Dynamic model calibration method and device based on simplex equiangular tight frame classifier

Abstract

The invention provides a dynamic model calibration method and device based on a simplex equiangular tight frame classifier, and the method comprises the steps: converting an input sample into a high-dimensional feature through a shared feature extraction module, and generating two groups of independent initial probability distribution vectors through a standard trainable classifier and a confidence adjustable module; performing weighted fusion on the two groups of probability distribution vectors, and dynamically adjusting a calibration weight factor according to the difference between the minimum prediction confidence and the actual accuracy; respectively calculating classification loss of the two groups of probability distribution vectors, and performing weighted fusion on the two groups of probability distribution vectors based on a calibration weight factor to obtain a total loss function for optimizing the network; and based on the calibration weight factor obtained by training, carrying out weighted fusion on the output of the standard trainable classifier and the confidence adjustable module to generate final prediction probability distribution. The method is applied to image classification of the deep neural network, the model calibration error can be minimized, the model prediction accuracy can be improved, and a prediction result with higher quality can be obtained.

Description

Technical Field

[0001] The present disclosure relates to the technical field of image classification using deep neural networks, and in particular to a dynamic model calibration method and device based on a simplex equiangular tight frame classifier. Background Art

[0002] In recent years, deep neural networks have achieved remarkable success in many fields such as image classification, object detection, and natural language processing. However, its reliability issues in practical applications have gradually emerged, especially the deviation between the model's predicted probability and the actual accuracy. This phenomenon of inconsistency between predicted probability and actual accuracy is called calibration error, which is usually manifested as the model's overconfidence in its output results. Calibration error poses severe challenges in many application scenarios that rely on model reliability, such as medical diagnosis and autonomous driving. The model's incorrect assessment of uncertainty may lead to serious safety hazards.

[0003] In order to solve this problem, improving the confidence calibration level of the model has become an important direction of deep learning research. Existing model calibration methods are mainly divided into two categories: post-processing methods and training-time methods. Post-processing methods adjust or rescale the output probability after the model training is completed without changing the internal characteristics of the model; training-time methods integrate the calibration mechanism during the training process to directly optimize the confidence estimation of the model, thereby improving the generalization performance. Recent studies have shown that training-time calibration methods are closely related to classifiers. Several training-time methods reduce the risk of overfitting by introducing weakly trained classification heads, thereby improving calibration performance. However, this type of method also has significant shortcomings. On the one hand, weak training may reduce the expressive power of the model on the training set, thereby affecting the prediction accuracy; on the other hand, these methods perform poorly when processing out-of-distribution samples, especially in tasks that require extensive generalization capabilities, which limits the practical application effect of the model.

[0004] Therefore, a new training-time calibration method is urgently needed. This method should avoid relying on weakly trained classification heads, and at the same time significantly reduce the calibration error and improve the generalization ability while ensuring the accuracy of the model, so as to overcome the limitations of existing methods in practical applications. Summary of the invention

[0005] In order to overcome the above-mentioned deficiencies of the prior art, the present disclosure provides a dynamic model calibration method and device based on a simplex equiangular tight frame classifier.

[0006] According to a first aspect of the present disclosure, a dynamic model calibration method based on a simplex equiangular tight frame classifier is provided, characterized in that it comprises the following steps:

[0007] S1. Obtain an original input sample, and convert the original input sample into a high-dimensional feature representation through a shared feature extraction module;

[0008] S2, inputting the high-dimensional features into a standard trainable classifier and a confidence adjustable module respectively, and outputting two independent probability distribution vectors as initial probability distribution vector results;

[0009] S3, performing weighted fusion on the two initial probability distribution vectors, and dynamically adjusting the calibration weight factor according to minimizing the difference between the prediction confidence and the actual accuracy of the fusion output;

[0010] S4, respectively calculating the classification losses of the two initial probability distribution vectors, and using the calibration weight factor to weight the two parts of the loss to obtain a total loss function for training optimization;

[0011] S5. Based on the calibration weight factor, weighted fusion is performed on the output probability distribution of the standard trainable classifier and the output probability distribution of the confidence adjustment module to generate and output a final classification result;

[0012] Preferably, the confidence adjustable module in S2 is composed of a fixed simplex equiangular tight frame classifier and an adapter, wherein the fixed simplex equiangular tight frame classifier can adjust its output confidence interval in the initialization stage, and the adapter is located in front of the fixed simplex equiangular tight frame classifier to process the input high-dimensional features to alleviate the conflict caused by the difference in feature requirements between the fixed simplex equiangular tight frame classifier and the standard trainable classifier;

[0013] The adapter adopts a bottleneck architecture, which includes two convolutional layers. The first convolutional layer reduces the dimension to compress the features, and the second convolutional layer restores the feature dimension to keep the output consistent with the input.

[0014] The high-dimensional features are processed by the adapter and then input into the fixed simplex equiangular tight frame classifier, and finally the output probability distribution vector of the confidence adjustable module is obtained;

[0015] Preferably, the fixed simplex equiangular tight frame classifier is obtained by randomly generating a simplex equiangular tight frame and fixing it, wherein the fixed simplex equiangular tight frame is independent of the gradient propagation of the network; for the K-classification problem, the fixed simplex equiangular tight frame is defined as a The set of vectors in space, represented by the matrix The columns are composed of , d is the dimension of the input feature, and the calculation method includes:

[0016]

[0017] In the formula, β is the scaling factor, U is a partially orthogonal matrix, that is, UT U=I, and d≥K, I is the identity matrix, is a vector of all 1s; at this time, the confidence of the final predicted label output by the fixed simplex equiangular tight frame classifier is positively correlated with the scaling factor β and negatively correlated with the number of categories K; the scaling factor β in the fixed simplex equiangular tight frame classifier is initialized to 1.0;

[0018] Preferably, the process of obtaining the two independent initial probability distribution vectors in S2 includes: obtaining the high-dimensional feature representation output by the shared feature extraction module; passing the high-dimensional feature representation through the standard trainable classifier to generate the original probability distribution vector p Sta ; The high-dimensional feature representation is passed through the confidence adjustable module to generate a calibration probability distribution vector p ETF ; Record two groups of probability distribution vectors respectively;

[0019] Preferably, the process of weighted fusion of the two initial probability distribution vectors in S3 includes:

[0020] p fud =γp Sta +(1-γ)p ETF ,

[0021] In the formula, p fused represents the mixed result of the two initial probability distribution vectors, γ∈[0,1] is a weight factor controlling the contribution of the two initial probability distribution vectors;

[0022] Preferably, the process of dynamically adjusting the calibration weight factor in S3 includes: in training round t, using different weight factors to mix the two sets of initial probability distribution vectors recorded in the round, and searching for the calibration weight γ of the t+1th training round by minimizing the difference between the prediction confidence and the actual accuracy. t+1 , the calculation methods include:

[0023]

[0024] In the formula, D = {(x, y)} represents the data set consisting of the input sample x and its corresponding true label y. δ is a careful hyperparameter used to balance the prediction accuracy and confidence of the model. t (D,γ) and acc t (D,γ) respectively represent the mixed output p calculated based on the weight factor γ in the tth training round fused The prediction confidence and actual accuracy of , the calculation method includes:

[0025]

[0026]

[0027] In the formula, It is an indicator function, which takes the value 1 when the condition in the brackets is met, otherwise it takes the value 0;

[0028] Preferably, in S4, the specific strategy of the total loss function includes: the total loss function is composed of the original classification loss and the calibration classification loss, the parameters {θ, W, φ} of the shared feature extractor, the standard trainable classifier and the adapter in the confidence adjustable module are optimized by minimizing the loss functions of both the original classification loss and the calibration classification loss, and combined with the calibration weight factor γ, and the calculation method includes:

[0029]

[0030] Loss Function represents the classification loss of the standard trainable classifier, and the calculation method includes:

[0031]

[0032] Loss Function represents the classification loss of the confidence adjustable module, and the calculation method includes:

[0033]

[0034] Preferably, the process of generating the final classification result in S5 includes: during training, using the optimal model with the lowest loss function value selected through multiple rounds of training iterations and based on performance evaluation on the validation set and obtaining the calibration weight factor γ of the corresponding round of the optimal model for reasoning, weighted fusion of the probability distribution vector output by the standard trainable classifier and the confidence adjustment module, the calculation method includes:

[0035] p=γp Sta +(1-γ)p ETF

[0036] Where p is the probability distribution vector of the final output of the model.

[0037] According to a second aspect of the present disclosure, a dynamic model calibration device based on a simplex equiangular tight frame classifier is provided, the device comprising:

[0038] A shared feature extractor module, used for obtaining original input samples and converting the original input samples into high-dimensional feature representations through the shared feature extraction module;

[0039] An initial probability distribution vector acquisition module is used to input the high-dimensional features into a standard trainable classifier and a confidence adjustable module respectively, and output two independent probability distribution vectors as initial probability distribution vector results;

[0040] A calibration weight factor dynamic adjustment module, used for weighted fusion of the two initial probability distribution vectors, and dynamically adjusting the calibration weight factor according to minimizing the difference between the prediction confidence and the actual accuracy of the fusion output;

[0041] A total loss function determination module, used to calculate the classification losses of the two initial probability distribution vectors respectively, and use the calibration weight factor to weight the two parts of the loss to obtain a total loss function for training optimization;

[0042] The classification result weighted output module is used to perform weighted fusion on the output probability distribution of the standard trainable classifier and the output probability distribution of the confidence adjustment module based on the calibration weight factor, generate a final classification result and output it.

[0043] The principle of the technical solution disclosed in this disclosure is:

[0044] In the training phase, the input samples are converted into high-dimensional features through a shared feature extraction module, and then two independent sets of initial probability distribution vectors are generated through a standard trainable classifier and an adjustable confidence module, which are used for subsequent calibration; the two sets of probability distribution vectors are weightedly fused, and the calibration weight factor is dynamically adjusted by minimizing the difference between the output prediction confidence and the actual accuracy; the classification loss is calculated for the two sets of probability distribution vectors respectively, and then they are weighted and fused based on the calibration weight factor to obtain the total loss function for optimizing the network; in the inference phase, the outputs of the standard trainable classifier and the adjustable confidence module are weighted and fused based on the calibration weight factor obtained through training to generate the final prediction probability distribution.

[0045] The technical effects of the present disclosure are:

[0046] The present invention utilizes the characteristic that the output confidence interval of a fixed simplex equiangular tight frame classifier can be adjusted during the initialization phase to generate a low-confidence calibration probability distribution that is opposite to the common overconfidence problem in the original probability distribution output by a standard trainable classifier, but has similar prediction results. Through a dynamic mixing module, the overconfident original probability distribution is mixed with the low-confidence calibration probability distribution, and the calibration error is minimized by dynamically adjusting the weight factor. By combining a standard trainable classifier with an adjustable confidence module, high-quality classification results are obtained. By adopting the method of the present invention, the calibration error can be effectively reduced on the basis of improving the prediction accuracy of the model, and the limitations of the existing classifier-based calibration method in generalization ability can be solved, thereby improving the practical application effect of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which:

[0048] Figure 1 A flow chart of a dynamic model calibration method based on a simplex equiangular tight frame classifier provided by the present disclosure;

[0049] Figure 2 A process framework diagram of a dynamic model calibration method based on a simplex equiangular tight frame classifier provided by the present disclosure;

[0050] Figure 3 A structural schematic diagram of a dynamic model calibration device based on a simplex equiangular tight frame classifier provided by the present disclosure;

[0051] Figure 4 A t-SNE visualization diagram of the last layer of features of a dynamic model calibration method based on a simplex equiangular tight frame classifier provided by the present disclosure on the CIFAR-10 dataset;

[0052] Figure 5 A schematic diagram of class activation on the Tiny-ImageNet dataset of a standard trainable classifier module and a confidence adjustable module of a dynamic model calibration device based on a simplex equiangular tight frame classifier provided in the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present disclosure clearer, the technical solution in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0054] like Figure 1 and Figure 2 As shown, in one embodiment, a dynamic model calibration method based on a simplex equiangular tight frame classifier is provided, comprising the following steps:

[0055] S1. Obtain an original input sample, and convert the original input sample into a high-dimensional feature representation through a shared feature extraction module;

[0056] The input sample is taken from the training data set D = {(x, y)} consisting of the input sample x and its corresponding true label y;

[0057] The input sample is input into a shared feature extraction module to extract a d-dimensional high-dimensional feature. The shared feature extraction module is composed of an input layer and multiple residual modules. The input layer includes a convolution kernel and a maximum pooling layer. Each residual block includes two or three convolution kernels, batch normalization BN and activation function ReLU. All residual blocks are connected through shortcuts to add the input directly to the output.

[0058] S2, inputting the high-dimensional features into a standard trainable classifier and a confidence adjustable module respectively, and outputting two independent probability distribution vectors as initial probability distribution vector results;

[0059] The high-dimensional features are input into a standard trainable classifier to generate the original probability distribution vector p Sta , the standard trainable classifier comprises a global average pooling layer and a fully connected layer;

[0060] The high-dimensional features are input into the confidence adjustable module to generate a calibration probability distribution vector p ETF The confidence adjustable module is composed of an adapter and a simplex equiangular tight frame classifier, the adapter includes two convolution kernels, batch normalization BN and activation function ReLU, and the simplex equiangular tight frame classifier includes a global average pooling layer and a fixed simplex equiangular tight frame matrix K is the number of categories;

[0061] Record two sets of probability distribution vectors respectively;

[0062] S3, performing weighted fusion on the two initial probability distribution vectors, and dynamically adjusting the calibration weight factor according to minimizing the difference between the prediction confidence and the actual accuracy of the fusion output;

[0063] In training round t, two sets of initial probability distribution vectors recorded in this round are obtained;

[0064] The two sets of initial probability distribution vectors are mixed using different weight factors:

[0065] p fused =γp Sta +(1-γ)p ETF ,

[0066] In the formula, p fused represents the mixed result of two initial probability distribution vectors, γ∈[0,1] is the weight factor that controls the contribution of both;

[0067] The calculated mixed output p fusedThe prediction confidence and actual accuracy of , the calculation method includes:

[0068]

[0069] In the formula, conf t (D,γ) and acc t (D,γ) respectively represent the mixed output p fused The prediction confidence and actual accuracy of It is an indicator function, which takes the value 1 when the condition in the brackets is met, otherwise it takes the value 0;

[0070] By minimizing the difference between the predicted confidence and the actual accuracy, the calibration weight γ for the t+1th training round is searched t+1 , the calculation methods include:

[0071]

[0072] In the formula, is a careful hyperparameter used to balance the predictive accuracy and confidence of the model;

[0073] S4, respectively calculating the classification losses of the two initial probability distribution vectors, and using the calibration weight factor to weight the two parts of the loss to obtain a total loss function for training optimization;

[0074] Obtaining the two sets of initial probability distribution vectors;

[0075] Calculate the original classification loss using the original probability distribution vector The calculation methods include:

[0076]

[0077] The calibration probability distribution vector is used to calculate the calibration classification loss, which is calculated by:

[0078]

[0079] The total loss function is obtained by weighted fusion of the original classification loss and the calibration classification loss. The model parameters are optimized by minimizing the two loss functions and combining the calibration weight factor γ. The calculation method includes:

[0080]

[0081] where {θ, W, φ} denotes the parameters of the shared feature extractor, the standard trainable classifier, and the adapter in the confidence-adjustable module, respectively;

[0082] S5. Based on the calibration weight factor, weighted fusion is performed on the output probability distribution of the standard trainable classifier and the output probability distribution of the confidence adjustment module to generate and output a final classification result;

[0083] The optimal model and its corresponding round calibration weight factor γ are used for reasoning, and the probability distribution vector output by the weighted fusion standard trainable classifier and the confidence adjustment module is calculated by:

[0084] p=γp Sta +(1-γ)p ETF .

[0085] Where p is the probability distribution vector of the final output of the model.

[0086] In some embodiments of the present disclosure, experiments are conducted on three datasets, CIFAR-10, CIFAR-100, and SVHN, and the results obtained using the method disclosed in the present disclosure are compared with the results obtained using other training methods for model calibration in the prior art, especially compared with the most relevant classifier design-based methods (PLP, TST, and V-TST) in the present disclosure. In order to objectively evaluate the comparison results, the results of the method are presented by three indicators when calculating each training, and the results are shown in Table 1. The smaller the objective evaluation index values ​​ECE and AECE in the table, the better the calibration effect, and the larger the ACC, the higher the classification accuracy. It can be observed from Table 1 that the calibration performance of the method disclosed in the present disclosure always obtains the best results, and its classification accuracy is also improved compared to Vanilla, which does not use any model calibration method during training.

[0087] Table 1

[0088]

[0089] The present disclosure also provides experiments on out-of-distribution sample detection, training a model on a data set (e.g., CIFAR-10) and testing it on other data sets (e.g., CIFAR-100 and SVHN). The results obtained using the method disclosed herein are compared with the results of other model calibration methods (PLP, TST, and V-TST) based on classifier design in the prior art. In order to objectively evaluate the comparison results, the results of each model method during training are calculated and presented by two indicators, and the results are shown in Table 2. The larger the objective evaluation index value AUROC in the table, the better the generalization performance, and the smaller the FPR95, the better the generalization performance. It can be observed from Table 2 that in some embodiments, the generalization performance of the method disclosed herein always obtains the best results.

[0090] Table 2

[0091]

[0092]

[0093] In some embodiments, Figure 3 As shown, a dynamic model calibration device 100 based on a simplex equiangular tight frame classifier is provided, comprising:

[0094] A shared feature extractor module 10, used to obtain original input samples and convert the original input samples into high-dimensional feature representations through a shared feature extraction module;

[0095] An initial probability distribution vector acquisition module 20 is used to input the high-dimensional features into a standard trainable classifier and a confidence adjustable module respectively, and output two independent probability distribution vectors as initial probability distribution vector results;

[0096] A calibration weight factor dynamic adjustment module 30, configured to perform weighted fusion on the two initial probability distribution vectors, and dynamically adjust the calibration weight factor according to minimizing the difference between the prediction confidence and the actual accuracy of the fusion output;

[0097] A total loss function determination module 40 is used to calculate the classification losses of the two initial probability distribution vectors respectively, and use the calibration weight factor to weight the two parts of the loss to obtain a total loss function for training optimization;

[0098] The classification result weighted output module 50 is used to perform weighted fusion on the output probability distribution of the standard trainable classifier and the output probability distribution of the confidence adjustment module based on the calibration weight factor, generate and output the final classification result.

[0099] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, ordinary technicians in this field may make various modifications and improvements to the technical solutions of the present invention, which should all fall within the scope of protection determined by the claims of the present invention.

Claims

1. A dynamic model calibration method based on a simplex equiangular tight frame classifier, characterized in that: The following steps are involved: S1. Obtain an original input sample, and convert the original input sample into a high-dimensional feature representation through a shared feature extraction module; S2, inputting the high-dimensional features into a standard trainable classifier and a confidence adjustable module respectively, and outputting two independent probability distribution vectors as initial probability distribution vector results; S3, performing weighted fusion on the two initial probability distribution vectors, and dynamically adjusting the calibration weight factor according to minimizing the difference between the prediction confidence and the actual accuracy of the fusion output; S4, respectively calculating the classification losses of the two initial probability distribution vectors, and using the calibration weight factor to weight the two parts of the loss to obtain a total loss function for training optimization; S5. Based on the calibration weight factor, perform weighted fusion on the output probability distribution of the standard trainable classifier and the output probability distribution of the confidence adjustment module to generate and output a final classification result.

2. The dynamic model calibration method based on simplex equiangular tight frame classifier according to claim 1 is characterized in that: The confidence adjustable module in S2 is composed of a fixed simplex equiangular tight frame classifier and an adapter, wherein the fixed simplex equiangular tight frame classifier can adjust its output confidence interval in the initialization stage, and the adapter is located in front of the fixed simplex equiangular tight frame classifier to process the input high-dimensional features to alleviate the conflict caused by the difference in feature requirements between the fixed simplex equiangular tight frame classifier and the standard trainable classifier; The adapter adopts a bottleneck architecture, which includes two convolutional layers. The first convolutional layer reduces the dimension to compress the features, and the second convolutional layer restores the feature dimension to keep the output consistent with the input. The high-dimensional features are processed by the adapter and then input into the fixed simplex equiangular tight frame classifier, and finally the output probability distribution vector of the confidence adjustable module is obtained.

3. The dynamic model calibration method based on simplex equiangular tight frame classifier according to claim 1 is characterized in that: The fixed simplex equiangular tight frame classifier is obtained by randomly generating a simplex equiangular tight frame and fixing it. The fixed simplex equiangular tight frame is independent of the gradient propagation of the network; for the K-classification problem, the fixed simplex equiangular tight frame is defined as a The set of vectors in space, represented by the matrix The columns are composed of , d is the dimension of the input feature, and the calculation method includes: In the formula, β is the scaling factor, U is a partially orthogonal matrix, that is, U T U=I, and d≥K, I is the identity matrix, is a vector of all 1s; at this time, the confidence of the final predicted label output by the fixed simplex equiangular tight frame classifier is positively correlated with the scaling factor β and negatively correlated with the number of categories K; the scaling factor β in the fixed simplex equiangular tight frame classifier is initialized to 1.

0.

4. The dynamic model calibration method based on simplex equiangular tight frame classifier according to claim 1, characterized in that: The process of obtaining the two independent initial probability distribution vectors in S2 includes: obtaining the high-dimensional feature representation output by the shared feature extraction module; passing the high-dimensional feature representation through the standard trainable classifier to generate the original probability distribution vector p Sta ; The high-dimensional feature representation is passed through the confidence adjustable module to generate a calibration probability distribution vector p ETF ; Record two sets of probability distribution vectors respectively.

5. The dynamic model calibration method based on simplex equiangular tight frame classifier according to claim 4 is characterized in that: The process of weighted fusion of the two initial probability distribution vectors in S3 includes: p fused =γp Sta +(1-γ)p ETF , In the formula, p fused represents the mixed result of the two initial probability distribution vectors, and γ∈[0,1] is a weight factor controlling the contribution of both the two initial probability distribution vectors.

6. The dynamic model calibration method based on simplex equiangular tight frame classifier according to claim 5, characterized in that: The process of dynamically adjusting the calibration weight factor in S3 includes: in training round t, two sets of initial probability distribution vectors recorded in the round are mixed using different weight factors, and the calibration weight γ of the t+1th training round is searched by minimizing the difference between the prediction confidence and the actual accuracy. t+1 , the calculation methods include: In the formula, D = {(x, y)} represents the data set consisting of the input sample x and its corresponding true label y. δ is a careful hyperparameter used to balance the prediction accuracy and confidence of the model. t (D,γ) and acc t (D,γ) respectively represent the mixed output p calculated based on the weight factor γ in the tth training round fused The prediction confidence and actual accuracy of , the calculation method includes: In the formula, It is an indicator function, which takes the value 1 when the condition in the brackets is met, otherwise it takes the value 0.

7. The dynamic model calibration method based on simplex equiangular tight frame classifier according to claim 6 is characterized in that: The specific strategy of the total loss function in S4 includes: the total loss function is composed of the original classification loss and the calibration classification loss, the parameters {θ, W, φ} of the shared feature extractor, the standard trainable classifier and the adapter in the confidence adjustable module are optimized by minimizing the loss functions of both the original classification loss and the calibration classification loss, and combined with the calibration weight factor γ, and the calculation method includes: Loss Function represents the classification loss of the standard trainable classifier, and the calculation method includes: Loss Function represents the classification loss of the confidence adjustable module, and the calculation method includes:

8. The dynamic model calibration method based on simplex equiangular tight frame classifier according to claim 7, characterized in that: The process of generating the final classification result in S5 includes: during training, using the optimal model with the lowest loss function value selected through multiple rounds of training iterations and based on performance evaluation on the validation set and obtaining the calibration weight factor γ of the corresponding round of the optimal model for reasoning, weighted fusion of the probability distribution vector output by the standard trainable classifier and the confidence adjustment module, and the calculation method includes: p=γp Sta +(1-γ)p ETF Where p is the probability distribution vector of the final output of the model.

9. A dynamic model calibration device based on a simplex equiangular tight frame classifier, the device comprising: A shared feature extractor module, used for obtaining original input samples and converting the original input samples into high-dimensional feature representations through the shared feature extraction module; An initial probability distribution vector acquisition module is used to input the high-dimensional features into a standard trainable classifier and a confidence adjustable module respectively, and output two independent probability distribution vectors as initial probability distribution vector results; A calibration weight factor dynamic adjustment module, used for weighted fusion of the two initial probability distribution vectors, and dynamically adjusting the calibration weight factor according to minimizing the difference between the prediction confidence and the actual accuracy of the fusion output; A total loss function determination module, used to calculate the classification losses of the two initial probability distribution vectors respectively, and use the calibration weight factor to weight the two parts of the loss to obtain a total loss function for training optimization; The classification result weighted output module is used to perform weighted fusion on the output probability distribution of the standard trainable classifier and the output probability distribution of the confidence adjustment module based on the calibration weight factor, generate a final classification result and output it.

Images