An adaptive weight multi-modal fusion biometric intelligent classification and prediction method

Abstract

The application discloses a kind of self-adapting weight multimodal fusion biological characteristic intelligent classification prediction method, belong to medical health information technology field, including the following steps: to the dimension alignment and standardization mapping of multimodal biological characteristics, obtain stable feature set;Then, realize redundancy optimization and feature fusion by dimension reduction strategy;Finally, the confidence score is generated by combining redundant operation count and fused feature input classifier for prediction, the self-adapting weight multimodal fusion biological characteristic intelligent classification prediction method provided in the application realizes efficient, robust, controllable intelligent classification, not only can make full use of the complementary information of multimodal feature, improve the discriminability of feature expression, but also can eliminate invalid or repeated information through dynamic redundancy detection and adaptive adjustment, reduce noise interference, so as to significantly improve classification accuracy and stability.At the same time, provide confidence quantification and uncertainty prompt function, make prediction result more reliable.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of healthcare information technology, in particular to a biological feature intelligent classification and prediction method based on adaptive weight multi-modal fusion. BACKGROUND

[0002] The existing multi-modal fusion biological feature intelligent classification and prediction technology collects multiple biological information modal data, such as clinical scales, movement behavior indicators, brain images or physiological signals, and performs basic quality inspection and numerical preprocessing on each modal data. The feature level or decision level fusion method is used to integrate the multi-modal data into a unified representation and input the machine learning or deep learning model for training and prediction. Common methods include support vector machine, random forest and multi-layer neural network, etc. Through such multi-modal fusion, the existing technology can fully utilize the complementarity of different modal information to improve the accuracy and robustness of classification.

[0003] For example, the Chinese invention patent with publication number CN112289441B discloses a medical biological feature information matching system based on multi-modal, which includes: an acquisition module that acquires human information data and corresponding disease information, and establishes a human information feature set and a disease feature set respectively; a model establishment module that performs multi-modal feature fusion according to the human information feature set and the disease feature set through a convolutional neural network, and establishes a matching model and a corresponding disease symptom model; a set establishment module that acquires the human information data to be matched, and establishes a human information feature set; and a matching module that establishes a canonical correlation analysis model, calculates the similarity between the matching model and the human information feature set according to the canonical correlation analysis model, and matches the human information feature set according to the similarity.

[0004] For example, the Chinese invention patent with publication number CN119889701B discloses a disease prediction method and system based on adaptive fusion of microbiome hierarchical features, which includes: using a trained disease prediction model to predict diseases based on microbiome data; the disease prediction model performs multi-level feature extraction on feature data of five classification levels through a hierarchical feature extraction module, and integrates the results into multi-level fusion features; a multi-scale feature fusion module is used to process the multi-level fusion features through three parallel paths of mode flow, context flow and content flow to realize comprehensive representation of microbiome data; a dynamic sampling module is introduced in the training process, and based on the category balance mechanism and the uncertainty evaluation strategy, the sample sampling weight is dynamically adjusted; finally, a classifier is used to predict diseases based on multi-scale fusion features.

[0005] The above-mentioned technology at least has the following technical problems:

[0006] In the multi-modal fusion biological feature intelligent classification, the features of multiple modalities are fused to produce a high-dimensional feature space, and there may be redundant information between different modalities, which not only reduces the utilization rate of effective information, but also significantly increases the classification complexity, and interferes with the discriminant accuracy of the modal weight allocation and matching model. SUMMARY

[0007] In one aspect, an adaptive weight multi-modal fusion biological feature intelligent classification prediction method is provided, which comprises:

[0008] After obtaining the multi-modal biological features, dimension alignment is performed to obtain an initial feature set, stable features are screened to obtain a stable feature set, and feature fusion of the stable feature set is performed.

[0009] The redundancy detection interval is determined, and thus the redundancy of the stable feature set is detected during the feature fusion process of the stable feature set, and a redundancy adjustment strategy is determined and executed.

[0010] After the feature fusion of the stable feature set is completed, the redundancy of the stable feature set is detected again, a redundancy secondary adjustment strategy is determined and executed.

[0011] After receiving the biological feature intelligent classification prediction executable signal, biological feature intelligent classification prediction is performed to generate a confidence score and a prediction result.

[0012] The technical scheme provided by the embodiments of the present application has at least the following beneficial effects:

[0013] 1. The present application realizes efficient, stable and controllable intelligent classification through preprocessing, stability screening, redundancy adjustment and fusion feature classification of multi-modal biological features, which not only makes full use of the complementary information of multi-modal features to improve the discriminant ability of feature expression, but also eliminates invalid or repetitive information through dynamic redundancy detection and adaptive adjustment to reduce noise interference, thereby significantly improving the classification accuracy and stability. At the same time, the confidence quantification and uncertainty prompt function are provided to make the prediction result more reliable and interpretable, and to support manual intervention on low confidence samples, thereby enhancing the safety and operability in practical application.

[0014] 2. The present application calculates the redundancy to identify the high-redundancy feature subset in real time and perform high-redundancy compression to prevent the interference of redundant information features on the model. The dynamic threshold learning and log tracking mechanism can optimize the compression strategy according to the redundancy adjustment trend to realize the rapid convergence and effective control of the redundancy, and also supports principal component analysis dimension reduction operation to map the high-redundancy features to a small number of principal components, which not only reduces the feature dimension but also preserves the core information, improves the compactness and discriminant ability of the fused features, provides high-quality and low-redundancy input for classification prediction, and at the same time avoids invalid or excessive adjustment to improve the running efficiency.

[0015] 3、The application records the number of operations for each feature compression or adjustment in real time through a redundant operation counter, avoids repeated or invalid operations by tracking the history of redundant adjustment, dynamically adjusts the threshold and compression ratio by judging the redundant decline slope after continuous operation, and realizes fine redundant control; the counter can also trigger the freezing mechanism to prevent over-adjustment of features or subsets and protect key information from being lost.

[0016] 4、The application quantifies the confidence of the output result, automatically outputs reliable prediction results by distinguishing the confidence level, prompts manual intervention for low confidence samples, ensures the safety and controllability of prediction, and dynamically optimizes the prediction strategy according to the historical adjustment record in combination with the data of the redundant operation counter, to realize closed-loop optimization of classification performance and feature management. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 is a self-adaptive weight multi-modal fusion biological feature intelligent classification prediction method flow chart provided by the embodiment of the application;

[0019] Figure 2 is a self-adaptive weight multi-modal fusion biological feature intelligent classification prediction method step flow chart provided by the embodiment of the application;

[0020] Figure 3 is a self-adaptive weight multi-modal fusion biological feature intelligent classification prediction method initialization feature flow chart provided by the embodiment of the application;

[0021] Figure 4 is a self-adaptive weight multi-modal fusion biological feature intelligent classification prediction method redundant secondary adjustment flow chart provided by the embodiment of the application. DETAILED DESCRIPTION

[0022] The technical solutions in the application will be described below in combination with the drawings.

[0023] In order to make the technical problems, technical solutions and advantages of the application clearer, the following will be described in detail in combination with the drawings and the biological feature intelligent classification prediction of Parkinson's disease patients, patients without typical dopamine deficiency symptoms and healthy individuals as specific embodiments.

[0024] In this embodiment, multimodal biometrics refers to single-photon emission computed tomography (SPECT) and clinical scale features. The initial feature set obtained after dimensional alignment refers to a unified, standardized, and comparable set of multimodal biometric vectors that integrates SPECT and clinical scales. Each feature corresponds to a vector, ensuring that different modal features are comparable at the same scale.

[0025] like Figure 1 The flowchart shown is a biometric intelligent classification and prediction method based on adaptive weighted multimodal fusion. The processing flow of this method can include the following steps: performing dimensional alignment and standardization mapping on multimodal biometrics, uniformly encoding each feature into a low-dimensional latent space, and then obtaining a stable feature set through multiple rounds of cross-validation and significance testing; next, calculating the correlation coefficient and redundancy between features, and performing dimensionality reduction on a subset of highly correlated features to achieve redundancy optimization and feature fusion; finally, inputting the fused features into a classifier for prediction, and generating a confidence score by combining redundancy operation counts. High confidence scores are directly output as prediction results, while low confidence scores are marked as uncertain for manual review.

[0026] In this embodiment, an adaptive weighted multimodal fusion-based intelligent classification and prediction method for biometrics can be used for intelligent diagnosis and classification of neurological diseases. It is particularly suitable for distinguishing between Parkinson's disease patients, patients without typical dopaminergic deficiency symptoms, and healthy individuals. The main differences between Parkinson's disease, suspected Parkinson's syndrome, and healthy individuals lie in motor and non-motor function performance and brain metabolic patterns. Parkinson's disease patients usually show obvious motor impairments (such as tremor, rigidity, and bradykinesia) and some non-motor symptoms (such as cognitive and sleep abnormalities) on clinical scales. At the same time, single-photon emission computed tomography (SPECT) shows decreased function of the substantia nigra dopaminergic pathway and local metabolic abnormalities. Although patients without typical dopaminergic deficiency symptoms may have some scale scores close to PD, SPECT usually does not show typical substantia nigra function decline, and their metabolic patterns are close to normal or mildly abnormal. Healthy individuals do not have obvious motor or non-motor abnormalities on scale scores, and SPECT shows normal metabolism in the substantia nigra and related brain regions.

[0027] After acquiring multimodal biometric features, dimensional alignment is performed to obtain an initial feature set. This initial feature set is then screened to obtain a stable feature set, and feature fusion of the stable feature set is then performed.

[0028] like Figure 3The flowchart shown is an initialization feature flowchart for an intelligent classification and prediction method for biometrics based on adaptive weighted multimodal fusion. It includes: normalizing the input multimodal biometrics to a standardized interval, calculating the mean and standard deviation and modeling them as Gaussian distributions, generating smooth features through reparameterized sampling, then projecting the input into an embedding mapping layer into a low-dimensional latent space and determining the dimension based on the feature mean, and finally performing dimension alignment to obtain a unified and comparable initial feature set for subsequent stability analysis, redundancy detection, and feature fusion.

[0029] Furthermore, after acquiring multimodal biometric features, dimensional alignment is performed to obtain an initial feature set. The specific analysis method is as follows:

[0030] A1 performs normalization processing on the input multimodal biometrics, mapping features of different dimensions to a normalized numerical range, thereby obtaining the normalized feature values.

[0031] It should be noted that the input single-photon emission computed tomography (SPECT) images and clinical scales are first normalized to map the feature values ​​of different modalities and scales to the standardized range of 0 to 1. The minimum and maximum values ​​of each feature are calculated, and then linear normalization is used to subtract the minimum value from the feature value and divide by the range (maximum value minus minimum value) to obtain the corresponding normalized value. For features with large distribution differences, logarithmic or Z-score transformations can be performed first to compress the influence of extreme values ​​before linear normalization is performed to eliminate the bias caused by inconsistent value ranges in subsequent analyses.

[0032] A2 calculates the mean and standard deviation of each normalized eigenvalue, models each normalized eigenvalue as a Gaussian distribution, and generates a smooth feature representation through reparameterized sampling.

[0033] It should be noted that after normalization, the normalized mean and standard deviation of each feature are calculated to characterize its central tendency and dispersion. Based on this, it is assumed that each normalized feature follows a Gaussian distribution, and a reparameterized sampling strategy is used for feature generation: first, noise terms are randomly sampled from the standard normal distribution, and then a smooth feature representation is generated.

[0034] A3 embeds the smooth feature representation into the mapping layer and projects it into a unified low-dimensional latent space. The dimension of this low-dimensional latent space is determined based on the mean of the feature values.

[0035] It should be noted that the smoothed feature representation is input into the embedding mapping layer, and the high-dimensional features are projected into a unified low-dimensional latent space through the autoencoder encoding layer.

[0036] In this embodiment, a pre-defined correspondence rule between the distribution range of the feature value mean and the dimension of the low-dimensional latent space is stored in a configuration file in the form of an interval mapping table or a function model for unified access. During actual operation, the mean distribution characteristics of the current feature set are first calculated to obtain the degree of concentration of the feature value mean. For each feature in the feature set, its mean over all samples is calculated to obtain a feature mean vector. Subsequently, statistical analysis is performed on all feature means, including calculating the overall mean, variance, or standard deviation of the means to measure the dispersion or concentration of the means. To more intuitively reflect the distribution characteristics, a histogram is further plotted to analyze the distribution range, peak value, and skewness of the feature means. The interquartile range of the means is then calculated as the degree of concentration. Based on this degree of concentration, the corresponding latent space dimension is found or calculated from the pre-defined rules. It should be noted that the more concentrated the distribution of the feature value mean, the higher the redundancy and the weaker the differences between features. Therefore, the latent space dimension is reduced to compress redundant information.

[0037] A4, based on the dimensions of the low-dimensional latent space, performs dimension alignment to obtain dimension-aligned multimodal biological features, which are denoted as the initial feature set.

[0038] It should be noted that after completing the latent space mapping, the features of different modalities are dimensionally aligned according to the determined latent space dimensions. Since the original features of each modality differ in dimensionality and statistical distribution, an interpolation mapping method is used to uniformly adjust the features of different modalities to the same dimensional structure and numerical scale.

[0039] After dimensional alignment is completed, the alignment features of each modality in the low-dimensional latent space are structurally integrated to form a multimodal biofeature set with uniform length and comparability, which is recorded as the initial feature set.

[0040] It should be explained that interpolation mapping is a numerical reconstruction method for aligning features of different modalities. Its core idea is to insert estimated values ​​between the original feature points so that the features maintain a smooth and continuous trend of change in the new dimensional space. In this embodiment, for the differences in the feature dimensions of each modality, the target alignment dimension and the corresponding index interval are first determined. Then, based on the value distribution of the original features in the interval, linear interpolation is used to smoothly extrapolate the feature values ​​at the missing or mismatched positions, thereby generating a new feature vector consistent with the target dimension.

[0041] It should be added that a significance test can also be performed on each feature value. Each feature is grouped by category to form a set of feature values ​​under each category. Then, a non-parametric significance test is performed on each feature: when comparing two categories, the Wilcoxon rank-sum test is used; when comparing three or more categories, the Kruskal-Wallis test is used. The significance level of each feature is obtained through the test and compared with a preset significance level threshold.

[0042] If the significance level of a certain feature is higher than the significance level threshold, it is marked as an inefficient feature and returned to the dimension alignment step. Dynamic dimension adjustment is performed on these features: the reduction value of the output dimension of the coding layer is determined based on the deviation value between the significance level and the significance level threshold.

[0043] If the significance level of a feature is lower than or equal to the significance level threshold, it is recorded as an efficient feature, and subsequent operations are performed.

[0044] like Figure 2 The flowchart shown is a biometric intelligent classification and prediction method based on adaptive weighted multimodal fusion. The processing flow of this method includes: obtaining a stable feature set after cross-validation of the initial feature set; performing redundancy detection on the stable feature set to determine whether to use a high redundancy compression strategy or continue feature fusion; high redundancy compression involves dividing the stable feature set into a highly relevant feature subset, then determining a compressible subset within the highly relevant feature subset, performing compression, and then performing feature fusion; after fusion, redundancy detection is performed again, which is divided into dimensionality reduction strategy and freezing strategy. If it is a dimensionality reduction strategy, classification is performed after dimensionality reduction; if it is a freezing strategy, classification is stopped and dynamic threshold learning is performed. After classification, a confidence score is generated.

[0045] Furthermore, a stable feature set is obtained through screening, and the specific analysis method is as follows:

[0046] The initial feature set is subjected to multiple rounds of random sampling and cross-validation to obtain the feature selection results of the initial feature set. Frequency statistics and weighted variance analysis are performed on the feature selection results obtained in each round of training to calculate the selection probability and contribution volatility of features among different feature selection results.

[0047] It should be noted that in this embodiment, the initial feature set is used as input data. Feature selection is performed through multiple rounds of random sampling and cross-validation. In each round, the training and validation sets are randomly divided. A feature selection algorithm is used to extract a subset of features that significantly contribute to the classification or prediction task. After multiple rounds of training, the feature selection results of each round are statistically analyzed. The number of times each feature is selected in different rounds is counted to obtain its selection probability, which is used to measure the stability of the feature in multiple experiments. Simultaneously, variance analysis is performed on the weight values ​​of the same feature in each round to reflect the degree of fluctuation in the importance of the feature under different training conditions, denoted as the contribution volatility.

[0048] It should be noted that, firstly, the features selected in each training round are recorded and aligned, and the number of times each feature is selected in all rounds is counted, which is statistical analysis. The selection probability is obtained by dividing the number of times by the total number of rounds, so as to quantify the stability of the feature in multiple experiments. At the same time, for each identical feature, its corresponding weight value sequence in each training round is collected, and the variance of these weights is calculated, which is analysis of variance, reflecting the fluctuation of the feature's contribution under different training subsets or random initialization.

[0049] If a feature has a selection probability greater than or equal to a preset selection probability threshold and a weight variance less than or equal to a weight variance threshold in multiple training sessions, it is marked as a stable feature; otherwise, it is marked as an unstable feature.

[0050] It should be noted that if a feature has a selection probability greater than or equal to a preset selection probability threshold and a weight variance less than or equal to a weight variance threshold in multiple training iterations, it means that the feature can maintain a high level of importance and consistent contribution under different sample partitioning and training conditions, and has good stability and reliability. Therefore, it is marked as a stable feature and can be used as an important input for subsequent fusion and classification models. Conversely, if the feature's performance is unstable in different training rounds, its importance is easily affected by sample distribution or random factors, and it is difficult to guarantee a continuous contribution to the model. Therefore, it is marked as an unstable feature.

[0051] Unstable features are removed from the initial feature set to obtain a stable feature set.

[0052] The redundancy detection interval is determined, and the redundancy of the stable feature set is detected during the feature fusion process of the stable feature set. A redundancy adjustment strategy is then determined and executed.

[0053] Furthermore, the redundancy detection interval is determined, and thus, during the feature fusion process of the stable feature set, redundancy detection of the stable feature set is performed. The specific analysis method is as follows:

[0054] After obtaining the stable feature set, obtain the number of features in the stable feature set.

[0055] Determine the redundancy detection interval duration of the stable feature set according to the number of features in the stable feature set, and start the redundancy detection of the stable feature set based on the detection interval duration.

[0056] In this embodiment, a corresponding rule between the number of features in the stable feature set and the redundancy detection interval duration is preset and stored in the configuration file in the form of an interval mapping table or a function model for unified call. During the actual operation process, first count the number of features in the current stable feature set, and then, according to this number of features, look up or calculate the corresponding redundancy detection interval duration from the preset rules. When the number of features is small, the information distribution may be insufficient, and the detection interval is correspondingly extended to reduce the calculation overhead. On the contrary, when the number of features is large, a shorter detection interval is adopted accordingly to increase the detection frequency.

[0057] Collect the redundancy dependence parameters of the stable feature set, including the average cross-correlation coefficient, the average mutual information content, and the average variance inflation factor.

[0058] Analyze the redundancy of the stable feature set according to the redundancy dependence parameters.

[0059] The average cross-correlation coefficient quantifies the linear or non-linear dependence relationship between features by calculating the Pearson correlation coefficient between each pair of features in the stable feature set. Subsequently, take the absolute value of the correlation coefficients of all feature pairs and calculate the average to obtain the average cross-correlation coefficient of the entire stable feature set, which is used to reflect the strength of the overall redundancy level between features, expressed as , where A represents the average cross-correlation coefficient, there are n features in the feature set, i and j represent the i-th feature and the j-th feature, i = 1, 2, 3,..., n, j = 1, 2, 3,..., n. To calculate the index between each pair, usually only consider i ≠ j and each pair is only calculated once without repeated calculation. The number of pairwise combinations of features is , directly sum all combinations and then calculate the average, and need to divide by the total number of combinations. To simplify the formula, convert the division to multiplication. The meaning of i < j is to only calculate each pair of features once. represents the Pearson correlation coefficient.

[0060] The average mutual information content calculates the joint probability distribution by discretizing continuous features or using the kernel density estimation method, and then calculates the mutual information value. Calculate the average of the mutual information values of all feature pairs, that is, obtain the average mutual information content, expressed as , represents the feature and the feature between the mutual information content, S represents the average mutual information content, , respectively represent the i-th and j-th features in the stable feature set.

[0061] For each feature in the stable feature set, a linear regression model is established between it and all other features. The variance inflation factor for that feature is calculated, where is the coefficient of determination of that feature's regression with respect to other features. Then, the average variance inflation factors of all features are averaged to obtain the average variance inflation factor, denoted as . D represents the average variance inflation factor, V i V represents the variance inflation factor of the i-th feature in the stable feature set. max This represents the maximum value of all Vᵢ in the stable feature set.

[0062] A higher average cross-correlation coefficient indicates a stronger linear dependence between features, usually accompanied by a higher average mutual information content, suggesting more shared information and higher redundancy among features. Simultaneously, a higher average variance inflation factor indicates increased multicollinearity, meaning the regression contribution of a single feature is more significantly influenced by other features. Therefore, as the average cross-correlation coefficient and average mutual information content increase, the average variance inflation factor tends to increase as well, reflecting an overall higher level of feature set redundancy and lower independent information content.

[0063] A higher average cross-correlation coefficient indicates stronger linear dependence and greater redundancy among features. A higher average mutual information indicates more shared information and richer redundant information among features. A higher average variance inflation factor indicates more severe multicollinearity among features and a higher overall level of redundancy. In summary, the higher these three indicators are, the greater the redundancy of the feature set, which may lead to a decrease in the utilization of effective information and a limitation in the model's generalization ability.

[0064] In this embodiment, three types of contribution coefficients are introduced to address the redundancy of the feature set: the average cross-correlation coefficient contribution coefficient, the average mutual information coefficient contribution coefficient, and the average variance inflation factor contribution coefficient, corresponding to the average cross-correlation coefficient, average mutual information, and average variance inflation factor, respectively. All three contribution coefficients range from 0 to 1, and are designed to ensure a weighted sum of 1, quantifying the relative contribution of each type of redundancy to the overall redundancy of the stable feature set. A mapping relationship between each redundancy index and its corresponding contribution coefficient is pre-established, and the mapping rules are stored in a database. During feature set analysis, the average cross-correlation coefficient, average mutual information, and average variance inflation factor, calculated in real time, are converted into their corresponding contribution coefficients through the mapping rules and normalized to ensure that the weighted sum of the three coefficients is always 1. This approach enables weighted fusion evaluation of linear redundancy, nonlinear redundancy, and single-feature multiple redundancy, providing a reliable quantitative basis for redundancy optimization and feature selection of the stable feature set, and guiding the implementation of subsequent feature adjustment or fusion strategies.

[0065] The three indicators—average cross-correlation coefficient, average mutual information, and average variance inflation factor—are uniformly normalized to ensure they are comparable and weighted on the same scale. The specific procedure is as follows: First, calculate the average value of each indicator across all feature pairs or subsets; then, apply linear normalization to each average value, mapping it to the standard interval [0,1], i.e., subtract the minimum value and divide by the range; for indicators with large distribution differences, logarithmic or Z-score transformations can be performed first to compress the influence of extreme values ​​before normalization; the normalized indicators can then be combined to calculate the overall redundancy.

[0066] The average cross-correlation coefficient reference value, the average mutual information reference value, and the average variance inflation factor reference value together constitute the core benchmarks for measuring the degree of redundancy information in a stable feature set. During the setup process, the average cross-correlation coefficient, average mutual information, and average variance inflation factor data of the stable feature set are collected at each round to quantify the linear relationships, information redundancy, and multicollinearity among features. Specifically, the following steps are taken: First, the average values ​​of the above three indicators are calculated for each feature subset in each round. Then, statistical and trend analyses are performed on the indicators under different feature combinations, sample distributions, and data complexities to extract the mean values. Subsequently, based on the operational patterns, an arithmetic mean is used to generate the reference benchmark range for each indicator, and an appropriate safety margin is added on top of this to form preset reference values ​​that can be used for judgment during redundancy detection and feature compression, ensuring the stability and reliability of feature fusion and redundancy reduction operations.

[0067] The redundancy of a stable feature set is a quantitative indicator of the overall redundancy information of the stable feature set, which is jointly expressed by the average cross-correlation coefficient, the average mutual information, and the average variance inflation factor. The specific analysis process is as follows: After weighting the average cross-correlation coefficient, the average mutual information, and the average variance inflation factor with their respective contribution coefficients, the redundancy of the stable feature set is obtained.

[0068] ,

[0069] Where Q represents the redundancy of the stable feature set, α represents the contribution coefficient of the average cross-correlation coefficient, A represents the average cross-correlation coefficient, A0 represents the reference value of the average cross-correlation coefficient, β represents the contribution coefficient of the average mutual information, S represents the average mutual information, S0 represents the reference value of the average mutual information, γ represents the contribution coefficient of the average variance inflation factor, D represents the average variance inflation factor, and D0 represents the reference value of the average variance inflation factor.

[0070] Furthermore, the redundancy adjustment strategy is determined, and the specific analysis method is as follows:

[0071] Extract the preset redundancy threshold from the database.

[0072] It should be noted that the preset redundancy threshold is based on historical data statistics and operational patterns to ensure the effectiveness and stability of redundancy detection. The specific process is as follows: During multiple training or validation cycles, the distribution of redundancy under different sample, image, or load conditions is collected; then, these data are statistically analyzed to extract the mean, typical intervals, and extreme value distributions, and preliminary thresholds are determined according to the importance of different indicators and fault tolerance requirements; next, the final threshold can be generated using the average, weighted average, or percentile methods, and a safety margin is added on it to prevent abnormal samples from affecting the judgment; finally, the generated threshold is stored in the configuration for subsequent redundancy adjustment and judgment, realizing robust control of the model in actual operation.

[0073] If the redundancy of the stable feature set is greater than or equal to the redundancy threshold, the redundancy adjustment strategy will be determined as high redundancy compression.

[0074] It should be noted that if the redundancy of the stable feature set is greater than or equal to the preset redundancy threshold, it means that there is high redundancy information in the feature set, that is, some features are highly correlated or information is repeated, which may lead to a decrease in the utilization rate of effective information, lengthy feature representation, increased model computational complexity, and may easily affect the generalization ability of the classifier or prediction model. Therefore, the redundancy adjustment strategy is determined to be high redundancy compression.

[0075] If the redundancy of the stable feature set is less than the redundancy threshold, the feature fusion process of the stable feature set continues.

[0076] It should be noted that when the redundancy of a stable feature set is less than the redundancy threshold, it means that there is little redundant information among the features in the set, and the features are relatively independent and have good information complementarity. This indicates that the feature set has high effective information utilization, the feature representation is concise and rich in diversity, which is conducive to the classifier or prediction model learning patterns more efficiently, improving the model's generalization ability and stability. At the same time, there is no need to perform additional redundancy adjustment on the feature set, and the feature fusion process of the stable feature set can continue.

[0077] It should be added that the specific operation process of feature fusion is as follows: the low-dimensional vectors obtained by the autoencoder layer of each modality in the stable feature set are concatenated in the dimensional direction to form a unified fusion vector. The low-dimensional vectors of each modality are then horizontally superimposed so that the information of different modalities can be fully expressed in the same vector.

[0078] Furthermore, the specific analysis method for high-redundancy compression is as follows:

[0079] Each highly correlated feature subset is obtained based on the stable feature set.

[0080] It should be noted that the Pearson correlation coefficient between features is calculated based on the stable feature set and is denoted as the feature correlation coefficient. Features whose feature correlation coefficient is greater than or equal to the preset feature correlation coefficient threshold in the database are grouped into the same set and denoted as each highly correlated feature subset.

[0081] Each highly relevant feature subset is compressed, and the basic compression ratio is determined by the deviation between the redundancy of the stable feature set and the redundancy threshold.

[0082] It should be noted that the redundancy deviation value is obtained by subtracting the redundancy threshold from the redundancy of the stable feature set.

[0083] In this embodiment, the base compression ratio is adaptively determined based on the redundancy deviation value. Specifically, the overall redundancy of the current stable feature set is first calculated to obtain the redundancy deviation value. Then, the corresponding base compression ratio is searched or calculated from a pre-set compression ratio mapping rule based on the redundancy deviation value. The larger the redundancy deviation value, the more redundant information the feature set contains, and the base compression ratio is increased accordingly to compress more redundant features.

[0084] Based on the redundancy operation counter, the number of redundant operations is obtained, and it is determined whether the number of redundant operations is zero. If the number of redundant operations is zero, the basic compression ratio is recorded as the execution compression ratio.

[0085] It should be noted that zero redundancy operation count means that no features requiring redundancy adjustment were found during the redundancy detection of the previous stable feature set. This indicates that the previous feature set structure was relatively stable and the feature information utilization rate was high. Therefore, there is no need to perform compression ratio adjustment operation this time, and the basic compression ratio is recorded as the execution compression ratio.

[0086] If the number of redundant operations is not zero and the number of redundant operations is greater than the threshold for the number of redundant operations, an early warning message will be generated.

[0087] It should be added that the preset threshold for redundant operation counts is based on historical data statistics and operational patterns to ensure the effectiveness and stability of redundant operation count detection. The specific process is as follows: During multiple training or validation cycles, the impact of operation counts under different sample, image, or load conditions is collected; then, these data are statistically analyzed to extract the mean, typical intervals, and extreme value distributions, and preliminary thresholds are determined according to the importance of different indicators and fault tolerance requirements; then, the final threshold can be generated using the average, weighted average, or percentile methods, and a safety margin is added on it to prevent abnormal samples from affecting the judgment; finally, the generated threshold is stored in the configuration for subsequent redundant operation count determination, realizing robust control of the model in actual operation.

[0088] It should be noted that if the number of redundant operations is not zero and the number of redundant operations is greater than the threshold for the number of redundant operations, it means that a large number of redundant features were found in the redundancy detection of the previous stable feature set, and the number of times these redundant features have been processed has exceeded the set safety or reasonable range. In this case, no further adjustment can be made and an early warning message will be generated.

[0089] In this embodiment, the warning message could be: "Attention! The number of redundant operations exceeds the threshold."

[0090] If the number of redundant operations is not zero and is less than or equal to the threshold for the number of redundant operations, query the cumulative adjustment level, determine the basic compression ratio optimization value based on the cumulative adjustment level and the number of redundant operations, and obtain the execution compression ratio based on the basic compression ratio optimization value and the basic compression ratio.

[0091] It should be noted that the execution compression ratio is obtained by adding the base compression ratio optimization value to the base compression ratio.

[0092] In this embodiment, the cumulative adjustment degree of the feature set during the previous redundancy adjustment process and the current number of redundancy operations are first counted to quantify the feature set redundancy status and historical adjustment intensity. Based on the cumulative adjustment degree and the number of redundancy operations, the corresponding basic compression ratio optimization value is searched or calculated from the pre-set compression ratio mapping rules. When the cumulative adjustment degree is large and the number of redundancy operations is high, it indicates that the feature set has a lot of redundancy or insufficient adjustment in the early stage. The basic compression ratio optimization value is increased accordingly to increase the compression intensity.

[0093] Based on the preset number of features to retain in the database, a compressible subset is selected.

[0094] It should be noted that the number of features retained is the number of features that are retained without compression or dimensionality reduction during feature compression or dimensionality reduction, in order to ensure the integrity of the feature information.

[0095] Obtain the number of features in each highly relevant feature subset, compare it with the preset number of features to be retained, and filter out the highly relevant feature subsets whose number of features is greater than the number of features to be retained, and record them as compressible subsets.

[0096] The compression ratio is evenly distributed to the compressible subset to perform high-redundancy compression. After the high-redundancy compression is completed, the result is recorded in the log, and the number of redundant operations in the redundancy operation counter is incremented by 1.

[0097] It should be noted that, firstly, the target compression dimension is obtained by multiplying the original feature dimension with the compression ratio. Then, a linear mapping matrix is ​​constructed. After obtaining the target compression dimension, a linear mapping matrix for feature dimensionality reduction is constructed through random orthogonal initialization, projecting high-dimensional features onto a low-dimensional subspace. Specifically, the process is as follows: First, the number of rows in the matrix equals the original feature dimension, and the number of columns equals the target compression dimension. Each row corresponds to a feature in the original feature vector, and each column corresponds to a dimension of the compressed subspace. Then, a set of column vectors with the same target dimension is randomly generated, and these are processed sequentially using the Gram-Schmidt orthogonalization method. Each vector is subtracted from its projection onto the already orthogonal vectors, making all column vectors mutually orthogonal. Next, each orthogonal vector is normalized to maintain scale consistency. Finally, the original feature vectors are multiplied by this orthogonal mapping matrix to obtain the low-dimensional projected features, thus completing the high-redundancy compression.

[0098] It should be added that when the redundancy of the stable feature set is lower than the redundancy threshold, or the number of redundant operations reaches the number of operations threshold, the system can end the current round of redundant operations when either condition is met. This ensures that feature compression is both effective and avoids excessive operations, while providing a stable input feature set for subsequent feature fusion and classification modeling.

[0099] like Figure 4 The flowchart shown is a secondary adjustment flowchart for the redundancy of an intelligent biometric classification and prediction method based on adaptive weighted multimodal fusion. It includes: when redundancy detection shows that the redundancy still exceeds the threshold, the number of operations is checked; if the threshold is reached, a freezing mechanism is triggered, compression is stopped and an early warning is generated, while dynamic threshold learning is initiated—the threshold is dynamically adjusted by analyzing the redundancy descent slope, and then an adaptive threshold is generated using an exponentially weighted average; if the number of operations does not reach the threshold, a dimensionality reduction strategy is executed: the redundancy deviation is calculated to determine the number of dimensionality reductions, the subsets that can be reduced in dimensionality are selected and sorted before dimensionality reduction is performed, and finally, an executable signal is generated through feature fusion.

[0100] After completing the feature fusion of the stable feature set, the redundancy of the stable feature set is detected again, and the redundancy secondary adjustment strategy is determined and executed.

[0101] Furthermore, redundancy detection of the stable feature set is performed again to determine the secondary redundancy adjustment strategy. The specific analysis method is as follows:

[0102] The redundancy of the stable feature set is detected again to obtain the redundancy of the stable feature set, which is denoted as the first redundancy.

[0103] If the first redundancy is less than the redundancy threshold, the redundancy secondary adjustment strategy is recorded as no adjustment is needed, the fused features are recorded as the final fused features, and an executable signal for intelligent classification and prediction of biometric features is generated.

[0104] It should be noted that a redundancy level less than the redundancy threshold means that there is less redundant information between the fused features, the features are relatively independent and have good information complementarity, that is, the feature set structure is relatively compact, the effective information utilization rate is high, and the generated feature set can cover the key information of multimodal data well. It can be directly used for subsequent fusion or training without additional compression or removal operations.

[0105] If the first redundancy is still greater than or equal to the redundancy threshold, the number of redundant operations is extracted. If the number of redundant operations is less than the redundancy operation threshold, the redundancy secondary adjustment strategy is recorded as the dimensionality reduction strategy.

[0106] It should be noted that if the first redundancy is greater than or equal to the redundancy threshold, it means that there is significant redundant information in the generated feature set, some features are highly correlated or have information duplication, which may lead to a decrease in the utilization rate of effective information, lengthy feature representations, and increased model training complexity and overfitting risk. If the number of redundant operations is less than the redundancy operation threshold, it means that the number of redundant features that need to be processed in the previous operation is small and has not yet reached the preset operation limit. In this case, the redundancy can be adjusted a second time. Therefore, the redundancy adjustment strategy is called the dimensionality reduction strategy.

[0107] If the number of redundant operations is greater than or equal to the threshold for the number of redundant operations, then the secondary adjustment strategy for redundancy is recorded as the freezing strategy.

[0108] It should be noted that if the number of redundant operations is greater than or equal to the redundancy operation threshold, it means that the previous high redundancy adjustment was performed many times and the preset operation limit has been reached. In order to avoid over-adjustment and loss of feature information, redundant operations cannot be performed this time, and the current feature set state can only be maintained. Therefore, the redundancy adjustment strategy is recorded as the freezing strategy.

[0109] Furthermore, the dimensionality reduction strategy is implemented as follows:

[0110] The first redundancy and the redundancy threshold are deviated to obtain the first redundancy deviation value.

[0111] The first redundancy deviation value is obtained by subtracting the redundancy threshold from the first redundancy value.

[0112] The number of dimensionality reductions is determined based on the first redundancy deviation value.

[0113] In this embodiment, the first redundancy deviation value of the stable feature set is calculated based on the first redundancy, which is the difference between the current redundancy and the redundancy threshold. Then, based on the first redundancy deviation value, the corresponding number of dimensionality reductions is searched or calculated from a pre-set dimensionality reduction mapping rule or function model. This mapping rule can be obtained through offline experiments or historical data analysis: dimensionality reduction experiments are conducted on the feature set at different redundancy levels, the optimal number of dimensionality reductions corresponding to each redundancy is statistically analyzed, and the results are organized into an interval mapping table or fitted into a continuous function model so as to quickly search or calculate the number of dimensionality reductions in actual operation. The specific logic is: when the first redundancy deviation value is large, it indicates that there is a lot of redundant information in the feature set, and the number of dimensionality reductions is increased accordingly to compress more redundant features.

[0114] Based on the preset number of features to retain, a subset that can be reduced in dimensionality is selected from each subset of highly relevant features.

[0115] Obtain the number of features of the reducible subsets, and sort the reducible subsets from largest to smallest based on this.

[0116] The dimensionality reduction process of the dimensionality-reducible subset is executed based on the order of the number of features in the dimensionality-reducible subset and the number of dimensionality reductions. After the dimensionality reduction is completed, the feature fusion of the stable feature set is re-executed to obtain the final fused features and generate an executable signal for intelligent classification and prediction of biometric features.

[0117] It should be noted that principal component analysis is used to map the dimensionality-reducible subset onto a small number of principal components, calculate the covariance matrix between features, extract its eigenvalues ​​and eigenvectors, and denote them as contributing eigenvalues ​​and variance contribution rates, respectively. The eigenvalues ​​are then sorted from largest to smallest to reflect the degree of contribution of each principal component to the overall variance. The number of principal components to be retained is determined based on the cumulative variance contribution rate. When the cumulative variance contribution rate reaches the preset variance contribution ratio, the corresponding number of principal components are selected to construct a new feature subspace. Finally, the original feature data of the highly correlated feature subset is projected onto this principal component space to generate the dimensionality-reduced feature vectors.

[0118] In one specific embodiment, there are 5 reducible subsets 1, 2, 3, 4, 5, with feature counts of 20, 8, 15, 23, and 13, respectively. The reducible subsets are sorted from largest to smallest as 23, 20, 15, 13, 8. The dimensionality reduction process is performed according to this order. If the dimensionality reduction count is 3, then the reducible subsets 4, 1, and 3 corresponding to 23, 20, and 15 are dimensionality reduced. If the dimensionality reduction count is 8, it means that the reducible subsets are insufficient for the dimensionality reduction process, and then the process jumps to the freezing strategy.

[0119] Furthermore, the freezing strategy is implemented as follows:

[0120] Stop performing intelligent biometric classification and prediction, generate early warning information, and simultaneously start dynamic threshold learning.

[0121] The dynamic threshold learning is initiated by calculating the average decrease slope of the redundancy of the stable feature set after the redundancy operation based on the number of redundancy operations.

[0122] It should be added that the convergence trend of redundancy elimination is quantified by calculating the redundancy change of the stable feature set after each compression or adjustment. The specific process is as follows: First, record the redundancy value after each redundancy operation. Then, according to the operation sequence, form a sequence of redundancy values ​​with the corresponding number of operations or time. Next, use linear regression to calculate the slope of the change of redundancy with the number of operations. Divide the decrease of adjacent redundancy by the detection interval and average it to obtain the average redundancy decrease slope of the stable feature set.

[0123] If the average decline slope of redundancy is less than or equal to the preset average decline slope threshold, the redundancy threshold increase value is determined based on the number of redundancy operations.

[0124] It should be added that the preset average descent slope threshold is based on historical data statistics and operational patterns to ensure the effectiveness and stability of the average descent slope of redundancy. The specific process is as follows: In multiple training or validation cycles, the distribution of the average descent slope of redundancy under different sample, image, or load conditions is collected; then, these data are statistically analyzed to extract the mean, typical intervals, and extreme value distributions, and preliminary thresholds are determined according to the importance of different indicators and fault tolerance requirements; then, the final threshold can be generated using the average, weighted average, or percentile methods, and a safety margin is added on it to prevent abnormal samples from affecting the judgment; finally, the generated threshold is stored in the configuration for subsequent redundancy threshold adjustment and judgment, realizing robust control of the model in actual operation.

[0125] It should be noted that if the average slope of redundancy decreases by less than or equal to the preset average slope threshold, it means that during continuous redundancy adjustment or compression, the overall redundancy of the feature set decreases by a small margin. In other words, the effect of compressing or eliminating redundant features is limited, and the adjustment operations of high redundancy compression and dimensionality reduction strategies do not significantly improve the redundancy level. This may indicate that the remaining redundant features of the feature set are difficult to compress further, so it is necessary to increase the redundancy threshold.

[0126] In this embodiment, the redundancy threshold increase value is determined based on the number of redundant operations in the stable feature set. Specifically, the number of redundant operations performed in the current redundancy adjustment process is first counted and compared with the operation count threshold to quantify the redundancy adjustment intensity. Then, the corresponding redundancy threshold increase value is searched or calculated from the preset threshold adjustment mapping rules or function models based on the operation count. When the number of redundant operations is large, close to or exceeds the threshold, it indicates that there are a large number of unadjustable redundant features in the feature set, and the redundancy threshold is increased accordingly.

[0127] If the average slope of redundancy decreases is greater than the preset average slope threshold, the redundancy threshold reduction value is determined based on the deviation between the slope of decrease and the average slope threshold.

[0128] It should be noted that if the average slope of redundancy decreases is greater than the preset average slope threshold, it means that the overall redundancy of the feature set decreases significantly during continuous redundancy adjustment or compression. In other words, the adjustment operations of high redundancy compression and dimensionality reduction strategies have a significant effect on improving the redundancy level. In other words, redundant features are effectively reduced and the feature set structure is rapidly optimized. This indicates that there is a lot of adjustable redundant information in the feature set. Therefore, the redundancy threshold is lowered to improve sensitivity.

[0129] The slope deviation is obtained by subtracting the average slope threshold from the slope of descent.

[0130] In this embodiment, the reduction value of the redundancy threshold is determined based on the deviation between the average decrease slope of the redundancy of the stable feature set and the preset average decrease slope threshold. Specifically, firstly, the deviation value of the decrease slope of redundancy during continuous redundancy adjustment or compression is calculated. Then, based on the decrease slope deviation value, the corresponding redundancy threshold reduction value is searched or calculated from the preset threshold adjustment mapping rules or function models. When the decrease slope deviation value is large, it indicates that the redundancy decreases too quickly or the adjustment force is too large. In order to avoid excessive compression leading to the loss of effective feature information, the redundancy threshold is increased accordingly.

[0131] Using an exponentially weighted average, based on the increase or decrease of the redundancy threshold, a smooth update of the redundancy threshold is performed to generate the redundancy threshold to be used in the next round of adjustment, and the update record is written to the log.

[0132] After receiving the biometric intelligent classification prediction executable signal, the system performs biometric intelligent classification prediction and generates confidence scores and prediction results.

[0133] Furthermore, intelligent classification and prediction based on biometric features are performed to generate confidence scores and prediction results. The specific analysis method is as follows:

[0134] After receiving the biometric intelligent classification prediction executable signal, the final fused features are input into the trained classifier for recognition and prediction.

[0135] The classifier calculates the probability values ​​of each category based on the spatial distribution and pattern differences of the fused features, thereby generating the corresponding prediction results and initial confidence scores.

[0136] It should be noted that a classifier is a model that maps fused features to discrete categories. By learning the discrimination rules between features and categories, it can automatically identify and predict unknown samples and output the corresponding confidence scores.

[0137] The classifier receives fused features as input and performs feature matching, pattern recognition, and class probability calculation on the input samples according to the feature mapping and discrimination rules learned during training, thereby generating the corresponding prediction results and initial confidence scores.

[0138] In a specific embodiment, a classification prediction model can be constructed using a cascaded ensemble learning framework. The models are arranged in a hierarchical manner. The first-layer random forest model performs preliminary classification or pattern extraction on the input features and outputs prediction results or intermediate features. Subsequent support vector machine layers use the output of the previous layer as input to further learn residual information or enhance discrimination ability, thereby gradually improving classification accuracy.

[0139] Taking the intelligent classification and prediction of biometrics for Parkinson's disease patients, patients without typical dopaminergic deficiency symptoms, and healthy individuals as an example, the fused multimodal biometrics are first input into the model, including single-photon emission computed tomography (SPECT) image features and clinical scale features. The model uses a multi-level base model combination and ensemble strategy to input the fused features into the first-layer ensemble model, a random forest, to perform preliminary classification of samples and generate class probabilities. Subsequently, the output of the first layer is combined with the original features and input into the second-layer ensemble model, a support vector machine, to further learn the complex relationships between features and residual information, thereby improving the discrimination accuracy of difficult-to-classify samples. Within each layer, the prediction results of multiple models are fused through voting to reduce single-model bias and variance. Finally, three prediction results are generated: Parkinson's disease patients, patients without typical dopaminergic deficiency symptoms, or healthy individuals, along with an initial confidence score.

[0140] Obtain the number of redundant operations. If the number of redundant operations is zero, determine the confidence index boost value based on the redundancy of the stable feature set and the redundancy threshold.

[0141] It should be noted that the first redundancy deviation value is obtained by subtracting the redundancy threshold from the redundancy of the stable feature set. In this embodiment, the first redundancy deviation value is used as the basis for determining the confidence index improvement value. The specific logic is as follows: a confidence improvement mapping rule or function model is established in advance through offline experiments or historical data analysis; the optimal confidence adjustment range corresponding to different redundancy deviation values ​​is statistically analyzed; and the results are organized into an interval mapping table or continuous function model for quick lookup or calculation during actual operation. During runtime, when the first redundancy deviation value is small, it indicates that the feature set has less redundancy, and the confidence index improvement value is increased accordingly to enhance the model's dependence on high-quality features and its discrimination stability.

[0142] If the number of redundant operations is not zero, the confidence reduction value is determined based on the number of redundant operations.

[0143] It should be noted that in this embodiment, the confidence reduction value is determined based on the number of redundant operations on the stable feature set. Specifically, firstly, the actual number of redundant operations performed during the current redundancy adjustment process is counted and compared with a preset operation count threshold to quantify the redundancy adjustment intensity of the feature set. Subsequently, based on this operation count, the corresponding confidence reduction value is searched or calculated from a pre-set confidence adjustment mapping rule or function model. The specific logic is as follows: when the number of redundant operations is high, it indicates that there is high redundancy in the feature set and the adjustment amplitude is large. To avoid excessive interference from redundant features in the model's judgment, the confidence reduction value is increased accordingly, so that the classifier reduces its dependence on highly redundant features in subsequent predictions.

[0144] The confidence score is determined based on the confidence increase or decrease value.

[0145] It should be noted that the initial confidence score is obtained by adding the confidence increase or confidence decrease value.

[0146] If the confidence score is higher than the preset confidence threshold in the database, the predicted category will be output directly.

[0147] It should be added that the preset confidence threshold is based on historical data statistics and operational patterns to ensure the effectiveness and stability of the confidence score. The specific process is as follows: During multiple training or validation cycles, the distribution of confidence scores under different sample, image, or load conditions is collected; then, these data are statistically analyzed to extract the mean, typical intervals, and extreme value distributions, and preliminary thresholds are determined according to the importance of different indicators and fault tolerance requirements; then, the final threshold can be generated using the average, weighted average, or percentile methods, and a safety margin is added on it to prevent abnormal samples from affecting the judgment; finally, the generated threshold is stored in the configuration for subsequent confidence score adjustment and judgment, realizing robust control of the model in actual operation.

[0148] It should be noted that a confidence score higher than the preset confidence threshold in the database means that the classifier's judgment result on the current input feature has high credibility. In other words, the model believes that the predicted category or output result is highly reliable and the possibility of incorrect prediction is low. At this time, the prediction result can be directly used for decision-making or output.

[0149] If the confidence score is lower than or equal to the confidence threshold, the sample is marked as an uncertain result and manual confirmation information is generated.

[0150] It should be noted that when the confidence score is lower than or equal to the preset confidence threshold, the prediction result of the sample is deemed unreliable and is therefore marked as an uncertain result. At the same time, corresponding manual confirmation information is generated to prompt manual review or verification of the prediction result of the sample to ensure the accuracy and security of the decision.

[0151] In this embodiment, the manual confirmation information could be: "Attention! The output category confidence score is lower than or equal to the confidence threshold."

[0152] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0153] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0154] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0155] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0156] Those skilled in the art will 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, or a combination of computer software and electronic hardware. 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.

[0157] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0158] In the embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0159] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0160] In addition, the functional units in the various embodiments of the present invention 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.

[0161] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0162] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An intelligent classification and prediction method for biometric features based on adaptive weighted multimodal fusion, characterized in that, The method includes: After acquiring multimodal biometric features, dimensional alignment is performed to obtain an initial feature set, which is then screened to obtain a stable feature set, and feature fusion of the stable feature set is performed. The process of obtaining a stable feature set through screening includes: obtaining the feature selection results of the initial feature set by passing multiple rounds of random sampling and cross-validation on the initial feature set, and performing frequency statistics and weighted variance analysis on the feature selection results obtained in each round of training to calculate the selection probability of the feature among different feature selection results; If a feature has a selection probability greater than or equal to a preset selection probability threshold and a weight variance less than or equal to a weight variance threshold in multiple training sessions, it is marked as a stable feature; otherwise, it is marked as an unstable feature. Unstable features are removed from the initial feature set to obtain a stable feature set; The redundancy detection interval is determined, and the redundancy of the stable feature set is detected during the feature fusion process of the stable feature set. The redundancy adjustment strategy is determined and executed, including: after obtaining the stable feature set, obtaining the number of features in the stable feature set. The redundancy detection interval of the stable feature set is determined based on the number of features in the stable feature set, and the redundancy detection of the stable feature set is started based on the detection interval. Collect redundant dependency parameters of the stable feature set, including average cross-correlation coefficient, average mutual information, and average variance inflation factor; The redundancy of the stable feature set is analyzed based on the redundancy dependency parameter. The redundancy of the stable feature set is a quantitative indicator of the overall redundancy information of the stable feature set, which is jointly expressed by the cross-correlation number, mutual information, and average variance inflation factor. The specific analysis process is as follows: the average cross-correlation number, average mutual information, and average variance inflation factor are weighted by combining their respective contribution coefficients to obtain the redundancy of the stable feature set; and the redundancy threshold is extracted. If the redundancy of the stable feature set is greater than or equal to the redundancy threshold, the redundancy adjustment strategy is determined to be high redundancy compression. High redundancy compression divides the stable feature set into a highly correlated feature subset, determines a compressible subset within the highly correlated feature subset, and performs feature fusion after compression. Among them, the Pearson correlation coefficient between features is calculated based on the stable feature set and is denoted as the feature correlation coefficient. Features whose feature correlation coefficient is greater than or equal to the preset feature correlation coefficient threshold in the database are grouped into the same set and denoted as the highly correlated feature subset. If the redundancy of the stable feature set is less than the redundancy threshold, the feature fusion process of the stable feature set continues. After completing the feature fusion of the stable feature set, the redundancy of the stable feature set is detected again, and the redundancy secondary adjustment strategy is determined and executed. After receiving the biometric intelligent classification prediction executable signal, the system performs biometric intelligent classification prediction and generates confidence scores and prediction results.

2. The adaptive weighted multimodal fusion-based intelligent classification and prediction method for biometric features according to claim 1, characterized in that, After acquiring multimodal biometrics, dimensional alignment is performed to obtain an initial feature set. The specific analysis method is as follows: A1 performs normalization processing on the input multimodal biometrics, mapping features of different dimensions to a normalized numerical range, thereby obtaining the normalized feature values. A2 calculates the mean and standard deviation of the eigenvalues ​​based on each normalized eigenvalue, models each normalized eigenvalue as a Gaussian distribution, and generates a smooth feature representation through reparameterized sampling; A3 embeds the smooth feature representation into the mapping layer and projects it into a unified low-dimensional latent space. The dimension of the low-dimensional latent space is determined based on the mean of the feature values. A4, based on the dimensions of the low-dimensional latent space, performs dimension alignment to obtain dimension-aligned multimodal biological features, which are denoted as the initial feature set.

3. The adaptive weighted multimodal fusion-based intelligent classification and prediction method for biometric features according to claim 1, characterized in that, The specific analysis method for the high-redundancy compression is as follows: Each highly correlated feature subset is obtained based on the stable feature set; Each highly relevant feature subset is compressed, and the basic compression ratio is determined by the deviation between the redundancy of the stable feature set and the redundancy threshold. Based on the redundancy operation counter, the number of redundant operations is obtained, and it is determined whether the number of redundant operations is zero. If the number of redundant operations is zero, the basic compression ratio is recorded as the execution compression ratio. If the number of redundant operations is not zero and the number of redundant operations is greater than the threshold for the number of redundant operations, an early warning message is generated. If the number of redundant operations is not zero and the number of redundant operations is less than or equal to the threshold of the number of redundant operations, query the cumulative adjustment degree, determine the basic compression ratio optimization value based on the cumulative adjustment degree and the number of redundant operations, and obtain the execution compression ratio based on the basic compression ratio optimization value and the basic compression ratio. Based on the preset number of features to retain, select compressible subsets; The compression ratio is evenly distributed to the compressible subset to perform high-redundancy compression. After the high-redundancy compression is completed, the result is recorded in the log, and the number of redundant operations in the redundancy operation counter is incremented by 1.

4. The adaptive weighted multimodal fusion-based intelligent classification and prediction method for biometric features according to claim 1, characterized in that, The redundancy detection of the stable feature set is performed again to determine the secondary redundancy adjustment strategy. The specific analysis method is as follows: The redundancy of the stable feature set is detected again to obtain the redundancy of the stable feature set, which is denoted as the first redundancy. If the first redundancy is less than the redundancy threshold, the redundancy secondary adjustment strategy is recorded as no adjustment is needed, the fused features are recorded as the final fused features, and an executable signal for intelligent classification and prediction of biometric features is generated. If the first redundancy is still greater than or equal to the redundancy threshold, the number of redundant operations is extracted. If the number of redundant operations is less than the redundancy operation threshold, the redundancy secondary adjustment strategy is recorded as the dimensionality reduction strategy. If the number of redundant operations is greater than or equal to the threshold for the number of redundant operations, then the secondary adjustment strategy for redundancy is recorded as the freezing strategy.

5. The adaptive weighted multimodal fusion-based intelligent classification and prediction method for biometric features according to claim 4, characterized in that, The specific implementation method of the dimensionality reduction strategy is as follows: The first redundancy and the redundancy threshold are deviated to obtain the first redundancy deviation value; The number of dimensionality reductions is determined based on the first redundancy deviation value; Based on the preset number of features to retain, select dimensionality-reducible subsets from each subset of highly relevant features; Obtain the number of features of the reducible subsets, and sort the reducible subsets from largest to smallest based on this number; The dimensionality reduction process of the dimensionality-reducible subset is executed based on the order of the number of features in the dimensionality-reducible subset and the number of dimensionality reductions. After the dimensionality reduction is completed, the feature fusion of the stable feature set is re-executed to obtain the final fused features and generate an executable signal for intelligent classification and prediction of biometric features.

6. The adaptive weighted multimodal fusion-based intelligent classification and prediction method for biometric features according to claim 4, characterized in that, The freezing strategy is implemented as follows: Stop performing intelligent biometric classification and prediction, generate early warning information, and simultaneously start dynamic threshold learning; The specific process of initiating dynamic threshold learning is as follows: based on the number of redundant operations, calculate the average decrease slope of the redundancy of the stable feature set after the redundant operations. If the average decrease slope of redundancy is less than or equal to the preset average decrease slope threshold, the increase value of the redundancy threshold is determined based on the number of redundancy operations. If the average decrease slope of redundancy is greater than the preset average decrease slope threshold, the redundancy threshold reduction value is determined based on the deviation between the decrease slope and the average decrease slope threshold. Using an exponentially weighted average, based on the increase or decrease of the redundancy threshold, a smooth update of the redundancy threshold is performed to generate the redundancy threshold to be used in the next round of adjustment, and the update record is written to the log.

7. The adaptive weighted multimodal fusion-based intelligent classification and prediction method for biometric features according to claim 1, characterized in that, The process of performing intelligent classification and prediction based on biometric features, generating confidence scores and prediction results, is analyzed in the following manner: After receiving the biometric intelligent classification and prediction executable signal, the final fused features are input into the trained classifier for recognition and prediction. The classifier calculates the probability values ​​of each category based on the spatial distribution and pattern differences of the fused features, thereby generating the corresponding prediction results and initial confidence scores; Obtain the number of redundant operations. If the number of redundant operations is zero, determine the confidence index boost value based on the redundancy of the stable feature set and the redundancy threshold. If the number of redundant operations is not zero, the confidence reduction value is determined based on the number of redundant operations. The confidence score is determined based on the confidence increase or confidence decrease value; If the confidence score is higher than the preset confidence threshold, the predicted category will be output directly. If the confidence score is lower than or equal to the confidence threshold, the sample is marked as an uncertain result and manual confirmation information is generated.

Images