A sample training optimization method, system and device for an image segmentation model

Abstract

The application provides a sample training optimization method, system and device for an image segmentation model, relates to the technical field of remote sensing image segmentation, and generates a comprehensive cognitive load score through a state decoding model; then, in combination with a professional portrait of a user, a cognitive calibration factor for current boundary correction is generated, the cognitive calibration factor is mapped into a confidence weight of a boundary correction label; a target vector boundary in an automatic segmentation result of a remote sensing image, corresponding confidence weights, corresponding remote sensing image data, geographical information data and environmental parameter data of the remote sensing image are associated, and a dynamic sample library is constructed; finally, in the process of training an image segmentation model by using the dynamic sample library, a training loss function supervised by a boundary correction label weighted based on the confidence weight is constructed, so as to optimize the segmentation precision of the image segmentation model in a boundary area. The application can optimize the performance of the model, thereby improving the segmentation accuracy and training quality.

Description

Technical Field

[0001] This application relates to the technical field of remote sensing image segmentation, and in particular to a sample training optimization method, system, and device for image segmentation models. Background Technology

[0002] In remote sensing image interpretation scenarios such as natural resource monitoring and urban planning, image segmentation models are the core tools for achieving automated ground feature extraction. The performance of these models is highly dependent on the quality, scale, and accuracy of the annotations of the training samples, especially for samples with difficult-to-learn boundary regions and easily confused ground feature categories. Therefore, optimization methods for the sample training process, particularly how to efficiently acquire high-quality labeled data and fully utilize its training value, have significant research significance and application prospects.

[0003] Currently, there are two main typical modes in related technical methods: First, there is the direct application of general-purpose models pre-trained on large benchmark datasets. While these models possess a certain generalization ability, their segmentation accuracy, especially the accuracy of boundary localization, often decreases significantly when facing specific regions, special time phases, or new land features, and the model performance degrades as data distribution shifts. Second, there is an interactive workflow combining model prediction and manual correction. Expert users review the model's segmentation results and manually correct boundaries to produce accurate vector data that meets the requirements of the current task.

[0004] However, existing interactive segmentation and sample training optimization methods have the following limitations: they treat user interaction correction as a homogeneous and reliable source of labels, ignoring the quality differences caused by the influence of users' cognitive state, professional level and task complexity during the labeling process. This results in the inability to differentiate and optimize corrected labels of different quality during model training, limiting the efficiency of sample utilization and the potential for improving model performance. Summary of the Invention

[0005] The purpose of this application is to provide a sample training optimization method, system, and device for image segmentation models, so as to solve the problems of low model segmentation accuracy and poor training quality caused by ignoring differences in annotation quality in the prior art.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a sample training optimization method for image segmentation models, comprising: Acquire boundary correction information based on the user's automatic segmentation results of remote sensing images, simultaneously capture the user's interaction behavior sequence, as well as the user's physiological state signals and neural response characteristics when performing the current boundary modification operation; By integrating the interaction behavior sequence, the physiological state signals, and the neural response features, a comprehensive cognitive load score is generated through a state decoding model. Based on the user's professional profile and the comprehensive cognitive load score, a cognitive calibration factor for current boundary correction is generated, and the cognitive calibration factor is mapped to the confidence weight of the boundary correction label. The target vector boundary, corresponding confidence weight, corresponding remote sensing image data, and geographic information data and environmental parameter data of the remote sensing image in the automatic segmentation result of the remote sensing image are associated to construct a dynamic sample library. During the training of the image segmentation model using the dynamic sample library, a training loss function supervised by the boundary correction labels weighted by the confidence weights is constructed to optimize the segmentation accuracy of the image segmentation model in the boundary region.

[0007] Optionally, by fusing the interaction behavior sequence, the physiological state signals, and neural response features, a comprehensive cognitive load score is generated through a state decoding model, including: The interaction behavior sequence is encoded to extract the behavior pattern feature vector, and the neural response features are encoded to extract the neural feature vector. Temporal analysis was performed on the eye-tracking data and mouse pressure change sequence in the physiological state signal to extract the corresponding cognitive features and obtain the cognitive feature vector; The behavioral pattern feature vector, the cognitive feature vector, and the neural feature vector are input into the spatiotemporal attention module to calculate the mutual information weights between different vectors in the time dimension, thereby obtaining the mutual information weight matrix. When the number of mutual information weights in the mutual information weight matrix that are less than a preset weight threshold exceeds a preset proportion, a cognitive inconsistency score is generated and output. The behavioral pattern feature vector, the cognitive feature vector, the neural feature vector, and the cognitive inconsistency score are input into the state decoding model. Through the attention-guided feature fusion layer in the state decoding model, the weight ratio of the behavioral pattern feature vector, the cognitive feature vector, and the neural feature vector in the fusion process is dynamically adjusted according to the cognitive inconsistency score. The fused features are then decoded to output a five-dimensional vector, which corresponds to five independent dimensions: attention concentration, cognitive load, decision certainty, cognitive consistency, and neural response stability. The five-dimensional vector is mapped to a comprehensive cognitive load score through a nonlinear mapping layer in the state decoding model.

[0008] Optionally, through the attention-guided feature fusion layer in the state decoding model, the weight ratios of the behavioral pattern feature vector, cognitive feature vector, and neural feature vector in the fusion process are dynamically adjusted based on the cognitive inconsistency score, including: The behavioral pattern feature vector, cognitive feature vector, and neural feature vector are projected onto a unified latent space to obtain the corresponding first latent feature, second latent feature, and third latent feature. Based on cognitive inconsistency scoring, multi-level attention adjustment is performed on different vectors to obtain the fusion weights of different vectors, and the corresponding latent features are weighted. The weighted first, second, and third latent features are input into a gating fusion unit. The gating fusion unit dynamically controls the contribution ratio of each latent feature in the fused features based on the predicted initial values ​​of the cognitive consistency dimension and the neural response stability dimension, so as to generate the fused features.

[0009] Secondly, this application provides a sample training optimization system for an image segmentation model, comprising: The acquisition module is used to acquire the boundary correction information of the user based on the automatic segmentation results of remote sensing images, and simultaneously capture the user's interaction behavior sequence, as well as the physiological state signals and neural response characteristics of the user when performing the current boundary modification operation. The fusion module is used to fuse the interaction behavior sequence, the physiological state signal, and the neural response features, and generate a comprehensive cognitive load score through a state decoding model. The generation module is used to generate a cognitive calibration factor for current boundary correction based on the user's professional profile and the comprehensive cognitive load score, and to map the cognitive calibration factor to the confidence weight of the boundary correction label. The first construction module is used to associate the target vector boundary, the corresponding confidence weight, the corresponding remote sensing image data, and the geographic information data and environmental parameter data of the remote sensing image in the automatic segmentation result of the remote sensing image to construct a dynamic sample library. The second construction module is used to construct a training loss function supervised by the boundary correction label weighted by the confidence weight during the training of the image segmentation model using the dynamic sample library, so as to optimize the segmentation accuracy of the image segmentation model in the boundary region.

[0010] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of a sample training optimization method for an image segmentation model as described in the first aspect above.

[0011] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of a sample training optimization method for an image segmentation model as described in the first aspect above.

[0012] The sample training optimization method for image segmentation models provided in this application has the following beneficial effects: First, a multimodal cognitive data acquisition system was constructed by simultaneously capturing the user's interactive behavior sequence, physiological state signals, and neural response characteristics during the boundary correction process. Next, the aforementioned multi-source heterogeneous data was innovatively fused, and a comprehensive cognitive load score was generated through a state decoding model. This enabled precise quantification of the user's cognitive load and decision confidence during labeling, transforming the originally subjective and implicit degree of hesitation into an objective and calculable indicator. Then, based on the user's professional profile and the aforementioned comprehensive cognitive load score, a cognitive calibration factor for the current boundary correction was generated and mapped to the confidence weight of the label. This intelligently distinguishes samples with different levels of confidence, providing a scientific basis for subsequent differentiated training. Subsequently, the generated target vector boundaries, corresponding confidence weights, remote sensing image data, and geographic environmental parameters are correlated to jointly construct a dynamic sample library. This not only accumulates training data but also provides each sample with its own quality certification, upgrading the sample library into an enhanced resource library with quality annotations. Finally, during the model training phase, a training loss function based on confidence weights is constructed. This design allows the model to prioritize learning accurate corrections with high confidence while effectively resisting interference from low-quality annotations, thereby improving the overall image segmentation model, especially in terms of segmentation accuracy and training efficiency in complex boundary regions.

[0013] Furthermore, this application achieves accurate quantitative assessment of users' cognitive states through multimodal feature fusion and intelligent decoding technology. Specifically: First, by performing feature encoding and time-frequency analysis on multi-source information such as interactive behavior, eye-tracking data, and physiological signals, representative feature vectors are extracted; then, a spatiotemporal attention mechanism is introduced to calculate cross-modal mutual information weights, effectively identifying the level of cognitive consistency between different information sources; when significant cognitive inconsistency is detected, an inconsistency score is automatically generated as an important moderating signal; through an attention-guided feature fusion layer, the weight ratio of each modality feature is dynamically adjusted to ensure higher reliability of the cognitive assessment results; the gating fusion unit further optimizes the feature fusion process, and finally outputs a comprehensive score containing five cognitive dimensions.

[0014] Therefore, this multi-level, adaptive fusion mechanism can effectively overcome the limitations of single-modal data and improve the accuracy and robustness of cognitive load assessment. By transforming implicit user cognitive states into quantifiable assessment indicators, it provides a reliable cognitive quantification basis for subsequent sample quality calibration and model training optimization. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a sample training optimization method for an image segmentation model provided in an embodiment of this application; Figure 2 A schematic diagram illustrating a specific implementation of a sample training optimization method for an image segmentation model provided in this application embodiment; Figure 3 A schematic diagram of the structure of a sample training and optimization system for an image segmentation model provided in this application embodiment; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] In applications such as natural resource monitoring, the training effect of remote sensing image segmentation models is crucial. However, existing methods have the following limitations: they treat user interaction corrections as a homogeneous and reliable source of labels, ignoring the quality differences caused by the influence of users' cognitive state, professional level, and task complexity during the labeling process. This results in the inability to differentiate and optimize correction labels of different quality during model training, limiting the efficiency of sample utilization and the potential for improving model performance.

[0018] To address this, this application proposes a sample training optimization method for image segmentation models. The core idea of ​​this method is to achieve a technological leap from manual annotation to cognitive empowerment through multimodal cognitive signal fusion and quantitative evaluation. Specifically, the system first simultaneously captures the user's interaction behavior sequence, physiological state signals, and neural response features, generating a comprehensive cognitive load score through a state decoding model. Then, based on the user's professional profile and cognitive score, a cognitive calibration factor is generated and mapped to the confidence weights of boundary correction labels. Finally, during the model training phase, a training loss function based on confidence weights is constructed, driving the model to prioritize learning high-confidence, accurate corrections while effectively resisting interference from low-quality annotations.

[0019] Therefore, the above method makes the implicit cognitive information in the human-computer interaction process explicit and structured, which not only improves the segmentation accuracy of the model in complex boundary areas, but also achieves a substantial breakthrough in training efficiency and effect, providing a new technical paradigm for intelligent interpretation of remote sensing images.

[0020] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The core of this application is to provide a sample training optimization method for image segmentation models, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes: S101. Obtain the boundary correction information of the user based on the automatic segmentation results of remote sensing images, and simultaneously capture the user's interaction behavior sequence, as well as the physiological state signals and neural response characteristics of the user when performing the current boundary modification operation.

[0022] Among them, boundary correction information is the correction result of the user's judgment on the boundary shape or category of the initial patch. This information may include the user's marking of the inaccuracies of the boundary or semantic label of the land cover category of the initial patch in the automatic segmentation result of remote sensing image. Interactive behavior sequence refers to the complete and orderly action log generated by the user during the above-mentioned correction operation, which is automatically recorded by the system background. It may include at least one of mouse hover trajectory, click interval duration, and local view zoom operation frequency. Physiological state signals include at least one of eye tracking data, mouse pressure change, and operation interval time. Neural response characteristics refer to the response signals generated by the central nervous system activity and captured by specific technical means when a user performs boundary modification operations. These technical means include, but are not limited to, electroencephalogram (EEG) signals recorded by EEG acquisition equipment or cerebral blood oxygenation signals recorded by functional near-infrared spectroscopy equipment.

[0023] Before performing step S101, this application can perform initial segmentation processing on the remote sensing image data of the area to be processed using an image segmentation model, generate automatic segmentation results of the remote sensing image, and display the automatic segmentation results of the remote sensing image to the user.

[0024] The automatic segmentation result of the remote sensing image is a plurality of initial patches covering the area to be processed. Each initial patch is associated with a semantic label of a land cover category, and the vector boundaries of the plurality of initial patches are spliced ​​together or adjacent to each other in space.

[0025] Among them, the semantic label of the land feature category refers to a category name automatically assigned by the system to each initial land feature, such as farmland, road, water system, vegetation, building, etc., which is used to initially describe the land feature type that the area may represent; the vector boundary refers to the outline line used to describe the shape and extent of each land feature in geospace. These boundary lines are connected to each other to ensure that all areas are covered without omission or overlap.

[0026] It should be noted that the structure of the image segmentation model can be a convolutional neural network and a support vector machine connected in sequence, and the image segmentation model can also adopt other known structures, which are not limited in this application.

[0027] During the execution of S101, the automatic segmentation results of remote sensing images can be presented to the user through a visual interface, allowing the user to view and judge them. When the user finds that the boundary of a certain patch is not drawn accurately or the assigned category label is incorrect, the user can make correction marks at the corresponding location through operations such as clicking, dragging, outlining, or selecting. At the same time, the system not only receives and understands the marking content itself used to indicate errors, but also automatically and continuously records the physiological state signals and neural response characteristics of the user during the marking process in the background. It uses a unified hardware clock to accurately timestamp the data streams from different sensors, such as mice, eye trackers, and EEG devices, to achieve synchronous capture and alignment of multimodal data, thereby forming a complete operation log.

[0028] S102. By integrating the interaction behavior sequence, the physiological state signal, and the neural response features, a comprehensive cognitive load score is generated through a state decoding model.

[0029] The state decoding model may include feature fusion layers, nonlinear mapping layers, etc. Model training can be conducted through supervised learning. For example, a set of user-annotated samples can be pre-collected, and experts can be invited or post-hoc accuracy verification can be used to annotate each sample with a cognitive load reference value representing its correction quality. This serves as a supervisory signal for training the model. It should be noted that this embodiment does not impose specific limitations on the model's structural design, parameter design, or the structural design within each layer; these can be set according to actual conditions.

[0030] The comprehensive cognitive load score is a scalar value within the range [0, 1] after model normalization. The higher the value, the greater the cognitive load, the more distracted the user is, and the lower the decision certainty when the system judges the user to be performing the current correction operation. This indicates a higher potential error risk and lower reliability of the correction result.

[0031] In one specific implementation, such as Figure 2 As shown, step S102 includes the following steps: Step 1021: Encode the interaction behavior sequence and extract the behavior pattern feature vector. At the same time, encode the neural response features and extract the neural feature vector.

[0032] Among them, the behavioral pattern feature vector may include the spectral characteristics of the operation speed, the fractal dimension of the operation trajectory, and the topological difference between the corrected path and the initial predicted boundary in the automatic segmentation result of the remote sensing image. Neural feature vectors can include the pupil's response delay to dynamic visual feedback, the synchronization error between mouse trajectory and boundary deformation animation, etc.

[0033] In step 1021, firstly, the interaction behavior sequence is encoded. Specifically, for the operation flow containing spatiotemporal information generated when the user corrects the boundary of remote sensing image patches, such as mouse movement, clicking, dragging vertices, etc., this embodiment can use a sequence model such as a recurrent neural network or a Transformer encoder to learn the operation flow. Then, the model captures the temporal dependencies generated by the user in the long-term, possibly intermittent boundary editing process, for example, the pattern difference between the smooth operation when continuously drawing smooth forest edges and the hesitation and pause when repeatedly considering the corner of a building, and outputs a behavior pattern feature vector that can summarize the user's operation pattern under the current boundary correction task. Simultaneously, neural response characteristics are encoded. Specifically, for the raw neural signals synchronously acquired by the EEG device when the user performs high-precision boundary correction, this embodiment of the application performs a standardized preprocessing procedure to adapt to computational analysis. The preprocessing includes filtering and segmentation operations. Filtering removes noise, ensuring signal quality suitable for cognitive state analysis. Segmentation, based on each specific boundary modification action performed by the user in the interactive interface, such as dragging a boundary point or erasing the precise start and end timestamps of an incorrectly segmented line segment, trims the continuous neural signals into segments corresponding to each operation. The time segments are aligned with the units; then, frequency domain features closely related to cognitive load and visual attention are extracted from these clean, operation-aligned signal segments, such as calculating the average power spectral density of each segment in four standard EEG bands of 1-4Hz, 4-8Hz, 8-13Hz, and 13-30Hz; finally, these frequency band power features reflecting different cognitive dimensions are concatenated and input into a shallow fully connected neural network for fusion and dimensionality reduction, thereby generating a fixed-dimensional neural feature vector that can characterize the user's neurocognitive state when performing a specific boundary correction action.

[0034] Step 1022: Perform time-domain analysis on the eye-tracking data and mouse pressure change sequence in the physiological state signal, extract the corresponding cognitive features, and obtain the cognitive feature vector.

[0035] In step 1022, the eye-tracking data is first processed. Specifically, for the visual behavior data generated by the user when reviewing and correcting the boundaries of remote sensing images, this embodiment of the application extracts temporal cognitive features that match the current specific boundary correction operation task window from continuous gaze point coordinate sequences and pupil diameter sequences; these features aim to quantify the user's visual attention allocation and cognitive effort when distinguishing complex textures (such as the transition zone between woodland and cultivated land) or accurately locating ambiguous boundaries (such as the edge of a water body).

[0036] The specific extraction process is as follows: Three core temporal indicators are extracted from the eye-tracking data. First, the duration of fixation is extracted, which reflects the depth of information processing. Second, the saccade amplitude between adjacent fixations is extracted, which reflects the visual search range. Third, the rate of change of pupil diameter relative to the user's resting baseline is extracted, which reflects the level of cognitive load. Subsequently, the mean, standard deviation, and maximum value of these three indicators are calculated within the current boundary correction operation time window, resulting in nine statistics. These nine statistics constitute a stable and interpretable set of eye-tracking cognitive features.

[0037] Secondly, the mouse pressure change sequence is processed. Specifically, for the pressure changes applied by the user when performing fine spatial operations such as dragging vector boundaries and fine-tuning vertices with the mouse, this application embodiment also uses a time-domain analysis method to evaluate the stability of the operation and decision confidence.

[0038] The specific extraction process is as follows: Read the numerical sequence recorded by the pressure sensor and calculate three key time-domain statistics throughout the corresponding operation process: First, calculate the pressure mean, which reflects the average intensity or degree of effort in the operation; second, calculate the pressure variation variance, which reflects the fluctuation of the operation intensity, with a large variance indicating hesitation or frequent adjustments; third, calculate the average slope of the rising edge of the pressure at key operation points, such as when the dragging peak begins, which reflects the decisiveness of the decision.

[0039] Finally, the 9-dimensional temporal statistical features extracted from eye-tracking data and the 3-dimensional temporal statistical features extracted from mouse pressure data are concatenated to form a comprehensive 12-dimensional cognitive feature vector. This vector can jointly characterize the user's real-time cognitive state when performing the current remote sensing image boundary correction task from both visual cognition and operational behavior dimensions.

[0040] Step 1023: Input the behavioral pattern feature vector, the cognitive feature vector, and the neural feature vector into the spatiotemporal attention module, calculate the mutual information weights between different vectors in the time dimension, and obtain the mutual information weight matrix.

[0041] In step 1023, the three time-series data—behavioral pattern feature vector, cognitive feature vector, and neural feature vector—are simultaneously input into the spatiotemporal attention module. This module first ensures that the time steps of the three vectors are aligned. Then, the sequence of values ​​of each feature vector, such as the behavioral pattern vector, across all time steps is treated as an observation of a random process. For each time step t, the mutual information between any two feature vectors, such as the behavioral pattern vector and the cognitive feature vector, is calculated. Mutual information is an indicator that measures the amount of information shared between two random variables. For example, its expression can be: ; in, It is the mutual information value, which is a measure of the interdependence between X and Y. It is the joint probability distribution of X and Y. , It is the corresponding marginal probability distribution, where x and y represent the values ​​of the two feature vectors at time step t, and X and Y are the value spaces of x and y, respectively. Obtain the mutual information values ​​between all feature vector pairs at all time steps. Then, the softmax function is used to normalize these mutual information values ​​to convert them into weights; these weights reflect the overall interdependence strength between different feature vector pairs. Finally, the mutual information weights of all feature vector pairs at all time steps are organized into a three-dimensional tensor or matrix form of mutual information weight matrix, where two dimensions of the three-dimensional tensor represent different feature vector pairs, and the third dimension represents the time step.

[0042] Step 1024: When the number of mutual information weights in the mutual information weight matrix that are less than a preset weight threshold exceeds a preset proportion, a cognitive inconsistency score is generated and output.

[0043] The preset weight threshold is the lower limit threshold of the absolute value of mutual information weight, such as 0.3. The preset ratio can be the threshold of the proportion of low mutual information weights, such as 0.4.

[0044] In step 1024, the mutual information weight matrix is ​​traversed, and the number of weights with mutual information weight values ​​lower than a preset weight threshold is counted. Then, the proportion of these low mutual information weights to the total number of weights in the matrix is ​​calculated, and this proportion is compared with a preset proportion threshold. If the proportion of low mutual information weights exceeds the proportion threshold, it is determined that the user has significant cognitive inconsistency. Based on the degree to which the proportion threshold is exceeded, an example function is used: Calculate a cognitive inconsistency score within a predetermined range. Where r is the proportion of low mutual information weights, This is a proportional threshold, such as 0.4, and The closer to 1, the more severe the cognitive inconsistency; if it does not exceed the proportional threshold... Cognitive inconsistency score =0.

[0045] Step 1025: Input the behavioral pattern feature vector, the cognitive feature vector, the neural feature vector, and the cognitive inconsistency score into the state decoding model. Through the attention-guided feature fusion layer in the state decoding model, the weight ratio of the behavioral pattern feature vector, the cognitive feature vector, and the neural feature vector in the fusion process is dynamically adjusted according to the cognitive inconsistency score. The fused features are then decoded to output a five-dimensional vector, which corresponds to five independent dimensions: attention concentration, cognitive load, decision certainty, cognitive consistency, and neural response stability.

[0046] For example, the feature fusion layer can be implemented by a module consisting of two linear transformation layers and a scaled dot product attention mechanism, with the key and value dimensions designed to be 64. It queries the cognitive inconsistency rating context after linear projection, thereby dynamically modulating the fusion weights of each modality feature.

[0047] Specifically, in step 1025, the attention-guided feature fusion layer in the state decoding model dynamically adjusts the weight ratios of the behavioral pattern feature vector, cognitive feature vector, and neural feature vector in the fusion process based on the cognitive inconsistency score, including: Step a1: Project the behavioral pattern feature vector, cognitive feature vector, and neural feature vector onto a unified latent space to obtain the corresponding first latent feature, second latent feature, and third latent feature.

[0048] In step a1, the behavioral pattern feature vector, cognitive feature vector, and neural feature vector are respectively input into three independent fully connected layers. These three fully connected layers have the same output dimension. Their function is to project the behavioral pattern feature vector, cognitive feature vector, and neural feature vector into a unified latent space so that they are converted into the first latent feature, the second latent feature, and the third latent feature, respectively. For example, the three fully connected layers can all be designed with an output dimension of 256 and without bias terms.

[0049] Step a2: Based on the cognitive inconsistency score, perform multi-level attention adjustment on different vectors to obtain the fusion weights of different vectors, and weight the corresponding latent features.

[0050] Step a2 includes the following process: Step a21: Determine the corresponding first-level weights based on the information entropy of each vector.

[0051] In step a21, the information entropy of the first latent feature, the second latent feature, and the third latent feature are calculated respectively. Information entropy is an indicator that measures the uncertainty of a random variable and is used in this application to evaluate the richness of information carried by each feature vector. Subsequently, the three information entropy values ​​are normalized using the softmax function, and the normalized result is the first-level weight corresponding to each latent feature. This weight reflects the proportion of each feature determined by its own information content when other contextual information is not considered.

[0052] Step a22: Based on the cognitive inconsistency score, calculate the decay factor of each vector. The decay factor is used to characterize the credibility of each vector when there is cognitive inconsistency. The decay factor corresponding to the behavioral pattern feature vector is negatively correlated with the cognitive inconsistency score, and the decay factor corresponding to the neural feature vector is positively correlated with the cognitive inconsistency score.

[0053] In step a22, based on the output cognitive inconsistency score, attenuation factors are calculated for the three latent features. For the first latent feature, which represents external behavior, the calculation of its attenuation factor makes it negatively correlated with the cognitive inconsistency score, that is, the higher the score, the smaller the attenuation factor value. For the second latent feature, its attenuation factor is usually set to 1. For the third latent feature, which represents internal neural activity, the calculation of its attenuation factor makes it positively correlated with the cognitive inconsistency score, that is, the higher the score, the larger the attenuation factor value.

[0054] It should be noted that the formula used in this embodiment for calculating the attenuation factor of each vector is not limited, and can be set accordingly based on the actual situation.

[0055] Step a23: Multiply the first-level weight by the corresponding modal confidence decay factor to obtain the second-level weight.

[0056] Step a23 dynamically modulates the credibility of the first-level weights based on the cognitive inconsistency detected in the current task, generating a more accurate and context-adaptive second-level weight.

[0057] Step a24: Calculate the third-level weights based on the mutual information values ​​between vectors, and then sum the second-level weights and the third-level weights to obtain the fusion weights of different vectors.

[0058] In step a24, for each latent feature, its mutual information values ​​with the other two latent features are summed to obtain a scalar value reflecting the overall correlation between the feature and other features. Then, the softmax function is used to normalize these scalar values ​​of correlation to obtain the third-level weight of each latent feature. Finally, the second-level weight of each latent feature and its third-level weight are linearly weighted and summed according to a preset mixing coefficient to obtain the final fusion weight of the latent feature.

[0059] Step a3: Input the weighted first latent feature, second latent feature and third latent feature into the gating fusion unit. The gating fusion unit dynamically controls the contribution ratio of each latent feature in the fused feature according to the predicted initial values ​​of the cognitive consistency dimension and the neural response stability dimension, so as to generate the fused feature.

[0060] The gating fusion unit can consist of a small, two-layer fully connected network used to generate gating coefficients and perform feature modulation.

[0061] In step a3, the weighted first, second, and third latent features are input into the gating fusion unit. This unit first uses a small fully connected network to preliminarily predict the initial values ​​of the cognitive consistency dimension and the neural response stability dimension based on the current weighted features. These two initial values ​​are used as important contextual information to generate a set of dynamic gating coefficients. The gating coefficients are vectors that match the input feature dimensions and determine the proportion of each weighted latent feature to be retained in the final output. Then, the gating fusion unit modulates the three weighted latent features using these gating coefficients, usually by element-wise multiplication. Finally, the three modulated feature vectors are added together to generate a unified fused feature.

[0062] Step 1026: Map the five-dimensional vector into a comprehensive cognitive load score through the nonlinear mapping layer in the state decoding model.

[0063] In step 1026, the output five-dimensional vector is input into the nonlinear mapping layer of the state decoding model. This layer typically consists of a two-layer fully connected network. The first layer maps the five-dimensional input to 16 dimensions and uses the ReLU activation function. The second layer maps the 16-dimensional input to 1 dimension and uses the Sigmoid activation function to restrict the output to between 0 and 1. Then, the fully connected network transforms and compresses the five-dimensional vector step by step in a forward propagation manner, and outputs a single scalar value in the last layer, namely the comprehensive cognitive load score. This score is the result of the model's high-level synthesis and trade-off of the information from the five independent cognitive dimensions previously decoded.

[0064] This application provides a scientific and reliable basis for the confidence weighting of subsequent training samples by transforming the annotation quality evaluation, which originally relied on subjective experience, into a computational process based on objective data. This fundamentally improves the intelligence and effectiveness of sample utilization in the active learning framework, thereby helping image segmentation models achieve more accurate learning and optimization in complex regions such as boundaries.

[0065] S103. Based on the user's professional profile and the comprehensive cognitive load score, generate a cognitive calibration factor for the current boundary correction, and map the cognitive calibration factor to the confidence weight of the boundary correction label.

[0066] The cognitive calibration factor is an intermediate moderating variable that combines the comprehensive cognitive load score, which reflects the immediate state, with the user's professional profile, which reflects long-term capabilities, to generate a more comprehensive and stable quality assessment value for this revision.

[0067] The boundary correction label represents the final result after user interaction in the data, and its format is consistent with the target vector boundary, such as polygon vector data in a geographic information system; during the training process, this label will be used as the ground truth of the supervision signal.

[0068] In one specific implementation, step 103 may include the following process: Step 1031: Based on the user's historical boundary correction records, construct a professional profile of the user.

[0069] In step 1031, all historical task records that the user has participated in are retrieved from the background database. For each historical record, the overlap index between the user's corrected boundary and the boundary verified by the gold standard is calculated to obtain the user's historical average annotation accuracy. Then, the user's specific accuracy in different land cover categories is statistically analyzed, and its operational efficiency index is analyzed. Finally, these standardized multi-dimensional indicators are combined and coded to form the user's professional profile.

[0070] Step 1032: Based on the professional profile and the comprehensive cognitive load score, generate preliminary cognitive calibration factors.

[0071] In step 1032, a preset fusion calculation module is invoked, which can be implemented by linear weighted summation. This module uses the core capability indicators in the professional profile as the basic trust value, and the comprehensive cognitive load score as the adjustment factor. The two are combined and calculated according to preset weight coefficients. For example, the weight coefficient of the professional profile is set to 0.6 and the weight coefficient of the comprehensive cognitive load score is set to 0.4 to output the preliminary cognitive calibration factor. The higher the value of this factor, the more reliable the quality of the correction is in the preliminary assessment.

[0072] Step 1033: Adjust the preliminary calibration factor to obtain the cognitive calibration factor.

[0073] In step 1033, this adjustment process can be based on two observations: First, adjustments should be made based on the objective complexity of the current boundary correction task itself. Specifically, the geometric complexity of the patch to be corrected and the degree of ambiguity of the surrounding terrain features should be assessed to form a quantifiable index of task complexity. When the task itself is relatively complex, the negative impact of high cognitive load exhibited by users on the initial calibration factor should be appropriately mitigated through preset rules or formulas. This is because a high cognitive load exhibited by users is a reasonable phenomenon under complex tasks and should not be entirely equated with a decline in annotation quality.

[0074] Second, adjustments are made based on the user's fatigue state resulting from continuous work. Specifically, a dynamic fatigue state model is maintained for each user. The structure of this model is not the focus and will not be elaborated here. This model estimates the user's continuous operation duration, operation intensity, and rest interval in the current work session and outputs a real-time ratio representing their remaining cognitive resource level. When it is determined that the user is in a fatigue state, that is, the remaining cognitive resource ratio is lower than a preset health threshold, the initial calibration factor is globally decayed to reflect the possible decrease in attention and increase in error rate caused by fatigue.

[0075] Combining the above two adjustments, the initial calibration factor is revised into a final cognitive calibration factor that more comprehensively and robustly reflects the actual reliability of this boundary correction. Step 1034: Calculate the confidence weights using a preset mapping function based on the cognitive calibration factor.

[0076] The preset mapping function can be a monotonically increasing non-linear function. For example, it can be the Sigmoid function. This embodiment does not limit the expression of this function.

[0077] In step 1034, the cognitive calibration factor is input into a predefined sigmoid function. The steepness coefficient in the sigmoid function can control the steepness of the function curve, thereby ensuring that high-quality input corresponds to high-weight output and low-quality input corresponds to low-weight output. After the function processes the cognitive calibration factor, the final confidence weight is output.

[0078] This application provides a precise basis for differentiated learning in model training by transforming subjective quality judgments into objective confidence weights, thereby effectively improving the contribution of high-quality samples and enhancing the model's resistance to labeled noise.

[0079] S104. Associate the target vector boundary, corresponding confidence weight, corresponding remote sensing image data, and geographic information data and environmental parameter data of the remote sensing image in the automatic segmentation result of the remote sensing image to construct a dynamic sample library.

[0080] The dynamic sample library constructed in this application has two main dynamic characteristics: first, the samples in the library and their associated confidence weights will continuously increase and accumulate during the human-computer interaction process; second, the library management strategy (such as spatial overlap optimization) will be dynamically adjusted according to the type and distribution of the samples, rather than being statically stored, making the sample library a training resource pool that can evolve on its own and has self-describing quality.

[0081] In one specific implementation, after step S104, the method may further include the following steps: Step 1041: Perform spatial overlap analysis and optimization on the samples in the dynamic sample library.

[0082] In step 1041, as samples are continuously added to the database, this application can periodically analyze the spatial relationship between all samples in the database, and evaluate the redundancy of the entire sample database in spatial distribution by calculating the overlapping area of ​​the corresponding regions of any two samples. For sample clusters with excessive overlap, an optimization process is initiated, such as screening based on indicators like sample generation time, quality score, or land cover integrity, to remove excessively redundant samples in order to maintain the spatial diversity and balance of the sample database.

[0083] Step 1042: Based on the type and scale of the land cover corresponding to the sample, dynamically adjust the spatial sampling strategy to control the overlap between samples within a preset range; wherein, the first overlap threshold of strip-shaped land cover is less than the second overlap threshold of area-like clustered land cover.

[0084] Strip-shaped features can be roads, rivers, linear forest belts, etc., while areal clustered features can be building complexes, lakes, concentrated and contiguous farmland, etc.

[0085] In step 1042, for strip-shaped features, due to their elongated shape, it is difficult for a single sample's image window to cover their entire length, and the continuity of their boundaries is crucial for image segmentation model learning. Therefore, this embodiment can set the maximum allowed spatial overlap threshold to 70%. This means that when the overlap area between the image region of a new sample and an existing sample in the database does not exceed 70%, it is allowed to be added to the database. This strategy ensures that a sufficient number of spatially continuous strip-shaped feature samples can be collected to support the image segmentation model in learning their complete structure.

[0086] For areal clusters of land features, their morphology is relatively compact, their local features are more representative, and they are prone to generating a large number of highly similar redundant samples in space. Therefore, this embodiment can set the maximum allowed spatial overlap threshold to 30%. This means that when the overlap area between a new sample and an existing sample in the database exceeds 30%, it will be rejected for inclusion in the database or a sample deduplication mechanism will be triggered. This strategy effectively avoids data redundancy and encourages samples to be distributed more spatially, thereby improving the diversity of training data and the generalization ability of the model.

[0087] This application, through this dynamic process, can adaptively maintain the overall spatial overlap of the sample library within the aforementioned threshold range, thereby optimizing the spatial distribution and preventing the training data from being overly concentrated in a specific region. This helps to improve the generalization ability of the image segmentation model in different geographical regions.

[0088] S105. During the training of the image segmentation model using the dynamic sample library, a training loss function supervised by the boundary correction label weighted by the confidence weight is constructed to optimize the segmentation accuracy of the image segmentation model in the boundary region.

[0089] In one specific implementation, step 105 includes: Step 1051: Construct a first function corresponding to pixel-level loss. The first function is used to compare the prediction result of each pixel with the user-corrected label, and the product of the first comparison result and the confidence weight corresponding to the pixel is used as the pixel-level loss.

[0090] The first function is used to measure the error of the model in pixel classification accuracy. This function can be a weighted cross-entropy loss function or other functions. This embodiment does not limit the expression of this function.

[0091] In step 1051, firstly, the image segmentation model performs forward propagation on the input remote sensing image and outputs the predicted probability distribution of each pixel belonging to each land cover category; simultaneously, it reads the user-corrected labels, which provide the ground truth value of the target category for each pixel; then, the first function compares the model's predicted probability distribution with the user-provided label ground truth value pixel by pixel and calculates a basic difference value that reflects the degree of prediction error of the pixel. This difference value measures the degree of prediction error of the model at that pixel. Finally, the first function multiplies the calculated base difference value with the confidence weight corresponding to the pixel position. This multiplication ensures that if the correction of the pixel comes from a user operation with high confidence, the resulting error will be fully used to correct the model; conversely, the pixel error from a low-confidence correction will have a weakened impact on the model update. The weighted error values ​​of all pixels obtained after the above calculation are summed to obtain the pixel-level loss.

[0092] Step 1052: Construct a second function corresponding to the boundary segment level loss. The second function is used to compare the prediction result on each continuous boundary segment with the user-corrected label, and the product of the second comparison result and the average confidence weight corresponding to the boundary segment is used as the boundary segment level loss.

[0093] The second function is used to measure the error of the model in maintaining the continuity and smoothness of the boundaries. This function can be the boundary regression loss function or other functions. This embodiment does not limit the expression of this function.

[0094] In step 1052, firstly, the complete outline boundary of each patch is extracted from the user-corrected label; then, the continuous outline is segmented into multiple continuous boundary segments using a corner detection algorithm, and for each boundary segment, the corresponding boundary pixel sequence in the model prediction result is obtained simultaneously; next, the shape difference between the predicted boundary segment and the real boundary segment is calculated, and a common method is to calculate the average distance between corresponding pixels of two line segments or to calculate the matching degree through distance transformation. Simultaneously, the average confidence weight of all pixels constituting the boundary segment is calculated to obtain the average confidence weight of the boundary segment. Finally, the calculated shape difference value is multiplied by the average confidence weight to obtain the weighted loss of the boundary segment. The weighted losses of all boundary segments are then summed to obtain the boundary segment-level loss.

[0095] Step 1053: Construct the third function corresponding to the patch-level loss. The third function is used to compare the prediction result of each complete patch with the user-corrected label in a skeletonized structure, and the product of the third comparison result and the average confidence weight corresponding to the patch is used as the patch-level loss.

[0096] The third function measures the error of the model in maintaining the correctness of the overall topology of the ground features. This function can be the Hausdorff distance loss function or other functions. This embodiment does not limit the expression of this function.

[0097] In step 1053, firstly, for a training sample in the dynamic sample library, the complete patch region defined by the user-corrected label and the corresponding patch region predicted by the model are extracted respectively. Next, the two binary regions are skeletonized, that is, the planar region is eroded to its single-pixel-wide centerline structure using a morphological thinning algorithm to obtain the predicted skeleton and the true skeleton. Then, the two skeleton lines are compared, including calculating the overlap of the two skeleton lines, comparing the similarity of their main branch structures, or calculating the Hausdorff distance based on the skeleton. At the same time, the average confidence weight of all pixels in the entire patch region is calculated to obtain the patch-level average confidence weight. Finally, the calculated structural difference value is multiplied by the patch-level average confidence weight to obtain the weighted loss of the patch, and this loss is added to all training samples to obtain the patch-level loss.

[0098] Step 1054: Weight the first function, the second function, and the third function to obtain the training loss function.

[0099] In step 1054, pixel-level loss, boundary segment-level loss, and patch-level loss are taken as inputs and combined using a weighted summation method to form the final training loss function. Specifically, a preset weight coefficient is assigned to each of the pixel-level loss, boundary segment-level loss, and patch-level loss. These three weight coefficients are hyperparameters used to balance the influence of supervision signals at different scales on model training. For example, the weight of the boundary segment-level loss can be increased to strengthen the model's learning of boundary shapes. Then, the three losses are multiplied by their corresponding weights and summed to obtain the training loss function. During the training iteration, the optimizer calculates the gradient of this loss with respect to the model parameters through the backpropagation algorithm and updates the parameters to drive the model to simultaneously optimize pixel classification accuracy, boundary localization accuracy, and patch structure integrity.

[0100] This application implements multi-level and differentiated supervision of image segmentation models, and makes model training not only dependent on pixel-level classification accuracy, but also forced to focus on the accuracy of boundary geometry and overall regional structure. At the same time, by modulating the confidence weights, the training process can automatically focus on the key information provided by high confidence labels, thereby effectively resisting the noise introduced by low-quality corrections and further improving the segmentation accuracy and robustness of the model in complex regions such as boundaries.

[0101] In one specific implementation, after step S105, the method may further include the following steps: Step 106: In the model application stage, when the segmentation accuracy of the model in a specific application scenario is detected to be lower than the preset accuracy threshold, remote sensing image data in the specific application scenario is collected as incremental sample data.

[0102] In step 106, after the trained model is deployed to the actual production environment, its segmentation output quality in different scenarios is continuously monitored. Once it is found that the evaluation accuracy of the automatic segmentation result of the model is consistently lower than the performance threshold set for that scenario when processing data of a specific new scenario, an alarm is triggered. Then, the data collection process for the problematic scenario is started to collect new and representative remote sensing image data in that scenario as incremental sample data.

[0103] Step 107: Perform self-calibration on the incremental sample data. Based on the remote sensing image data and the actual segmentation results corresponding to the remote sensing image data in the self-calibrated incremental sample data, update some or all parameters of the image segmentation model through the backpropagation algorithm.

[0104] In step 107, the incremental image data undergoes automated or semi-automated correction processing to generate labels that can be used for supervised learning. This process is called self-correction. Self-correction may use the current model for initial segmentation, and then combine heuristic rules or minimal human intervention for correction and purification. After self-correction, a batch of incremental training samples containing image input and relatively reliable labels will be obtained. Subsequently, using this batch of new samples as training data, the backpropagation algorithm is used to fine-tune the parameters of the existing pre-trained model. The fine-tuning can choose to update only some layer parameters of the network to adapt to the specific features of the new scene.

[0105] It should be noted that the backpropagation algorithm is a well-known technology in this field, and its specific implementation process will not be described in detail in the embodiments of this application.

[0106] Step 108: With some parameters updated, the updated model parameters are fused with the original parameters to obtain the optimized image segmentation model.

[0107] In step 108, after incremental learning is completed using a partial parameter update strategy, the model will contain two sets of parameters: the original parameters that remain mostly unchanged and a small number of recently updated parameters. Then, in order to form a unified and stable new version model, the two sets of parameters are fused using a weighted average, that is, different fusion weights are assigned to the original parameters and the updated parameters, thereby obtaining the optimized image segmentation model.

[0108] This application integrates user interaction feedback and composite loss function to train the model, dynamically adjusts the learning focus based on comprehensive cognitive load score, and combines incremental optimization mechanism to continuously improve the model segmentation accuracy and adapt to new scenarios.

[0109] Figure 3This is a schematic diagram illustrating a specific implementation of a sample training and optimization system for an image segmentation model provided in this application. (Refer to...) Figure 3 The system may include: The acquisition module 31 is used to acquire the boundary correction information of the user based on the automatic segmentation results of remote sensing images, and simultaneously capture the user's interactive behavior sequence, as well as the physiological state signals and neural response characteristics of the user when performing the current boundary modification operation.

[0110] The fusion module 32 is used to fuse the interactive behavior sequence, the physiological state signal and the neural response features, and generate a comprehensive cognitive load score through the state decoding model.

[0111] The generation module 33 is used to generate a cognitive calibration factor for current boundary correction based on the user's professional profile and the comprehensive cognitive load score, and to map the cognitive calibration factor to the confidence weight of the boundary correction label.

[0112] The first construction module 34 is used to associate the target vector boundary, the corresponding confidence weight, the corresponding remote sensing image data, and the geographic information data and environmental parameter data of the remote sensing image in the automatic segmentation result of the remote sensing image to construct a dynamic sample library.

[0113] The second construction module 35 is used to construct a training loss function supervised by the boundary correction label weighted by the confidence weight during the training of the image segmentation model using the dynamic sample library, so as to optimize the segmentation accuracy of the image segmentation model in the boundary region.

[0114] This application provides a sample training optimization system for an image segmentation model to implement the aforementioned sample training optimization method for an image segmentation model. Therefore, the specific implementation of the sample training optimization system for an image segmentation model can be found in the embodiment section of the sample training optimization method for an image segmentation model described above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0115] like Figure 4 As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of the sample training optimization method for an image segmentation model described above.

[0116] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the sample training optimization method for an image segmentation model described above.

[0117] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0118] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the sample training optimization method for image segmentation models.

[0119] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0120] The foregoing has provided a detailed description of a sample training optimization method, system, and device for image segmentation models provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A sample training optimization method for image segmentation models, characterized in that, include: Acquire boundary correction information based on the user's automatic segmentation results of remote sensing images, simultaneously capture the user's interaction behavior sequence, as well as the user's physiological state signals and neural response characteristics when performing the current boundary modification operation; By integrating the interaction behavior sequence, the physiological state signals, and the neural response features, a comprehensive cognitive load score is generated through a state decoding model. Based on the user's professional profile and the comprehensive cognitive load score, a cognitive calibration factor for current boundary correction is generated, and the cognitive calibration factor is mapped to the confidence weight of the boundary correction label. The target vector boundary, corresponding confidence weight, corresponding remote sensing image data, and geographic information data and environmental parameter data of the remote sensing image in the automatic segmentation result of the remote sensing image are associated to construct a dynamic sample library. During the training of the image segmentation model using the dynamic sample library, a training loss function supervised by the boundary correction label weighted by the confidence weight is constructed to optimize the segmentation accuracy of the image segmentation model in the boundary region. The process of fusing the interaction behavior sequence, physiological state signals, and neural response features to generate a comprehensive cognitive load score through a state decoding model includes: The interaction behavior sequence is encoded to extract the behavior pattern feature vector, and the neural response features are encoded to extract the neural feature vector. Temporal analysis was performed on the eye-tracking data and mouse pressure change sequence in the physiological state signal to extract the corresponding cognitive features and obtain the cognitive feature vector; The behavioral pattern feature vector, the cognitive feature vector, and the neural feature vector are input into the spatiotemporal attention module to calculate the mutual information weights between different vectors in the time dimension, thereby obtaining the mutual information weight matrix. When the number of mutual information weights in the mutual information weight matrix that are less than a preset weight threshold exceeds a preset proportion, a cognitive inconsistency score is generated and output. The behavioral pattern feature vector, the cognitive feature vector, the neural feature vector, and the cognitive inconsistency score are input into the state decoding model. Through the attention-guided feature fusion layer in the state decoding model, the weight ratio of the behavioral pattern feature vector, the cognitive feature vector, and the neural feature vector in the fusion process is dynamically adjusted according to the cognitive inconsistency score. The fused features are then decoded to output a five-dimensional vector, which corresponds to five independent dimensions: attention concentration, cognitive load, decision certainty, cognitive consistency, and neural response stability. The five-dimensional vector is mapped to a comprehensive cognitive load score through a nonlinear mapping layer in the state decoding model.

2. The method according to claim 1, characterized in that, Through the attention-guided feature fusion layer in the state decoding model, based on the cognitive inconsistency score, the weight ratios of the behavioral pattern feature vector, cognitive feature vector, and neural feature vector in the fusion process are dynamically adjusted, including: The behavioral pattern feature vector, cognitive feature vector, and neural feature vector are projected onto a unified latent space to obtain the corresponding first latent feature, second latent feature, and third latent feature. Based on cognitive inconsistency scoring, multi-level attention adjustment is performed on different vectors to obtain the fusion weights of different vectors, and the corresponding latent features are weighted. The weighted first, second, and third latent features are input into a gating fusion unit. The gating fusion unit dynamically controls the contribution ratio of each latent feature in the fused features based on the predicted initial values ​​of the cognitive consistency dimension and the neural response stability dimension, so as to generate the fused features.

3. The method according to claim 2, characterized in that, Based on cognitive inconsistency scores, multi-level attention adjustments are performed on different vectors to obtain fusion weights for the different vectors, including: The first-level weights are determined based on the information entropy of each vector. Based on the cognitive inconsistency score, a decay factor is calculated for each vector. The decay factor is used to characterize the credibility of each vector when there is cognitive inconsistency. The decay factor corresponding to the behavioral pattern feature vector is negatively correlated with the cognitive inconsistency score, and the decay factor corresponding to the neural feature vector is positively correlated with the cognitive inconsistency score. Multiply the first-level weight by the corresponding modal confidence decay factor to obtain the second-level weight; The third-level weights are calculated based on the mutual information values ​​between vectors, and the second-level weights are weighted and summed with the third-level weights to obtain the fusion weights of different vectors.

4. The method according to claim 1, characterized in that, Based on the user's professional profile and the comprehensive cognitive load score, a cognitive calibration factor for current boundary correction is generated. This cognitive calibration factor is then mapped to the confidence weights of the boundary correction label, including: Based on the user's historical boundary correction records, construct a professional user profile; Based on the aforementioned professional profile and the aforementioned comprehensive cognitive load score, preliminary calibration factors are generated; The initial calibration factor is adjusted to obtain the cognitive calibration factor; Based on the cognitive calibration factor, the confidence weight is calculated using a preset mapping function.

5. The method according to claim 1, characterized in that, Constructing a training loss function supervised by the boundary-corrected labels weighted by the confidence scores includes: A first function corresponding to pixel-level loss is constructed. The first function is used to compare the prediction result of each pixel with the user-corrected label, and the product of the first comparison result and the confidence weight corresponding to the pixel is used as the pixel-level loss. Construct a second function corresponding to the boundary segment level loss. The second function is used to compare the prediction result on each continuous boundary with the user-corrected label, and the product of the second comparison result and the average confidence weight corresponding to the boundary segment is used as the boundary segment level loss. A third function corresponding to the patch-level loss is constructed. The third function is used to compare the prediction result of each complete patch with the user-corrected label in a skeletonized structure, and the product of the third comparison result and the average confidence weight corresponding to the patch is used as the patch-level loss. The first function, the second function, and the third function are weighted and combined to obtain the training loss function.

6. The method according to claim 1, characterized in that, After constructing the training loss function supervised by the boundary-corrected labels weighted by the confidence weights, the method further includes: During the model application phase, when the segmentation accuracy of the model in a specific application scenario is detected to be lower than the preset accuracy threshold, remote sensing image data in the specific application scenario is collected as incremental sample data. The incremental sample data is self-calibrated. Based on the remote sensing image data and the actual segmentation results corresponding to the remote sensing image data in the self-calibrated incremental sample data, some or all parameters of the image segmentation model are updated through the backpropagation algorithm. With some parameters updated, the updated model parameters are fused with the original parameters to obtain an optimized image segmentation model.

7. The method according to claim 1, characterized in that, After constructing a dynamic sample library by associating the target vector boundaries, corresponding confidence weights, corresponding remote sensing image data, and geographic information and environmental parameter data of the remote sensing images in the automatic segmentation results of the remote sensing images, the following steps are also included: Spatial overlap analysis and optimization are performed on the samples in the dynamic sample library; Based on the type and scale of the land cover corresponding to the sample, the spatial sampling strategy is dynamically adjusted to control the overlap between samples within a preset range; among them, the first overlap threshold of strip-shaped land cover is less than the second overlap threshold of areal clustered land cover.

8. A sample training and optimization system for an image segmentation model, characterized in that, include: The acquisition module is used to acquire the boundary correction information of the user based on the automatic segmentation results of remote sensing images, and simultaneously capture the user's interaction behavior sequence, as well as the physiological state signals and neural response characteristics of the user when performing the current boundary modification operation. The fusion module is used to fuse the interactive behavior sequence, the physiological state signal, and the neural response features, and generate a comprehensive cognitive load score through a state decoding model. Specifically, the fusion module encodes the interactive behavior sequence to extract behavioral pattern feature vectors, and simultaneously encodes the neural response features to extract neural feature vectors. It performs temporal analysis on the eye-tracking data and mouse pressure change sequence in the physiological state signal to extract corresponding cognitive features, obtaining cognitive feature vectors. The behavioral pattern feature vectors, cognitive feature vectors, and neural feature vectors are input into a spatiotemporal attention module to calculate the mutual information weights between different vectors in the time dimension, obtaining a mutual information weight matrix. When the mutual information weight matrix contains values ​​less than a preset weight threshold... When the number of mutual information weights exceeds a preset ratio, a cognitive inconsistency score is generated and output. The behavioral pattern feature vector, the cognitive feature vector, the neural feature vector, and the cognitive inconsistency score are input into the state decoding model. Through the attention-guided feature fusion layer in the state decoding model, the weight ratios of the behavioral pattern feature vector, the cognitive feature vector, and the neural feature vector in the fusion process are dynamically adjusted according to the cognitive inconsistency score. The fused features are then decoded to output a five-dimensional vector, corresponding to five independent dimensions: attention concentration, cognitive load, decision certainty, cognitive consistency, and neural response stability. Through the nonlinear mapping layer in the state decoding model, the five-dimensional vector is mapped to a comprehensive cognitive load score. The generation module is used to generate a cognitive calibration factor for current boundary correction based on the user's professional profile and the comprehensive cognitive load score, and to map the cognitive calibration factor to the confidence weight of the boundary correction label. The first construction module is used to associate the target vector boundary, the corresponding confidence weight, the corresponding remote sensing image data, and the geographic information data and environmental parameter data of the remote sensing image in the automatic segmentation result of the remote sensing image to construct a dynamic sample library. The second construction module is used to construct a training loss function supervised by the boundary correction label weighted by the confidence weight during the training of the image segmentation model using the dynamic sample library, so as to optimize the segmentation accuracy of the image segmentation model in the boundary region.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of a sample training optimization method for an image segmentation model as described in any one of claims 1 to 7 when executing the computer program.

Images