A flow view classification method and apparatus, a terminal device, and a medium

Abstract

The application provides a stream view classification method and device, terminal equipment and medium, the method comprises the following steps: receiving stream view data arriving in a streaming order; inputting a data view arriving at a current time into a pre-constructed evidence deep neural network model to generate a current evidence vector through a non-negative activation transformation; performing linear transformation on the current evidence vector by using a preset label transition matrix to obtain a noise-adaptive evidence vector; calculating a group of Dirichlet distribution parameters based on the noise-adaptive evidence vector, and generating a subjective logical opinion of the data view according to the Dirichlet distribution parameters; synthesizing the subjective logical opinion and a comprehensive opinion output at a previous time through a fusion operator defined based on the subjective logic to obtain a new comprehensive opinion; and determining a final classification identifier of an entity according to a belief quality vector in the new comprehensive opinion. The application can improve the accuracy of stream view classification.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a streaming view classification method, apparatus, terminal device, and medium. Background Technology

[0002] In technological research, sensor data (such as time-series image sequences, layered scan slices, dynamic high-resolution scan images, etc.) are typically generated and stored in the form of continuous streams. For example, a dynamic monitoring process can generate a sequence of multiple images of the same target object at different points in time; a panoramic scan can generate multiple local high-resolution images transmitted in a streaming manner.

[0003] With the development of sensing devices, observational data are characterized by high dimensionality and temporal sequence. For example, dynamic augmentation observation generates sequential images of the same target at different time points; layered scanning imaging can produce continuous multi-layer slices; and full-frame scanning systems transmit high-resolution local images of the entire sample in a streaming manner. These multi-angle data, arriving in temporal or spatial order and belonging to the same observation target (such as a specific region), constitute the "stream view" problem in analysis.

[0004] Accurate and robust automated classification of this type of flow view data is a key technology for building intelligent analytics systems. However, existing solutions face several challenges:

[0005] Firstly, data labeling is highly dependent on the experience of professionals, and subjective differences, accidental oversights, and disagreements among experts can lead to significant label noise in the training data. Traditional deep learning-based classification models (which typically use the cross-entropy loss function) are prone to overfitting to noisy labels, thus impairing their generalization performance and practical usability on clean data.

[0006] Secondly, existing methods for processing multi-view data often employ strategies such as majority voting, probability averaging, or simply stitching together features from all views for classification. These methods fail to explicitly model the quality differences and complementary information relationships between different views, and cannot effectively address the issues of missing or unreliable view information caused by acquisition artifacts, local interference, or occlusion. Simple fusion methods are also easily affected by low-quality views.

[0007] Thirdly, mainstream classification models typically output a normalized class probability distribution, which combines the model's data-driven "cognitive uncertainty" with the problem's inherent "random uncertainty," but cannot explicitly separate them. In decision-making scenarios, understanding the model's confidence level (cognitive uncertainty) in a particular prediction is crucial. When faced with poor data quality, atypical features, or incomplete training data, high-probability mispredictions are more harmful than low-probability mispredictions. Existing models lack an effective mechanism to quantify and output this "cognitive uncertainty."

[0008] Fourthly, data streams often contain important temporal evolution or spatial correlation information. Existing methods typically treat streaming views as independent and identically distributed samples, ignoring the dynamic change patterns between views and failing to fully utilize sequence context information to improve the accuracy and consistency of classification.

[0009] In summary, there is an urgent need for a streaming view classification method that can simultaneously handle annotation noise, effectively integrate multi-view information, explicitly quantify decision uncertainty, and adapt to the characteristics of streaming data. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to provide a flow view classification method, apparatus, terminal equipment and medium to improve the accuracy of flow view classification.

[0011] In a first aspect, the present invention provides a flow view classification method, the method comprising the following steps:

[0012] Receive streaming view data arriving in a streaming order; the streaming view data includes multiple time-series data views corresponding to the same entity; the streaming view data includes one or more of image, audio, and video data;

[0013] The data view arriving at the current moment is input into a pre-built evidence deep neural network model, and a current evidence vector is generated through a non-negative activation transformation; the current evidence vector contains a set of non-negative values, where each value independently represents the strength of the data view supporting the entity's belonging to a target category;

[0014] A linear transformation is performed on the current evidence vector using a pre-defined label transition matrix to obtain a noise-adapted evidence vector; wherein, the label transition matrix is ​​used to characterize the conditional probability relationship between the true label of the target category and the noisy observation label;

[0015] A set of Dirichlet distribution parameters is calculated based on the noise-adapted evidence vector, and subjective logical opinions of the data view are generated based on the Dirichlet distribution parameters. The subjective logical opinions include a belief quality vector and an uncertainty parameter. The belief quality vector represents the support for each target category, and the uncertainty parameter represents the overall uncertainty of the classification decision due to missing view information or noise.

[0016] A fusion operator based on subjective logic is used to synthesize subjective logical opinions with the comprehensive opinions output at the previous time step to obtain new comprehensive opinions. The synthesis process weights and reduces the view information according to the uncertainty parameters in each opinion to achieve effective accumulation of cross-view information under the uncertainty framework. The new comprehensive opinions are iteratively passed in streaming processing.

[0017] The final classification identifier for an entity is determined based on the belief quality vector in the new synthesized opinion.

[0018] Optionally, the output layer of the evidence deep neural network model is constructed to directly output the current evidence vector, and the activation function of the evidence deep neural network model is configured to output only non-negative values, thereby avoiding the competition and forced normalization constraints between different categories of evidence values ​​caused by using the Softmax function.

[0019] The nonnegative activation transformation uses the Softplus function or the ReLU function to ensure that the generated evidence vector physically possesses nonnegative counting properties.

[0020] Optionally, a linear transformation is performed on the current evidence vector using a preset label transition matrix to obtain a noise-adapted evidence vector, including:

[0021] Set the label transition matrix as ;in, The elements in the matrix represent the total number of target categories. The true category of a medical entity is At that time, its initial observation label was incorrectly labeled as a category. The estimated probability;

[0022] Through calculation formula The noise adaptation evidence vector is obtained. ;in, Represents the label transition matrix transpose, This represents the current evidence vector.

[0023] Optionally, a set of Dirichlet distribution parameters is calculated based on the noise-adapted evidence vector, and a subjective logical opinion of the data view is generated based on the Dirichlet distribution parameters, specifically including:

[0024] For each entity, adapt the evidence vector based on the entity's noise. Calculate the Dirichlet distribution parameters ;in, , , , ;

[0025] Through calculation formula To obtain the total strength of evidence ;

[0026] Through calculation formula The belief quality vector is obtained. ;in, , This indicates that the entity belongs to the target category. Reliability quality;

[0027] Through calculation formula To obtain the uncertainty parameters ;

[0028] Based on the belief quality vector and uncertainty parameters Constructing subjective logical opinions ;in, .

[0029] Optionally, a new comprehensive opinion is obtained by synthesizing the subjective logical opinion with the comprehensive opinion output at the previous time step using a fusion operator based on the subjective logic definition, including:

[0030] Let the overall opinion at the previous moment be... The subjective logical opinion at the current moment is ;

[0031] For any target category Through the calculation formula:

[0032]

[0033] Gain new belief quality ;in, The conflict factor is used to measure the degree of contradiction between historical consensus and current viewpoints.

[0034] Calculation formula:

[0035]

[0036] New uncertainty parameters are obtained New comprehensive opinions , .

[0037] Optionally, the preset label transition matrix is ​​learned through the model training process, and the total loss function of this training process consists of at least the following parts:

[0038] Noise-robust cross-entropy loss term This is used to minimize the difference between the expected prediction made by the model based on the noise-adapted evidence vector and the observed noisy label, thereby driving the model to learn to adapt to noise.

[0039] Uncertainty regularization loss term This is used to constrain the cognitive uncertainty of model predictions by minimizing the KL divergence between the predicted Dirichlet distribution and a prior uniform distribution, thus preventing the model from making overconfident and erroneous predictions on noisy data.

[0040] Transition matrix regularization loss term This is used to constrain the structure of the label transition matrix. By minimizing the volume of the simplex spanned by its column vectors, the learned transition matrix is ​​made to have clear probabilistic semantics and good numerical stability.

[0041] Optionally, during streaming processing, a time consistency constraint is also imposed on the tag transition matrix;

[0042] Imposing time consistency constraints on the label transition matrix includes:

[0043] Calculation formula:

[0044]

[0045] Obtain the label transition matrix learned at the current time step. Label transition matrix from the previous time step The difference norm between ;

[0046] By minimizing the difference norm This ensures that the model's estimation of the label noise pattern remains stable throughout the continuous view stream, avoiding incorrect estimations of the noise distribution due to data fluctuations in a single view.

[0047] In a second aspect, the present invention provides a flow view classification device, comprising:

[0048] The data acquisition module is used to receive streaming view data arriving in a streaming order; the streaming view data includes multiple time-series data views corresponding to the same entity; the streaming view data includes one or more of image, audio, and video data;

[0049] The evidence generation module is used to input the data view that arrives at the current time into a pre-built evidence deep neural network model and generate the current evidence vector through non-negative activation transformation. The current evidence vector contains a set of non-negative values, where each value independently represents the strength of the data view supporting the entity to belong to a target category.

[0050] The evidence optimization module is used to perform a linear transformation on the current evidence vector using a preset label transition matrix to obtain a noise-adapted evidence vector; wherein, the label transition matrix is ​​used to characterize the conditional probability relationship between the true label of the target category and the noisy observation label;

[0051] The opinion generation module is used to calculate a set of Dirichlet distribution parameters based on the noise-adapted evidence vector, and generate subjective logical opinions for the data view based on the Dirichlet distribution parameters. The subjective logical opinions include a belief quality vector and an uncertainty parameter. The belief quality vector represents the support for each target category, and the uncertainty parameter represents the overall uncertainty of the classification decision due to missing view information or noise.

[0052] The opinion fusion module is used to synthesize subjective logical opinions with the comprehensive opinions output at the previous time step using a fusion operator based on subjective logic definition to obtain new comprehensive opinions. The synthesis process weights and reduces the view information according to the uncertainty parameters in each opinion to achieve effective accumulation of cross-view information under the uncertainty framework. The new comprehensive opinions are iteratively passed in streaming processing.

[0053] The view classification module is used to determine the final classification identifier of an entity based on the belief quality vector in the new comprehensive opinion.

[0054] Thirdly, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0055] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0056] The present invention has at least the following beneficial effects:

[0057] By introducing a learnable label transition matrix to explicitly model the labeling noise process and using it throughout forward inference and backward training, the model can learn the true category feature distribution from noisy labeled data, significantly improving classification stability in complex real-world labeling environments and enhancing the accuracy of stream view classification. A fusion operator based on subjective logic dynamically adjusts the weight of each view opinion when synthesizing it with historical information based on its own uncertainty parameter. Opinions with high uncertainty automatically reduce their contribution to the final decision, achieving intelligent weighted fusion of multi-source heterogeneous information, enhancing the overall robustness of the system, and improving the accuracy of stream view classification. The model output includes not only a classification label but also a quantified uncertainty parameter, clearly distinguishing between "high-confidence decisions based on sufficient evidence" and "high-probability outputs exhibited by the model due to sufficient training." This provides crucial confidence evidence for subsequent manual review or integration with other systems, increasing the system's security and credibility. Attached Figure Description

[0058] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0059] Figure 1 This is a flowchart of a flow view classification method in one embodiment of this application;

[0060] Figure 2 This is one of the comparison diagrams showing the implementation results of the flow view classification method provided by the present invention and seven baseline methods in another embodiment of this application;

[0061] Figure 3 This is the second comparison chart of the implementation results of the flow view classification method and the seven baseline methods in another embodiment of this application;

[0062] Figure 4 This is a comparison chart of the implementation results of the flow view classification method and the multi-view method TMNR in another embodiment of this application; wherein, Figure 4 (a) is a comparison chart of implementation results on the Handwritten dataset. Figure 4 (b) A comparison chart of implementation results on the FlickrAnimal dataset. Figure 4 (c) is a comparison chart of implementation results on the NH dataset; the vertical axis ACC represents the accuracy, and the horizontal axis Noise Ratio represents the noise ratio;

[0063] Figure 5 This is the third comparison chart showing the implementation results of the flow view classification method and the seven baseline methods in another embodiment of this application;

[0064] Figure 6 This is a schematic diagram of the flow view classification device in one embodiment of this application;

[0065] Figure 7 This is a schematic diagram of the structure of a terminal device in one embodiment of this application. Detailed Implementation

[0066] The technical solution of the present invention will now be described in detail and completely with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0067] In the description of this invention, it should be noted that the terms "upper", "lower", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0068] Example 1

[0069] like Figure 1 As shown, the flow view classification method provided by the present invention includes steps 11 to 16.

[0070] Step 11: Receive the streaming view data that arrives in streaming order.

[0071] The streaming view data includes multiple time-series arriving data views corresponding to the same entity; the streaming view data includes one or more of image, audio, and video data. Specifically, the... The features of each view are represented as follows: ,in Corresponding to the The first sample One view, Indicates the number of samples. It is the first The feature dimensions of each view.

[0072] In one feasible implementation, the flow view data includes multiple time-arriving inspection images (data views) corresponding to the same target part (entity). The data comes from multiple high-resolution industrial cameras on the production line, arranged in the order of part delivery, capturing images of the same part from different angles (e.g., top, side, bottom). The goal is to classify the part appearance into several predefined categories (e.g., Category 1: qualified; Category 2: scratch defects; Category 3: stain defects; Category 4: deformation defects). This classification information provides a basis for decision-making in automated sorting systems, enabling automated screening of product quality.

[0073] In another feasible implementation, the flow chart data can also be a multi-phase sequence of dynamic contrast-enhanced MRI (DCE-MRI) of the liver. The goal is to classify the imaging features into several predefined imaging pattern categories (e.g., category 1: marked enhancement in the arterial phase; category 2: persistent enhancement in the portal venous phase; category 3: clearance in the delayed phase). This imaging pattern classification information can provide important reference for physicians to determine the nature of lesions. It should be noted that the method provided by this invention cannot be used to directly diagnose diseases in pathological medicine; it only provides auxiliary analytical reference for professional medical personnel. The final diagnosis of the disease should be made by qualified medical personnel based on clinical manifestations, multiple examination results, and the auxiliary information provided by this invention.

[0074] In actual implementation, for each view ( Standardization preprocessing is performed: spatial resolution is uniformly sampled to... Pixels; pixel intensity values ​​are normalized to the [0,1] range after window width and window level adjustments. The preprocessed view data is then stored in a first-in-first-out (FIFO) buffer to ensure sequential processing.

[0075] Step 12: Input the data view that arrives at the current time into the pre-built evidence deep neural network model, and generate the current evidence vector through non-negative activation transformation.

[0076] In this embodiment of the invention, the current evidence vector contains a set of non-negative values, where each value independently represents the strength of the data view supporting the current entity belonging to a target category.

[0077] In one feasible implementation, the backbone feature extractor of the evidence deep neural network model can adopt a standard convolutional neural network architecture (e.g., ResNet-50). It should be noted that, unlike traditional methods, the flow view classification method provided in this invention constructs the output layer of the evidence deep neural network model to directly output the current evidence vector. The activation function of this evidence deep neural network model is configured to output only non-negative values, thereby avoiding the competition and forced normalization constraints between different categories of evidence values ​​caused by using the Softmax function. Specifically, the output layer of the evidence deep neural network model removes the Softmax classification layer and replaces it with a feature extractor that... indivual( A fully connected layer is formed from the output neurons (corresponding to the number of categories), and a non-negative activation function (Softplus or ReLU) is applied after this fully connected layer to ensure that the generated evidence vector physically possesses non-negative counting properties. Taking the Softplus function as an example, its expression is: . This indicates the total number of target categories.

[0078] In actual implementation, the current moment view Inputting this into the network yields the original evidence vector. ,in, , Each This can be intuitively understood as the current view being a category. The amount of supporting evidence provided. In one possible implementation, for a given arterial phase view... The network output may be This indicates that the features of this view strongly support the imaging pattern of "significant enhancement in the arterial phase" (Category 1), while providing weak support for "persistent enhancement in the portal venous phase" (Category 2).

[0079] It should be noted that, unlike Softmax which forces the output to be normalized to a probability and 1, evidence vectors allow the model to express situations where "evidence in all categories is weak" (image is blurry) or "evidence in multiple categories is strong" (image features are between multiple categories). This provides a direct input basis for subsequent quantification of uncertainty and is a key prerequisite for achieving uncertainty perception.

[0080] Step 13: Perform a linear transformation on the current evidence vector using a preset label transition matrix to obtain a noise-adapted evidence vector.

[0081] To overcome the influence of annotation noise in the training data, this invention introduces a learnable label transition matrix to characterize the conditional probability relationship between the true label of the target category and the noisy observed label. Specifically, the label transition matrix is ​​as follows: The elements in the matrix. The actual category of the entity is Its initial observation labels were incorrectly labeled as categories. The estimated probability, .

[0082] In one feasible implementation, through calculation formula The noise adaptation evidence vector is obtained. ;in, Represents the label transition matrix transpose, This represents the current evidence vector. This operation can be understood as correcting the currently observed evidence based on noise confusion patterns learned from history.

[0083] For example, assuming that the learned The matrix is:

[0084]

[0085] This matrix shows that when the true pattern is category 1, there is a 90% probability of it being correctly labeled, a 7% probability of it being mislabeled as category 2, and a 3% probability of it being mislabeled as category 3. If... ,but As can be seen, due to The matrix model modeled the possible mislabeling between categories, and the corrected evidence distribution was slightly adjusted to be closer to the evidence distribution that should be observed under the hypothesis of "clean" labeling.

[0086] This step elevates noise modeling from the loss function level to the data representation level. By performing explicit transformations in the feature (evidence) space, the model acquires "denoising" capabilities during forward inference. This is more direct and effective than simply using a robust loss function, and it integrates naturally with the evidence theory framework.

[0087] It should be noted that, in this embodiment of the invention, the preset label transition matrix is ​​learned through the model training process, and the training objective is to jointly optimize the parameters of the evidence neural network and the label transition matrix end-to-end.

[0088] In one feasible implementation, the expression for the total loss function is:

[0089]

[0090]

[0091] in, This represents the total loss value. This represents the noise-robust cross-entropy loss term, used to minimize the difference between the expected prediction made by the model based on the noise-adapted evidence vector and the observed noisy label, driving the model to learn to adapt to noise. The uncertainty regularization loss term is used to constrain the cognitive uncertainty of the model's predictions. By minimizing the KL divergence between the predicted Dirichlet distribution and a prior uniform distribution, it prevents the model from making overconfident and erroneous predictions on noisy data. The regularization loss term for the transition matrix is ​​used to constrain the structure of the label transition matrix. By minimizing the volume of the simplex spanned by its column vectors, the learned transition matrix is ​​made to have clear probabilistic semantics and good numerical stability. This represents a time consistency constraint. This represents the label transition matrix learned at the current time step. This represents the label transition matrix at the previous time step. It is achieved by minimizing the difference norm. This ensures that the model's estimation of the label noise pattern remains stable throughout the continuous view stream, avoiding incorrect estimations of the noise distribution due to data fluctuations in a single view.

[0092] It should be noted that when facing complex technical problems such as label noise, decision uncertainty quantification, and multi-view fusion in streaming view classification, the conventional approach for those skilled in the art is usually a divide-and-conquer strategy: for example, using label smoothing or noise-robust loss functions to mitigate the impact of noise; adding a probability calibration layer after the model or training a confidence network separately to estimate uncertainty; designing manually weighted rules (such as based on image quality scoring) or multi-branch networks for view fusion. These approaches decompose the problem into independent sub-modules for sequential optimization, often neglecting the inherent coupling relationship between noise, uncertainty, and fusion. The inventiveness of this invention lies in constructing an end-to-end unified training framework, the core difference of which is: introducing a learnable label transition matrix and embedding it into the forward computation graph, enabling deep fusion of the noise adaptation process and evidence generation; directly mapping evidence to endogenous subjective logical opinions containing uncertainty components through Dirichlet distribution parameterization, and explicitly constraining and calibrating this uncertainty with KL divergence regularization; finally, these well-calibrated opinions with built-in uncertainty measures are stream-synthesized through uncertainty-based fusion operators, with the fusion rules themselves being theory-driven rather than heuristically designed. The better effect of this integrated design is that the model is jointly optimized from the training stage to solve all problems in a collaborative manner: noise adaptation improves the purity of evidence, uncertainty regularization ensures the reliability of the evidence-to-opinion mapping, thereby enabling streaming fusion to dynamically adjust weights based on highly credible, quantified uncertainty, achieving simultaneous improvement in three dimensions: noise robustness, decision interpretability, and multi-view fusion robustness, resulting in performance gains and system synergy effects far exceeding those of conventional serial schemes.

[0093] Step 14: Calculate a set of Dirichlet distribution parameters based on the noise-adapted evidence vector, and generate subjective logical opinions for the data view based on the Dirichlet distribution parameters.

[0094] In this embodiment of the invention, the subjective logical opinion includes a belief quality vector and an uncertainty parameter. The belief quality vector represents the support for each target category, while the uncertainty parameter represents the overall uncertainty regarding the classification decision due to missing view information or noise.

[0095] Specifically, step 14 includes steps 14.1 to 14.5.

[0096] Step 14.1: For each entity, adapt the evidence vector based on the entity's noise. Calculate the Dirichlet distribution parameters .

[0097] in, , , , .

[0098] Step 14.2, using the calculation formula To obtain the total strength of evidence .

[0099] Step 14.3, using the calculation formula The belief quality vector is obtained. .

[0100] in, , This indicates that the entity belongs to the target category. The reliability quality.

[0101] Step 14.4, using the calculation formula To obtain the uncertainty parameters .

[0102] Step 14.5, based on the belief quality vector and uncertainty parameters Constructing subjective logical opinions .

[0103] in, For the current moment , .

[0104] In one feasible implementation method, This can be interpreted as follows: Based on the current view, the confidence level (belief) for classifying the image of this area as belonging to category 1 is 6.1%, to category 2 is 59.0%, and to category 3 is 11.4%. However, due to potential limitations in the view information itself (such as noise or partial volumetric effects), the model holds an overall uncertainty of 23.5% regarding the classification decision. The sum of beliefs (0.765) and uncertainties (0.235) equals 1.

[0105] This step maps evidence to a cognitive state. Uncertainty Explicit quantification is key to this invention. It directly depends on the total strength of evidence. When the view is clear and the features are well-defined, the evidentiary value is high. big, A small evidence value indicates strong confidence in the decision-making process. Conversely, a large evidence value indicates a weak evidence value when the view is unclear or the features are contradictory. Small, A large value indicates that the model is "aware" of its low confidence in making judgments. This provides a basis for the next step of intelligent integration.

[0106] Step 15: Using a fusion operator based on subjective logic definition, the subjective logical opinion is synthesized with the comprehensive opinion output at the previous time step to obtain a new comprehensive opinion.

[0107] In this embodiment of the invention, the synthesis process weights and reduces information based on uncertainty parameters in each opinion, thereby achieving effective accumulation of cross-view information within an uncertainty framework. New synthesized opinions are iteratively transmitted in streaming processing.

[0108] Specifically, the process of synthesizing the subjective logical opinion with the comprehensive opinion output at the previous time step by using a fusion operator based on the subjective logic definition to obtain a new comprehensive opinion includes steps 15.1 to 15.3.

[0109] Step 15.1, set the overall opinion from the previous moment as... The subjective logical opinion at the current moment is .

[0110] Step 15.2, for any target category Through the calculation formula:

[0111]

[0112] Gain new belief quality .in, The conflict factor is used to measure the degree of contradiction between historical consensus and current viewpoints.

[0113] Step 15.3, using the calculation formula:

[0114]

[0115] New uncertainty parameters are obtained New comprehensive opinions , .

[0116] In one feasible implementation, it is assumed that after processing the first two data views, the overall opinion is as follows: This indicates that the current overall judgment leans towards category 2, with low uncertainty. The opinion of the current third data view is... The opinion itself is highly uncertain, and there is a relatively higher belief in category 3. The calculated conflict factor... After integration, the new comprehensive opinions In the middle, the belief in category 2 may be slightly lowered from 0.70, while the belief in category 3 may be higher from 0.10, while overall uncertainty... The increase from 0.15 may be slight, reflecting the impact of newly introduced uncertain information.

[0117] It is worth mentioning that uncertainty ( In this invention, it acts as a natural weight. The more certain the historical opinion, the more ( The smaller the value of a viewpoint, the higher its weight in the fusion process; the more uncertain the current viewpoint, the greater its weight in the fusion process. The higher the quality and certainty of a view, the lower its belief weight. This achieves the following: 1) High-quality, high-certainty views dominate the fusion direction; 2) Low-quality, high-uncertainty views are automatically reduced to avoid interfering with established reliable judgments; 3) When new and old opinions conflict, the uncertainty after fusion will reasonably increase, reflecting the increased difficulty of decision-making. This dynamic weighting mechanism based on uncertainty is the core of achieving robust multi-view fusion.

[0118] Step 16: Determine the final classification identifier of the entity based on the belief quality vector in the new comprehensive opinion.

[0119] Specifically, when all After processing each data view, a final comprehensive opinion is obtained. The final classification label output consists of two parts: the classification label and the confidence score. The classification label specifies the category. This indicates the category with the highest belief quality. The confidence assessment results are expressed in terms of uncertainty. Or the confidence score derived from it (such as...) Output in the form of ).

[0120] In one feasible implementation, the final classification identifier is: "Imaging pattern classification result: persistent enhancement in the portal venous phase; classification confidence: 85%". This output, as structured information, can be integrated into the imaging reporting system or directly displayed to doctors as an auxiliary reference for their image description and judgment.

[0121] Example 2

[0122] To verify the model's advancement, the streaming view classification method (RSCLN) provided in this invention was compared with seven baseline methods in this embodiment, including non-robust and noisy robust methods: (1) Non-robust methods: SVL is a streaming view classification method, ABAD is a single-view classification method, and AWDRM and MCCD are multi-view classification methods. (2) Noisy robust methods: Cycle-T is a single-view label noise classification method, and LNRMC and TMNR are multi-view label noise methods. All comparison methods were configured using the recommended parameters. Given the single-pass nature of the streaming view scenario, the single-view classification method was evaluated based on its performance on the final view. For the multi-view method, to maintain consistency with the streaming view setup, it was assumed that the inherited information was data from the previous view, and therefore its results on the last two views were reported.

[0123] In practice, the various methods were evaluated on seven datasets, and the final results were expressed in terms of accuracy (ACC) and F1 (mean). The variance value is displayed, such as Figure 2 and Figure 3 Best performance is highlighted in bold. The NH dataset contains three views with dimensions of 2000, 3304, and 6750, containing 550 samples across 5 categories. The Handwritten dataset consists of six views (dimensions of 240, 6, 47, 64, 76, and 216), representing 2000 handwritten digit samples across 10 categories. The FlickrAnimal dataset contains 3411 samples from 13 animal categories, each represented by 6 feature views with dimensions of 64, 500, 144, 73, 128, and 225. The Forest dataset is an image dataset containing a 7-category subset of Caltech101, with 441 samples and 6 views of varying dimensions (1302, 48, 512, 100, 256, and 256). The Wiki dataset is a two-view dataset consisting of image-text pairs, containing 2866 samples across 10 categories. The Sensit dataset contains 3,000 samples across 3 classes, with two views of equal dimensions (50 features each).

[0124] Depend on Figure 2 , Figure 3It can be known that:

[0125] (1) Although the performance of almost all models gradually decreases with increasing noise, the streaming view classification method (RSCLN) provided by this invention can still achieve optimal performance in most cases.

[0126] (2) Due to the lack of noise handling mechanisms, non-robust methods lag behind noisy robust methods in most cases. Although SVL is designed for streaming views, it has two main limitations: sensitivity to label noise and reliance on sufficient views. These two factors lead to performance degradation in real-world scenarios. In particular, on datasets such as Wiki and Sensit, SVL struggles to train effective base models or even fails, which may be attributed to the insufficient number of views in these datasets. On some datasets, noisy robust methods exhibit performance comparable to RSCLN, possibly because subsequent arriving views contain a large amount of valuable information, making the lack of some historical view information less of an issue.

[0127] (3) The multi-view-based approach is superior to the single-view-based approach in most cases, which may be because two views provide more information than a single view. This fully demonstrates the positive role of storing view data in model building, and also indirectly shows that the inheritance strategy we designed in the method achieves an effect comparable to directly storing view data.

[0128] To further verify the outstanding performance of the streaming view classification method (RSCLN method) provided by this invention in preserving effective information of historical views, in this embodiment of the invention, the streaming view method RSCLN, which stores less information, is directly compared with the multi-view method TMNR, which directly stores all view data (it is known that both TMNR and RSCLN adopt a classification strategy based on uncertainty). The implementation comparison results for different datasets are as follows: Figure 4 (a) Figure 4 (b) Figure 4 As shown in (c), the shaded area represents the standard deviation range, and the horizontal axis, Noise Ratio, represents the noise ratio. Figure 4 (a) Figure 4 (b) Figure 4(c) It can be seen that, regardless of the streaming view method or the multi-view method, the performance tends to decrease as the noise ratio increases. Furthermore, it is evident that RSCLN and TMNR have a large overlap in most cases, and their performance is comparable. It can be considered that RSCLN achieves a performance comparable to the multi-view method, fully demonstrating that the streaming view classification method provided by this invention does indeed retain effective information from historical views. Even without retaining the original data, it can achieve a performance comparable to the multi-view method through effective inheritance and update strategies. It is worth noting that although both methods employ the same uncertainty-based classification strategy, RSCLN requires significantly less storage data to achieve this goal.

[0129] The results above show that the performance of both the streaming view method and the multi-view method decreases with increasing noise levels, which is consistent with objective facts. Furthermore, it is evident that in most cases, there is significant overlap between the RSCLN method and TMNR, indicating that their performance is comparable. This demonstrates that the RSCLN method achieves performance on par with the multi-view method, fully proving that the streaming view classification method provided by this invention effectively preserves valuable information from historical views. Even without retaining the original data, it can achieve performance comparable to the multi-view method through effective inheritance and update strategies.

[0130] Example 3

[0131] To further verify the effectiveness of the streaming view classification method (RSCLN method) provided by this invention in real-world scenarios, experiments were conducted on two medically significant datasets in this embodiment: the schizophrenia dataset COBRE and the invasive breast cancer dataset BRCA. This experiment aimed to evaluate the method's ability to address real-world challenges related to complex biomedical data analysis. In this embodiment, the streaming view classification method (RSCLN) provided by this invention was compared with seven baseline methods, including non-robust and noise-robust methods: (1) Non-robust methods: SVL is a streaming view classification method, ABAD is a single-view classification method, and AWDRM and MCCD are multi-view classification methods. (2) Noise-robust methods: Cycle-T is a single-view label noise classification method, and LNRMC and TMNR are multi-view label noise methods. All comparison methods were configured using the recommended parameters.

[0132] The COBRE dataset contains raw 3D-T1 MRI scans from 61 patients with schizophrenia and 74 healthy controls. Each data sample includes three measurements: sulcus depth, cortical thickness, and curvature, with corresponding dimensions of 2587, 4005, and 2587, respectively. The BRCA dataset contains samples from 875 patients diagnosed with five PAM50 subtypes (LumA, LumB, Basal, Her2, and Normal). Each sample contains preprocessed miRNA, mRNA, and DNA methylation data, with corresponding dimensions of 503, 1000, and 1000, respectively. The results are as follows... Figure 5 As shown.

[0133] like Figure 5 As shown, the streaming view classification method provided by this invention achieves superior performance in most cases. In particular, the advantages of the RSCLN method become increasingly apparent as the noise ratio increases, demonstrating its robustness in handling noise issues. It is worth noting that the SVL method still fails when processing the BRCA dataset, indicating that with limited views, SVL may not be able to acquire enough information to train an effective base model.

[0134] Example 4

[0135] This invention discloses a flow view classification device, such as... Figure 6 As shown, the device 600 includes:

[0136] The data acquisition module 601 is used to receive streaming view data arriving in a streaming order; the streaming view data includes multiple time-series data views corresponding to the same entity; the streaming view data includes one or more of image, audio, and video data;

[0137] The evidence generation module 602 is used to input the data view that arrives at the current time into a pre-built evidence deep neural network model and generate the current evidence vector through non-negative activation transformation; the current evidence vector contains a set of non-negative values, where each value independently represents the strength of the data view supporting the entity to belong to a target category;

[0138] The evidence optimization module 603 is used to perform a linear transformation on the current evidence vector using a preset label transition matrix to obtain a noise-adapted evidence vector; wherein, the label transition matrix is ​​used to characterize the conditional probability relationship between the true label of the target category and the noisy observation label;

[0139] The opinion generation module 604 is used to calculate a set of Dirichlet distribution parameters based on the noise-adapted evidence vector, and generate subjective logical opinions of the data view based on the Dirichlet distribution parameters; the subjective logical opinions include a belief quality vector and an uncertainty parameter; the belief quality vector represents the support for each target category, and the uncertainty parameter represents the overall uncertainty of the classification decision due to missing view information or noise;

[0140] The opinion fusion module 605 is used to synthesize subjective logical opinions with the comprehensive opinions output at the previous time step through a fusion operator based on subjective logic definition to obtain new comprehensive opinions; the synthesis process weights and reduces the view information according to the uncertainty parameters in each opinion to achieve effective accumulation of cross-view information under the uncertainty framework; the new comprehensive opinions are iteratively passed in streaming processing;

[0141] The view classification module 606 is used to determine the final classification identifier of an entity based on the belief quality vector in the new comprehensive opinion.

[0142] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. Their specific functions and technical effects can be found in the method embodiments section, and will not be repeated here. Those skilled in the art will understand that, for the sake of convenience and brevity, the division of the above-mentioned functional units and modules is only used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0143] like Figure 7 As shown, embodiments of the present invention provide a terminal device, such as... Figure 7 As shown, the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 7The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 executes the computer program D102 to implement the steps in any of the above method embodiments.

[0144] Specifically, when processor D100 executes computer program D102, it receives streaming view data arriving in a streaming order; it inputs the data view arriving at the current moment into a pre-built evidence deep neural network model, generating a current evidence vector through non-negative activation transformation; it performs a linear transformation on the current evidence vector using a preset label transition matrix to obtain a noise-adapted evidence vector; it calculates a set of Dirichlet distribution parameters based on the noise-adapted evidence vector, and generates a subjective logical opinion of the data view based on the Dirichlet distribution parameters; it synthesizes the subjective logical opinion with the comprehensive opinion output at the previous moment through a fusion operator based on the subjective logic definition to obtain a new comprehensive opinion; and it determines the final classification label of the entity based on the belief quality vector in the new comprehensive opinion. In this process, by introducing a learnable label transition matrix to explicitly model the annotation noise process and using it throughout forward inference and backward training, the model can learn the true category feature distribution from noisy labeled data, significantly improving classification stability in complex real-world annotation environments and contributing to improved accuracy of streaming view classification; the fusion operator based on subjective logic dynamically adjusts the weight of each view opinion when synthesizing it with historical information according to its own uncertainty parameters. Opinions with high uncertainty automatically contribute less to the final decision, achieving intelligent weighted fusion of multi-source heterogeneous information, enhancing the overall robustness of the system, and improving the accuracy of flow view classification. The model output includes not only a classification label but also a quantified uncertainty parameter, which clearly distinguishes between "high-confidence decisions with sufficient evidence" and "high-probability outputs exhibited by the model due to sufficient training." This provides key confidence criteria for subsequent manual review or integration with other systems, increasing the system's security and credibility.

[0145] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0146] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0147] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0148] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.

[0149] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0150] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A stream view classification method characterized by, The method comprises the following steps: receiving stream view data arriving in a streaming order; the stream view data comprises multiple time-arriving data views corresponding to the same entity; the stream view data comprises one or more of image, audio, and video data; inputting a data view arriving at a current time into a pre-constructed evidence deep neural network model to generate a current evidence vector through a non-negative activation transformation; the current evidence vector comprises a set of non-negative numerical values, each of which independently represents the intensity of the data view supporting the entity belonging to a target category; performing linear transformation on the current evidence vector by using a preset label transition matrix to obtain a noise-adaptive evidence vector; the label transition matrix is used to represent the conditional probability relationship between the true label of the target category and the observation label containing noise; the preset label transition matrix is learned through a model training process, and the total loss function of the model training process at least comprises: a noise-robust cross-entropy loss term for minimizing the difference between the prediction expectation of the model based on the noise-adaptive evidence vector and the observation label, and driving the model to learn to adapt to noise; an uncertainty regularization loss term for constraining the cognitive uncertainty of the model prediction, and preventing the model from making wrong predictions on noisy data by minimizing the KL divergence between the Dirichlet distribution of the prediction expectation and a prior uniform distribution; a transition matrix regularization loss term for constraining the structure of the label transition matrix, and promoting the learned label transition matrix to have probability semantics and numerical stability by minimizing the simplex volume spanned by the column vectors of the label transition matrix; calculating a set of Dirichlet distribution parameters based on the noise-adaptive evidence vector, and generating a subjective logical opinion of the data view according to the Dirichlet distribution parameters; the subjective logical opinion comprises a belief quality vector and an uncertainty parameter; the belief quality vector represents the support degree of each target category, and the uncertainty parameter represents the overall uncertainty of the classification decision due to missing or noisy view information; synthesizing the subjective logical opinion and the comprehensive opinion output at the last time through a fusion operator defined based on subjective logic to obtain a new comprehensive opinion; the synthesis process weights and reduces the view information according to the uncertainty parameters in each opinion to realize effective accumulation of cross-view information in an uncertainty framework; the new comprehensive opinion is iteratively passed in the streaming processing; in the streaming processing, a time consistency constraint is also imposed on the label transition matrix; comprising: determining the final classification label of the entity according to the belief quality vector in the new comprehensive opinion. obtaining a difference norm between the label transition matrix learned at the current time and the label transition matrix at the previous time and the label transition matrix at the previous time ;​ by minimizing the difference norm ensuring that the model's estimate of the label noise pattern remains smooth across consecutive view streams, avoiding false estimates of the noise distribution due to data fluctuations in individual views; The output layer of the evidence deep neural network model is configured to directly output the current evidence vector, and the activation function of the evidence deep neural network model is configured to output only non-negative values, thereby avoiding the competition and forced normalization constraint between the evidence values of each category caused by the use of a Softmax function.

2. The stream view classification method of claim 1, wherein, ​ The non-negative activation transformation adopts a Softplus function or a ReLU function to ensure that the generated evidence vector has a physically non-negative count attribute.

3. The stream view classification method of claim 2, wherein, The linear transformation of the current evidence vector by using a preset label transition matrix obtains a noise-adaptive evidence vector, including: Setting the label transition matrix as ; where, denotes the total number of target classes, and the element in the matrix denotes the estimated probability that the medical entity with the true class is mislabeled as the class . The noise-adaptive evidence vector is obtained by the calculation formula ; wherein, ; wherein, represents the transpose of the label transition matrix , and represents the current evidence vector.

4. The stream view classification method of claim 3, wherein, The calculation of a group of Dirichlet distribution parameters based on the noise-adaptive evidence vector, and the generation of subjective logical opinions of the data view according to the Dirichlet distribution parameters, specifically include: For each entity, a Dirichlet distribution parameter is computed according to the noise-adaptive evidence vector of the entity ; wherein , , , ; The total evidence strength is obtained by the calculation formula ;​ Through calculation formula The belief quality vector is obtained. ;in, , This indicates that the entity belongs to the target category. The reliability quality; The uncertainty parameter is obtained by the calculation formula ​ According to the belief quality vector and the uncertainty parameter constructing a subjective logical opinion ; wherein .

5. The stream view classification method of claim 4, wherein, The synthesis of the subjective logical opinions and the comprehensive opinions output at the last moment by using a fusion operator defined based on subjective logic to obtain new comprehensive opinions, including: Set the comprehensive opinion of the last time as , and the subjective logical opinion of the current time as ; For any target class by the computational formula get new belief quality ; wherein, represents a conflict factor, used to measure the degree of contradiction between the historical integrated opinion and the current view opinion; The calculation formula a new uncertainty parameter is obtained ; the new consolidated opinion , .

6. A stream view classification apparatus characterized by comprising: Including: A data acquisition module is configured to receive stream view data arriving in a streaming order; the stream view data includes a plurality of time-sequentially arriving data views corresponding to the same entity; the stream view data includes one or more of image, audio, and video data; An evidence generation module is configured to input data views arriving at a current moment into a pre-constructed evidence deep neural network model to generate a current evidence vector by using a non-negative activation transformation; the current evidence vector includes a group of non-negative numerical values, each of which independently represents the intensity of the data view supporting the entity belonging to a target category; An evidence optimization module is configured to linearly transform the current evidence vector by using a preset label transition matrix to obtain a noise-adaptive evidence vector; the label transition matrix is used to represent the conditional probability relationship between the true label of the target category and the observation label containing noise; the preset label transition matrix is learned through a model training process, and the total loss function of the model training process at least includes: A noise-robust cross-entropy loss term is used to minimize the difference between the prediction expectation of the model based on the noise-adaptive evidence vector and the observation label, and drive the model to learn to adapt to noise; An uncertainty regularization loss term is used to constrain the cognitive uncertainty of the model prediction, and by minimizing the KL divergence between the Dirichlet distribution of the prediction expectation and a prior uniform distribution, the model is prevented from making wrong predictions on noisy data; A transition matrix regularization loss term is used to constrain the structure of the label transition matrix, and by minimizing the simplex volume spanned by the column vectors of the label transition matrix, the learned label transition matrix is prompted to have probability semantics and numerical stability; An opinion generation module is configured to calculate a group of Dirichlet distribution parameters based on the noise-adaptive evidence vector, and generate subjective logical opinions of the data view according to the Dirichlet distribution parameters; the subjective logical opinions include a belief quality vector and an uncertainty parameter; the belief quality vector represents the support degree of each target category, and the uncertainty parameter represents the overall uncertainty of the classification decision due to missing or noisy view information; an opinion fusion module configured to synthesize the subjective logic opinion and the comprehensive opinion output at the previous time by a fusion operator defined based on subjective logic, to obtain a new comprehensive opinion; the synthesis process weights and reduces view information according to uncertainty parameters in each opinion, to realize effective accumulation of cross-view information under an uncertainty framework; the new comprehensive opinion is iteratively passed in a streaming process; in the streaming process, a time consistency constraint is imposed on a label transition matrix; and by a calculation formula obtaining a difference norm between the label transition matrix learned at the current time and the label transition matrix at the previous time and the label transition matrix at the previous time ;​ by minimizing the difference norm ensuring that the model's estimate of the label noise pattern remains smooth across consecutive view streams, avoiding false estimates of the noise distribution due to data fluctuations in individual views; a view classification module configured to determine a final classification identifier of the entity according to a belief quality vector in the new comprehensive opinion.

7. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the method of any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, the computer-readable storage medium comprising: The computer program is executed by the processor to implement the method of any one of claims 1 to 5.

Images