment, the natural image segmentation dataset can be derived from publicly available natural image segmentation datasets, such as BSDS500, PASCAL VOC, COCO-Stuff, or Cityscapes, or from collected and labeled natural scene image data. The natural image segmentation dataset includes a small number of labeled natural images and a large number of unlabeled natural images. The labeled natural images are used for supervised loss calculation, while the unlabeled natural images are used for consistency learning and pseudo-label risk control. It should be noted that the aforementioned publicly available datasets are only used to illustrate optional implementations of the method of the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0040] S102, the natural image to be segmented is segmented using a natural image segmentation model to obtain the natural image segmentation result and pixel-level reliability map.

[0041] Figure 2 This is a schematic diagram of the overall structure of the dual-evidence modeling algorithm in this invention. The framework is based on the Mean Teacher semi-supervised learning paradigm, setting up two complementary evidence branches: the Fisher-EDL branch and the R-EDL branch. It utilizes the uncertainty and prediction discrepancies of the dual branches to achieve regional risk control.

[0042] The following is combined with Figure 2 This paper explains the evidence learning, dual-branch architecture, reliable region discrimination enhancement, conservative constraints on uncertain regions, and regional risk map construction process in this invention.

[0043] (a) Foundations of Evidence Learning Traditional natural image segmentation models typically use a softmax function to convert the network output into predicted probabilities for each category, selecting the category with the highest probability as the final segmentation result. However, the softmax probability mainly reflects the relative strength of responses between different categories and does not directly indicate how much valid evidence the model possesses for that prediction. Therefore, in regions with blurred boundaries, category confusion, or strong noise interference, the model may still give a high softmax probability, resulting in a "high-confidence but incorrect" prediction. This phenomenon is particularly critical in semi-supervised segmentation tasks because erroneous high-confidence pseudo-labels, once included in training, will continue to propagate and be reinforced in subsequent iterations.

[0044] Evidential Deep Learning (EDL) provides an effective approach for modeling prediction uncertainty. Unlike softmax, which directly outputs class probabilities, EDL interprets the network output as the strength of "evidence" corresponding to different classes and further describes the uncertainty of class probabilities using a Dirichlet distribution. For a given... In class segmentation tasks, the network outputs non-negative evidence for each pixel. And thus construct the Dirichlet distribution parameters:

[0045] ; in, The model indicates that the pixel belongs to the first... The strength of evidence for a class This represents the corresponding Dirichlet distribution parameter. The sum of all category parameters is denoted as:

[0046] ; Based on the Dirichlet distribution, the expected predicted probability for each category can be obtained: ; Furthermore, EDL can also estimate the uncertainty of the prediction based on the total amount of evidence. Generally, the larger the total amount of evidence, the more sufficient the model's judgment on the current pixel, and the lower the uncertainty; conversely, when the evidence in each category is weak, it indicates that the model lacks sufficient evidence, and the prediction result should have high uncertainty. Commonly used uncertainty measures can be expressed as:

[0047] ; in, The larger the value, the more uncertain the model is about the current prediction. A smaller value indicates that the model has more sufficient evidence to support it.

[0048] In natural image segmentation tasks, EDL (Easy-to-Learn) offers the advantage of not only providing pixel-level class predictions but also simultaneously outputting uncertainty estimates for each pixel. For structurally clear regions such as the interior of the target or the main body of the target, the model typically accumulates strong evidence and provides low uncertainty. However, for regions with blurred boundaries, transitional classes, or poor image quality, the model lacks sufficient evidence, resulting in increased uncertainty. Therefore, EDL provides more direct information than softmax probabilities for assessing the reliability of pseudo-labels.

[0049] Based on the above characteristics, this embodiment introduces the concept of evidence learning into a semi-supervised natural image segmentation framework to model the evidence strength and pixel-level uncertainty behind the model's predictions. Building upon this, two complementary evidence branches, Fisher-EDL and R-EDL, are further constructed to enhance the discrimination of reliable regions and impose conservative constraints on uncertain regions, respectively, providing a foundation for regional risk map construction and differential supervision of pseudo-labels.

[0050] (II) Overall Structure of Dual Evidence Branch This invention proposes a dual-EDL branch Mean Teacher semi-supervised natural image segmentation framework, the overall structure of which is as follows: Figure 2 As shown, this framework is based on the Mean Teacher semi-supervised learning paradigm, constructing two parallel teacher-student branches: the Fisher-EDL branch and the R-EDL branch. Both branches optimize segmentation results for the same semi-supervised learning objective, but characterize the reliability of the model's predictions from different evidence modeling perspectives. Specifically, the Fisher-EDL branch focuses on enhancing the discriminative evidence representation in reliable regions, enabling the model to form more stable predictions in regions with clear boundaries and distinct categories. The R-EDL branch, on the other hand, uses relatively conservative evidence constraints to suppress situations where evidence distribution is overly concentrated and prediction uncertainty is low in low-evidence regions, thereby improving the model's reliability modeling ability in ambiguous boundaries, structurally similar regions, and noisy regions.

[0051] For labeled data, images are input into two student models, and supervised training is performed using ground truth annotations. Both branches output corresponding predictions and evidence representations, which are optimized jointly by EDL Loss and Dice Loss. On one hand, Dice Loss directly constrains the distance between the segmentation result and the ground truth label, optimizing the network's segmentation accuracy; on the other hand, EDL Loss constrains the network to learn a reasonable evidence distribution, enabling the model not only to provide segmentation predictions but also to model the uncertainty of the prediction results. Therefore, labeled data is mainly used to improve the basic segmentation capabilities of both branches and to provide a stable evidence representation basis for judging the reliability of pseudo-labels in subsequent unlabeled data.

[0052] For unlabeled data, this embodiment employs a strong-weak augmentation strategy to construct a teacher-student consistency learning process. Specifically, unlabeled images are strongly augmented before being input into the student model, and weakly augmented before being input into the corresponding EMA teacher model. Within each branch, the intra-module consistency loss constrains the student model and teacher model to maintain consistent predictions on the same unlabeled sample, thereby fully utilizing the latent structural information in the unlabeled data. Simultaneously, the teacher model is updated by the student model through exponential moving average, providing relatively stable pseudo-label predictions and mitigating the impact of student model prediction fluctuations on unlabeled supervision during training.

[0053] Building upon this foundation, this embodiment further introduces inter-module consistency constraints to achieve cross-branch mutual teaching between the two EDL branches. Specifically, the Fisher-EDL branch and the R-EDL branch generate prediction results and uncertainty estimates from different evidence modeling mechanisms. Their predictions on unlabeled data possess both consistent information and complementary reliability judgments. By constraining the predictions of the student model in one branch to be consistent with those of the teacher model in the other branch, the model can utilize the complementary information between the two branches, reducing the impact of pseudo-label bias from a single teacher model on the training process, thereby improving the stability and reliability of unlabeled sample utilization.

[0054] Furthermore, this embodiment utilizes multi-source information from the outputs of the two branches for uncertainty estimation and fusion. Specifically, this includes the uncertainty of the Fisher-EDL branch, the uncertainty of the R-EDL branch, and the prediction discrepancy and evidence strength information between the two branches. Compared to methods relying solely on a single softmax confidence level or a single branch uncertainty, multi-source prediction fusion can evaluate the credibility of pseudo-labels from multiple perspectives: when the predictions of the two branches are consistent and the evidence is sufficient, the region is considered to have high reliability; when there are significant differences between the predictions of the two branches or insufficient evidence, it indicates that the region may have high uncertainty. The resulting fused uncertainty map can be further used to adjust the consistency supervision strength of unlabeled data, reducing the negative impact of unreliable pseudo-labels on model training.

[0055] Ultimately, the overall optimization objective of this embodiment consists of three parts: supervised loss on labeled data, intra-module teacher-student consistency loss on unlabeled data, and inter-module consistency loss between the two EDL branches. Through this design, the model can simultaneously learn segmentation representations of natural images and pixel-level reliability estimation under the condition that a small amount of labeled data and a large amount of unlabeled data participate in training. The final framework not only outputs segmentation results but also generates corresponding uncertainty maps, providing auxiliary basis for the credibility analysis of natural image segmentation results.

[0056] (III) Reliable Region Discrimination Enhancement Based on Fisher-EDL To enhance the model's class discrimination capability in reliable regions, this embodiment introduces a Fisher-EDL branch into the dual-evidence modeling framework, using it as a discrimination enhancement branch for reliable regions. The basic idea of ​​Fisher-EDL originates from Fisher Information-based Evidential Deep Learning. Its core is to use the Fisher information matrix to characterize the amount of information carried by different classes in the Dirichlet evidence distribution, and adaptively weight the prediction error accordingly. Compared to traditional EDL, which uses the same penalty for all class errors, Fisher-EDL can adjust the learning process based on the amount of information in the evidence distribution, making the model focus more on class information with discriminative value, thereby improving the class discrimination capability in reliable regions.

[0057] In this embodiment, the Fisher-EDL branch corresponds to a complete teacher-student network. The student network is used for normal training and updates, while the teacher network, obtained by the exponential moving average of the student network parameters, is used to provide relatively stable prediction results for unlabeled samples. For the input image... The Fisher-EDL branch first outputs the category response corresponding to each pixel. Then, class evidence is obtained through a non-negative activation function:

[0058] ; in, Indicates the number of categories. The model represents the first The strength of evidence for a class. Further construction of the Dirichlet distribution parameters:

[0059] And define the total amount of evidence: ; The class prediction probability and pixel-level uncertainty of the Fisher-EDL branch can be obtained from the Dirichlet distribution: ; in, Indicates that the current pixel belongs to the first... Predicted probability of class This indicates uncertainty in prediction. When the total amount of evidence... A larger value indicates that the model has sufficient evidence to support the current prediction, corresponding to lower uncertainty; conversely, when the evidence is insufficient, the uncertainty increases. Therefore, the Fisher-EDL branch can not only output segmentation predictions but also provide pixel-level reliability estimates.

[0060] On labeled data, the Fisher-EDL branch receives both segmentation supervision and evidence supervision. First, the Dice loss is used to constrain the prediction of region consistency between the segmentation results and the manually labeled data.

[0061] ; in, This indicates accurate labeling. This represents the predicted probability output by the Fisher-EDL branch. While using only the Dice loss can optimize the overlap of segmented regions, it cannot explicitly constrain the distribution of evidence behind the model's predictions. Therefore, this embodiment further introduces the Fisher-EDL evidence loss.

[0062] According to Dirichlet parameters The Fisher information matrix can be represented as: ; in, Here, is the trigamma function. The Fisher information matrix is ​​used to measure the impact of different categories of evidence on the predicted distribution. Based on this matrix, the Fisher weighted error term is defined as:

[0063] ; This term weights the prediction error based on the information content of different categories of evidence, enabling the model to form a more explicit category evidence representation in reliable regions. Simultaneously, to avoid overly concentrated evidence distribution in regions with insufficient evidence or category confusion, a Fisher information matrix determinant related term is introduced:

[0064] ; In addition, to suppress the accumulation of erroneous evidence in non-true categories, this embodiment retains the KL regularization terms commonly used in EDL: ; in, ; This operation preserves the evidence constraint on non-true categories without penalizing the accumulation of evidence on true categories. Combining the above, the Fisher-EDL evidence loss can be expressed as:

[0065] ; in, Used to adjust the weights of the Fisher weighted error term. Used to control the effect of Fisher's determinant terms. The KL annealing weights gradually increase during training. In the early stages of training, the model prioritizes learning basic segmentation capabilities. As training progresses, the KL regularization is gradually enhanced to further suppress misclass evidence and overconfident predictions.

[0066] Therefore, the supervised loss of the Fisher-EDL branch on labeled samples is: ; in, This represents the dynamic weights of the Fisher-EDL evidence loss. Through this loss design, the Fisher-EDL branch not only optimizes the segmentation result itself, but also further constrains the distribution of predicted evidence, enabling the model to form a more stable and explicit category representation in reliable regions.

[0067] On unlabeled data, the Fisher-EDL branch participates in teacher-student consistency learning. The teacher model first generates predicted probabilities for the unlabeled images. and uncertainty The student model then generates predicted probabilities for the same unlabeled image after strong enhancement. To reduce the impact of unreliable pseudo-labels on training, this embodiment constructs consistency weights based on the uncertainty of the teacher model:

[0068] ; Based on this, the intra-branch consistency loss of the Fisher-EDL branch is defined as: ; When the teacher model has low uncertainty about a certain region, it indicates that the false labels in that region are relatively reliable and its consistency supervision weight is large; when the uncertainty is high, the supervision effect of that region is automatically weakened, thereby reducing the risk of false label propagation.

[0069] Furthermore, the Fisher-EDL branch also participates in subsequent cross-branch consistency learning. Specifically, the predictions of the Fisher student branch... Will it be related to the prediction of the R-EDL teacher branch? The weighted form of consistency constraints can be expressed as:

[0070] ; in, The two-branch teacher prediction consistency, uncertainty, and branch prediction divergence are jointly determined, and this part will be described in more detail in subsequent chapters.

[0071] (iv) Conservative Constraints on Uncertain Regions Based on R-EDL Unlike the Fisher-EDL branch, which focuses on enhancing the class discrimination ability in reliable regions, the R-EDL branch is mainly used to conservatively constrain uncertain regions, preventing the model from producing results with overly concentrated evidence distribution and low prediction uncertainty in regions with ambiguous boundaries, class confusion, or insufficient evidence. Traditional EDL typically fixes the prior weights in the Dirichlet distribution to the number of classes and includes a variance minimization term in the optimization objective. This may cause the model to continuously accumulate evidence of the target class, making the Dirichlet distribution overly sharp, thereby weakening the model's ability to express prediction risk. R-EDL relaxes the fixed prior weights and variance minimization constraints, treating the prior weights as adjustable hyperparameters and directly optimizing the projected probability corresponding to the expected value of the Dirichlet distribution, thus obtaining a more robust uncertainty estimate.

[0072] In this embodiment, the R-EDL branch corresponds to the second complete teacher-student network. The student network is used for parameter updates, and the teacher network, obtained by the exponential moving average of the student network parameters, is used to generate relatively stable pseudo-labels on unlabeled samples. For the input image... The R-EDL branch first outputs the category response. Then, class evidence is obtained through a non-negative activation function:

[0073] ; in, Indicates the number of categories. The model represents the first The strength of evidence for a class. This differs from the approach used in basic EDL. R-EDL introduces adjustable relaxed prior parameters to replace the fixed prior form. And construct the Dirichlet distribution parameters:

[0074] ; Further define the total amount of evidence: ; The predicted probabilities of the R-EDL branch can be obtained from the Dirichlet distribution: In this embodiment, the uncertainty of the R-EDL branch is expressed in the form of vacuum under relaxed prior weights: ; in, Used to control the magnitude of prior weights. When the value is small, the model prediction relies more on the proportion of class evidence; when When the value is large, the predicted probability is more influenced by the total amount of evidence. Therefore, The model can adjust how it uses the proportion and magnitude of evidence, allowing the R-EDL branch to maintain a more conservative predictive behavior in the uncertain region.

[0075] On labeled data, the R-EDL branch employs a combination of Dice loss and R-EDL evidence loss for optimization. The Dice loss is used to constrain the consistency between the predicted segmentation results and the manually labeled regions.

[0076] ; in, This indicates accurate labeling. This represents the predicted probability output by the R-EDL branch.

[0077] The core idea of ​​R-EDL evidence loss is to directly optimize the projected probability corresponding to the expected value of the Dirichlet distribution, rather than calculating the expected value MSE of all class probabilities following the Dirichlet distribution as in traditional EDL. Traditional EDL's optimization objective includes a variance minimization term, which pushes the Dirichlet distribution towards a sharper form, potentially leading to over-concentration of evidence in regions of insufficient evidence and reducing the reliability of uncertainty estimation. R-EDL removes this variance minimization term, constraining only the difference between the projected probability and the true label. Its data term can be represented as:

[0078] ; This loss directly constrains the consistency between the Dirichlet distribution mean prediction and the true label, while avoiding additional compression of the Dirichlet distribution variance, thus allowing the model to retain some uncertainty when sufficient evidence is lacking. To suppress the accumulation of anomalous evidence on the wrong categories, this embodiment further introduces a KL regularization term. Unlike the basic EDL, which constrains non-true categories to a uniform Dirichlet prior, the R-EDL branch uses a Dirichlet prior consistent with the relaxed prior:

[0079] ; For the real category, construct a relaxed form. : ; In this case, the Dirichlet parameter corresponding to the real category is reset to... Non-true categories retain current prediction evidence. Therefore, the KL regularization term of the R-EDL branch can be written as:

[0080] ; Combining the projection probability error term and the KL regularization term, the R-EDL evidence loss is defined as: ; in, The annealing weights are gradually increased during training. In the early stages of training, the model primarily learns basic segmentation capabilities. As training progresses, the KL regularization is gradually enhanced to suppress erroneous evidence on non-true categories and mitigate the model's overconfident predictions in uncertain regions.

[0081] Therefore, the supervised loss of the R-EDL branch on labeled samples is: ; in, The dynamic weights represent the evidence loss. Through the above design, the R-EDL branch can learn the basic segmentation results and maintain a more conservative evidence representation in regions with insufficient evidence or class confusion, avoiding the result of excessively concentrated evidence distribution and low prediction uncertainty caused by the model relying solely on hard label supervision.

[0082] On unlabeled data, the R-EDL branch also participates in teacher-student consistency learning. The teacher model generates predicted probabilities for unlabeled images. and uncertainty The student model then generates predicted probabilities for the same unlabeled image after strong enhancement. To reduce the impact of unreliable pseudo-labels on training, this embodiment constructs consistency weights based on the uncertainty of the teacher network predictions:

[0083] ; The internal consistency loss of a branch is defined as: ; When the teacher model has low uncertainty in a certain region, the pseudo-labels in that region are considered more reliable, and their consistency supervision weight is higher; when the uncertainty is high, the supervision contribution of that region is automatically reduced. This avoids the R-EDL branch from overlearning unreliable pseudo-labels in regions with ambiguous boundaries or confused categories.

[0084] Furthermore, the R-EDL branch also participates in cross-branch consistency learning together with the Fisher-EDL branch. Specifically, the predictions of the R-EDL student branch... Predictions with the Fisher-EDL teacher branch Apply consistency constraints:

[0085] ; in, The factors are determined by the consistency of bibranch teacher predictions, the uncertainty of the uniform scale, and the divergence of bibranch probabilities, which will be discussed in detail in the next chapter.

[0086] (V) Construction of Uncertainty and Regional Risk Maps with Two Branches and Unified Scale Because the Fisher-EDL branch and the R-EDL branch use different Dirichlet prior settings, the uncertainty of their direct outputs is not on a completely consistent numerical scale.

[0087] R-EDL branches adopt a relaxed prior approach: ; when Even if two branches extract similar levels of evidence for the same pixel, their original uncertainty values ​​may exhibit systematic differences. Directly comparing or weighting the original uncertainties of the two branches could lead to a bias in risk assessment towards one branch, thus affecting the stability of cross-branch consensus learning. Therefore, in a two-branch framework, it is necessary to uniformly scale uncertainties under different prior settings.

[0088] To address the aforementioned issues, this embodiment introduces a unified scale, vacuum, during the cross-branch risk assessment phase. This scale applies to the category response of any branch output. First, class evidence is obtained through a non-negative activation function:

[0089] ; Then calculate the total amount of evidence: ; Based on this, the uncertainties of different branches are uniformly mapped to the base EDL reference ruler: ; in, Indicates the number of categories. This represents the total amount of evidence for the current pixel. This definition does not directly use the prior parameters of Fisher-EDL or R-EDL, but instead uses the same reference prior for uncertainty calculation. Therefore, regardless of the original branch... still Their uncertainties can all be transformed into a unified scale for comparison.

[0090] After obtaining the unified scale uncertainty, this embodiment further constructs a risk map by combining the bi-branch prediction divergence. Let the predictions of the Fisher-EDL teacher branch and the R-EDL teacher branch on unlabeled samples be respectively... and The corresponding prediction category is:

[0091] ; First, construct the bi-branch prediction consistency mask: ; Where i represents the pixel position. When two teacher branches predict the same position consistently, it indicates that the region has high false label confidence; when the two predictions are inconsistent, it indicates that the region may have class confusion or boundary uncertainty, and therefore it is not advisable to directly apply strong supervision.

[0092] In addition to the consistency of hard predictions, this embodiment also introduces a two-branch probability divergence metric. Let the probability distributions obtained by applying softmax to the logits of the two teacher branches be as follows: and Then the branch prediction divergence can be expressed as:

[0093] ; in, The larger the value, the more significant the difference in predictions between the two branches at that location, and the higher the risk of false labels in that region; conversely, The smaller the value, the more consistent the predictions of the two branches are, and the more reliable the region is.

[0094] In cross-branch consensus learning, the Fisher-EDL student branch primarily learns the pseudo-labels provided by the R-EDL teacher branch; therefore, its risk weight should be more influenced by the uncertainty of the R-EDL teacher branch. Conversely, the R-EDL student branch primarily learns the pseudo-labels provided by the Fisher-EDL teacher branch; therefore, its risk weight should be more influenced by the uncertainty of the Fisher-EDL teacher branch. Thus, this embodiment defines two-directional regional risk maps:

[0095] ; ; in, This indicates the risk when Fisher-EDL student branches accept false labels from R-EDL teachers. This represents the risk when the R-EDL student branch accepts the Fisher-EDL teacher pseudo-label. The design considers both the teacher branch's own uncertainty and the difference in predictions between the two branches: a lower risk value indicates lower teacher prediction uncertainty and smaller difference in predictions between the two branches; a higher risk value indicates higher teacher uncertainty or significant divergence between the two branches.

[0096] Furthermore, this embodiment transforms the risk graph into cross-branch consistency loss weights: ; ; in, This is the risk suppression coefficient, used to control the impact of risk values ​​on monitoring weights. When the risk is low, A value close to 1 indicates that the region is given a higher weight during cross-branch consistency training; when the risk is high, the weight is automatically reduced. Thus, the model can adaptively model the credibility of pseudo-labels in different regions.

[0097] Based on the above weights, the cross-branch consistency loss in both directions can be expressed as follows: ; ; in, and These represent the prediction results of the two student branches on the strongly enhanced unlabeled image, respectively. and These represent the prediction results of the two teacher branches on weakly enhanced unlabeled images. Through the aforementioned weighted consistency loss, low-risk regions can provide stronger training signals, while the supervisory role of high-risk regions is suppressed, thereby reducing the continuous propagation of false labels during training.

[0098] Based on the same inventive concept, this invention also provides a semi-supervised natural image segmentation system. The system includes a data acquisition module for reading labeled natural images, unlabeled natural images, and corresponding training configurations; a dual-evidence modeling module for constructing Fisher-EDL and R-EDL branches; a training module for performing supervised training, intra-module consistency training, and cross-branch consistency training; a risk control module for generating uniform scale uncertainty, prediction divergence, and region risk maps; and a segmentation module for outputting a segmentation mask and pixel-level reliability map for the natural image to be segmented.

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

[0100] The above describes a natural image segmentation method based on dual-evidence modeling, provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding natural image segmentation system based on dual-evidence modeling, such as... Figure 3 As shown.

[0101] Figure 3 This invention provides a schematic diagram of a natural image segmentation system based on dual-evidence modeling, the system comprising: The data acquisition module 301 is used to acquire unlabeled natural image data; The model building module 302 is used to build a dual-evidence segmentation network model based on the Mean Teacher model; the dual-evidence segmentation network model includes: parallel Fisher-EDL branches and R-EDL branches; both the Fisher-EDL branches and the R-EDL branches include parallel student model branches and teacher model branches; The model training module 303 is used to input natural image data into the two-branch teacher model and student model respectively, output the prediction results and pixel-level uncertainty of each pixel in the unlabeled natural image; and generate pseudo-labels for each pixel in the unlabeled natural image based on the prediction results of the two-branch teacher model; determine the prediction consistency mask and prediction divergence between the two-branch teacher model according to the prediction results output by the two-branch teacher model; determine the region risk map used to characterize the reliability of the pseudo-labels based on the prediction divergence and the pixel-level uncertainty output by the two-branch teacher model, and map the region risk map to cross-branch consistency loss weights using the prediction consistency mask; and construct a cross-branch consistency loss function based on the prediction results and cross-branch consistency weights output by the two-branch teacher model and student model to train the dual-evidence segmentation network model. The image segmentation module 304 is used to acquire the natural image to be segmented and input it into the trained dual-evidence segmentation network model, and output the image segmentation result of the natural image to be segmented.

[0102] For specific limitations regarding the natural image segmentation system based on dual-evidence modeling, please refer to the limitations of the natural image segmentation method based on dual-evidence modeling mentioned above, which will not be repeated here. Each module in the aforementioned natural image segmentation system based on dual-evidence modeling can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0103] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 This paper presents a natural image segmentation method based on dual-evidence modeling.

[0104] The present invention also provides Figure 4 The schematic diagram of the computer device shown is as follows: Figure 4 As shown, at the hardware level, this computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 1 This paper presents a natural image segmentation method based on dual-evidence modeling.

[0105] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

Claims

1. A natural image segmentation method based on dual-evidence modeling, characterized in that, include: Acquire unlabeled natural image data; Construct a dual-evidence segmentation network model based on the Mean Teacher model; The dual-evidence segmentation network model includes: a parallel Fisher-EDL branch and an R-EDL branch; both the Fisher-EDL branch and the R-EDL branch include a parallel student model branch and a teacher model branch. Natural image data is input into the two-branch teacher model and student model respectively, and the prediction results and pixel-level uncertainties of each pixel in the unlabeled natural image are output; and pseudo-labels of each pixel in the unlabeled natural image are generated based on the prediction results of the two-branch teacher model. Based on the prediction results output by the two-branch teacher model, a prediction consistency mask and prediction divergence between the two-branch teacher models are determined; based on the prediction divergence and the pixel-level uncertainty output by the two-branch teacher model, a regional risk map is determined to characterize the reliability of the pseudo-label, and the regional risk map is mapped to cross-branch consistency loss weights using the prediction consistency mask; based on the prediction results output by the two-branch teacher model and the student model and the cross-branch consistency weights, a cross-branch consistency loss function is constructed to train the dual-evidence segmentation network model. The natural image to be segmented is input into a trained dual-evidence segmentation network model, which outputs the image segmentation result of the natural image to be segmented.

2. The natural image segmentation method based on dual-evidence modeling as described in claim 1, characterized in that, The student and teacher models in both the Fisher-EDL and R-EDL branches are constructed based on evidence-based deep learning models. The Fisher-EDL branch is obtained by introducing the Fisher information matrix into the evidence deep learning model; the R-EDL branch is obtained by introducing adjustable relaxed prior parameters into the fixed prior Dirichlet distribution parameters of the evidence deep learning model.

3. The natural image segmentation method based on dual-evidence modeling as described in claim 1, characterized in that, The step of determining the prediction consistency mask and prediction divergence between the two-branch teacher models based on their prediction results includes: Based on the prediction results of the Fisher-EDL branch teacher model and the R-EDL branch teacher model for each pixel region of the unlabeled natural image, the prediction category of each pixel region of the natural image is determined. Based on whether the prediction categories of each pixel region are consistent between the two branch teacher models, a prediction consistency mask is generated between the two branch teacher models. The prediction divergence is determined based on the difference in prediction probabilities between the prediction results of the two-branch teacher models for each pixel region.

4. The natural image segmentation method based on dual-evidence modeling as described in claim 1, characterized in that, The determination of the regional risk map, based on the predicted divergence and the pixel-level uncertainty output by the two-branch teacher model, to characterize the reliability of the pseudo-label specifically includes: The pixel-level uncertainty output by the Fisher-EDL branch teacher model and the R-EDL branch teacher model is transformed to a unified scale to obtain the unified scale uncertainty of natural image data in the two branches. Based on the uniform scale uncertainty and prediction discrepancy corresponding to the R-EDL branch teacher model, a regional risk map of the Fisher-EDL branch teacher model is determined; the regional risk map of the Fisher-EDL branch teacher model is used to characterize the reliability of the Fisher-EDL branch student model in learning the pseudo-labels corresponding to the R-EDL branch teacher model. Based on the uniform scale uncertainty and prediction discrepancy corresponding to the Fisher-EDL branch teacher model, a regional risk map of the R-EDL branch teacher model is determined; the regional risk map of the R-EDL branch teacher model is used to characterize the reliability of the R-EDL branch student model in learning the pseudo-labels corresponding to the Fisher-EDL branch teacher model.

5. The natural image segmentation method based on dual-evidence modeling as described in claim 1, characterized in that, The construction of a cross-branch consistency loss function based on the prediction results output by the two-branch teacher model and the student model, and the cross-branch consistency weights, specifically includes: Based on the cross-branch consistency weights corresponding to the Fisher-EDL branch, the prediction results output by the student model in the Fisher-EDL branch, and the prediction results output by the teacher model in the R-EDL branch, the cross-branch consistency loss function corresponding to the Fisher-EDL branch is constructed, and its expression is: ； In the formula, The cross-branch consistency loss corresponding to the Fisher-EDL branch. This represents the prediction results of the Fisher-EDL branch student model on natural image data. This represents the prediction results of the R-EDL branch teacher model for natural image data. The cross-branch consistency weights corresponding to the Fisher-EDL branches. To prevent extremely small positive numbers with a denominator of zero; Based on the cross-branch consistency weights corresponding to the R-EDL branches, the prediction results output by the student model in the R-EDL branch, and the prediction results output by the teacher model in the Fisher-EDL branch, the cross-branch consistency loss function corresponding to the R-EDL branches is constructed, and its expression is: ； In the formula, This represents the cross-branch consistency loss corresponding to the R-EDL branch. The cross-branch consistency weights corresponding to the R-EDL branches. This represents the prediction results of the R-EDL branch student model on natural image data. This represents the prediction results of the Fisher-EDL branch teacher model for natural image data.

6. The natural image segmentation method based on dual-evidence modeling as described in claim 1, characterized in that, Training the dual-evidence segmentation network model also includes: Acquire labeled natural image data; Labeled natural image data is input into a two-branch student model, and the class evidence corresponding to the two branches of the student model is output respectively. Based on the category evidence, construct the Dirichlet distribution parameters for the two branches; and based on the Dirichlet distribution parameters, determine the prediction results of the student model for the labeled natural image for the two branches. Based on the prediction results of the two-branch student model on labeled natural image data and their corresponding real labels, Dice loss function and evidence loss are constructed respectively to conduct supervised training of the two-branch student model.

7. The natural image segmentation method based on dual-evidence modeling as described in claim 6, characterized in that, The process of constructing the evidence loss for the Fisher-EDL branch student model specifically includes: Construct the Fisher information matrix based on the Dirichlet distribution parameters of the Fisher-EDL branch student model; Based on the difference between the prediction results of the two-branch student model on labeled natural image data and their corresponding true labels, as well as the Fisher information matrix, the Fisher weighted error term is determined. A KL regularization term is introduced, and the evidence loss of the Fisher-EDL branch student model is constructed based on the Fisher weighted error term, the determinant of the Fisher information matrix, and the KL regularization term.

8. The natural image segmentation method based on dual-evidence modeling as described in claim 6, characterized in that, The process of constructing the evidence loss for the R-EDL branch student model specifically includes: An adjustable relaxation prior parameter is introduced, and the Dirichlet distribution parameters of the R-EDL branch are constructed based on the adjustable relaxation prior parameter and the category evidence output by the R-EDL branch student model. The prediction results of the R-EDL branch student model for the labeled natural image are determined based on the Dirichlet distribution parameters. Based on the difference between the prediction result and the true label of the labeled natural image data, a projection probability error term is determined; Based on the Dirichlet distribution parameters and the relaxed prior parameters, construct the relaxed form Dirichlet parameters corresponding to the true label; determine the KL regularization term of the R-EDL branch based on the KL divergence between the Dirichlet distribution parameters and the relaxed form Dirichlet parameters. Based on the projection probability error term and the KL regularization term of the R-EDL branch, construct the evidence loss of the R-EDL branch student model.

9. The natural image segmentation method based on dual-evidence modeling as described in claim 1, characterized in that, Training the dual-evidence segmentation network model also includes: For the teacher and student models within any branch, based on the prediction results and pixel-level uncertainties output by the teacher and student models, an intra-branch consistency loss is constructed, specifically including: For the teacher model and student model within any branch, the prediction difference between the teacher model and the student model is determined based on the prediction results of the student model on natural image data and the prediction results of the teacher model on natural image data. Based on the pixel-level uncertainty of the teacher model output, construct the consistency weights within this branch; Based on the prediction differences within the branch and the consistency weight within the branch, an intra-branch consistency loss is constructed.

10. A natural image segmentation system based on dual-evidence modeling, characterized in that, include: The data acquisition module is used to acquire unlabeled natural image data; The model building module is used to build a dual-evidence segmentation network model based on the Mean Teacher model. The dual-evidence segmentation network model includes: parallel Fisher-EDL branches and R-EDL branches. Both the Fisher-EDL branches and the R-EDL branches include parallel student model branches and teacher model branches. The model training module is used to input natural image data into the two-branch teacher model and student model respectively, output the prediction results and pixel-level uncertainty of each pixel in the unlabeled natural image; and generate pseudo-labels for each pixel in the unlabeled natural image based on the prediction results of the two-branch teacher model; determine the prediction consistency mask and prediction divergence between the two-branch teacher model according to the prediction results output by the two-branch teacher model; determine the region risk map to characterize the reliability of the pseudo-labels based on the prediction divergence and the pixel-level uncertainty output by the two-branch teacher model, and map the region risk map to cross-branch consistency loss weights using the prediction consistency mask; and construct a cross-branch consistency loss function based on the prediction results and cross-branch consistency weights output by the two-branch teacher model and student model to train the dual-evidence segmentation network model. The image segmentation module is used to acquire the natural image to be segmented and input it into the trained dual-evidence segmentation network model, and output the image segmentation result of the natural image to be segmented.

Images