Image classification method and fundus image classification method based on uncertainty measurement

Abstract

The present invention discloses an image classification method based on uncertainty measurement, including obtaining an existing data set and marking the classification; constructing a basic classification network, inputting the data set into the basic classification network and measuring the uncertainty information according to the output; optimizing the basic classification network to obtain a final classification model; using the obtained classification model to classify the target image and outputting the confidence of the corresponding classification result. The present invention also discloses a fundus image classification method including the image classification method based on uncertainty measurement. The present invention continuously overcomes the uncertainty of the model through repeated training based on the uncertainty information of the measurement model, so that the classification result of the model is more accurate and reliable, so the method of the present invention has high reliability and good accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of image processing, and in particular relates to an image classification method and a fundus image classification method based on uncertainty measurement. Background Art

[0002] With the development of economy and technology, artificial intelligence technology has been widely used in people's production and life, bringing endless convenience to people's production and life. Artificial intelligence technology represented by deep learning has great superiority in accuracy, which enables artificial intelligence technology to replace complicated manual methods in some application fields.

[0003] Image classification methods based on artificial intelligence technology have been widely used in clinical medicine, basic medical research, industrial production and other fields. However, since existing image classification methods rely heavily on training data, they often over-interpret the target image data in specific applications, resulting in overconfidence in the classification process. The key factor causing this phenomenon is that existing image classification methods cannot timely discover and measure the uncertainty factors caused by insufficient data and errors in data collection during the prediction process. The existence of these uncertain factors has led to a large degree of blindness in the specific application of existing image classification methods, causing unnecessary misjudgments, which in turn seriously affects the accuracy and reliability of the classification methods. Summary of the invention

[0004] One of the purposes of the present invention is to provide an image classification method based on uncertainty measurement with high reliability and good accuracy.

[0005] A second objective of the present invention is to provide a fundus image classification method including the image classification method based on uncertainty measurement.

[0006] The image classification method based on uncertainty measurement provided by the present invention comprises the following steps:

[0007] S1. Obtain existing data sets and classify and mark them;

[0008] S2. Build a basic classification network, input the data set obtained in step S1 into the basic classification network, and measure the uncertainty information according to the output of the network;

[0009] S3. According to the uncertainty information measured in step S2, the basic classification network is optimized to obtain the final classification model;

[0010] S4. Use the classification model obtained in step S3 to perform image classification of the target image and output the confidence of the corresponding classification result.

[0011] The step S2 specifically includes the following steps:

[0012] A. Use VGG16 as the basic classification network;

[0013] B. Input the data set obtained in step S1 into the basic classification network, and measure the uncertainty information according to the output of the basic classification network;

[0014] C. Sample step B several times to measure deeper uncertainty information.

[0015] The step A specifically comprises the following steps:

[0016] Use VGG16 as the basic classification network;

[0017] The cross entropy loss is used as the basic loss, and the following formula is used as the loss formula:

[0018]

[0019] In the formula Cross entropy loss value; y i is the true label of the i-th sample; p i is the probability of the network's predicted classification for the i-th sample; log() represents the logarithm with base 2.

[0020] The step B specifically converts the output of the fully connected layer of the basic classification network into evidence information, and measures the uncertainty information based on the evidence information; specifically includes the following steps:

[0021] B1. Use the output value of the network for each category as the support for the prediction: For the binary classification task, the output of the network for each category is The class activation value for the network output classified as a positive sample, The network outputs the class activation value classified as a negative sample; in order to ensure the non-negativity of the evidence corresponding to each category, the support opinion value corresponding to each category is passed through the ReLU activation function to generate the classification subjective evidence corresponding to each category The evidence value for the network output classification as a positive sample, Output the evidence value for the network to classify as a negative sample;

[0022] B2. For the binary classification task, the confidence score corresponding to the positive sample is The confidence score corresponding to the negative sample is The uncertainty fraction of the sample is u i , then the confidence score and uncertainty score of the sample satisfy the relationship The evidence corresponding to the uncertainty of each category of the sample is assigned a fixed value of 1, combined with the subjective evidence corresponding to each category Calculate the total evidence S corresponding to the sample i for

[0023] B3. Calculate the predicted uncertainty score u of the sample i for At the same time, the confidence score corresponding to the positive sample is calculated for Confidence scores corresponding to negative samples for

[0024] The step C specifically comprises the following steps:

[0025] C1. Set dropout in the last two fully connected layers of the network, and set the dropout probability to 0.3;

[0026] C2. Perform several random samplings on the fully connected layer parameters to obtain several sets of network output values Thus, the corresponding uncertainty score u is obtained for each set of sampled output values i ; Finally, through sampling, a set of uncertainty scores are collected for each sample Where N is the number of sampling times, is the uncertainty value of the network prediction for the i-th sample at the j-th sampling;

[0027] C3. The uncertainty fraction obtained by sampling, statistical standard deviation σ(U i ), variance σ 2 (U i ) and mean μ(U i ), where variance σ 2 (U i ) is a measure of network uncertainty, and in is the mean value μ(U i ),and

[0028] The step S3 specifically includes the following steps:

[0029] a. Use evidence information to parameterize the Dirichlet distribution, thereby reducing uncertainty;

[0030] b. Minimize the variance, standard deviation, and mean of the network predictions obtained during the sampling process to reduce the uncertainty of the network;

[0031] c. Construct the final constraints to optimize the basic classification network.

[0032] The step a specifically comprises the following steps:

[0033] a1. Obtain the evidence information corresponding to each category obtained in step S2 and α i is the total evidence that the network classifies the i-th sample as a positive sample, β i is the total evidence that the network classifies the i-th sample as a negative sample; then the obtained evidence information is used to parameterize the Dirichlet distribution, and the parameterized distribution is expressed as in is the probability that the network predicts the sample as a positive sample, α i ' is the total evidence corresponding to the true category of the sample being a positive sample, β i ' is the total evidence corresponding to the true category of the sample being a negative sample;

[0034] a2. KL divergence is used to measure the difference between the Dirichlet distribution parameterized by evidence and the Dirichlet distribution under 0 evidence. The difference information is expressed by the following formula:

[0035]

[0036] Where KL(A|B) is the KL divergence value between distribution A and distribution B; is the Dirichlet distribution under 0 evidence;

[0037] a3. Use the following formula to construct the constraints for network training and overcome the uncertainty information in the network by minimizing the constraints:

[0038]

[0039] Where L i EDL (θ) is the constraint term for network training; θ represents all network parameters; y i is the true label of the i-th sample; p i is the network prediction probability of the i-th sample; || ||2 is the sum of squares; D(x|α i ,β i ) is the i-th sample with respect to the continuous variable x, and the parameter is (α i ,β i ) of the Dirichlet distribution; λ is a dynamically changing coefficient, and t is the current training epoch number; when y i =0<α i ′,β i′>=<1,β i >, when y i =1<α i ′,β i ′>=<α i ,1>.

[0040] The step b specifically comprises the following steps:

[0041] b1. Set the standard deviation σ(U i ), variance σ 2 (U i ) and mean μ(U i ) as the optimization target;

[0042] b2. Use the following formula as the optimization target formula:

[0043]

[0044] Where λ1, λ2 and λ3 are the set parameters; To optimize the target value, the uncertainty in the sampling process is reduced by minimizing the target value.

[0045] The step c specifically comprises the following steps:

[0046] The final constraint term Loss is constructed using the following formula:

[0047] Loss = L i CE +L i EDL (θ)+L i MS

[0048] Where L i CE is the cross entropy loss corresponding to the i-th sample; L i EDL (θ) is the constraint term for the constructed network training; L i MS The optimization target value for the construction is minimized, thereby improving the reliability of the network confidence prediction.

[0049] The S4 specifically comprises the following steps:

[0050] Obtain the support value output by the network's fully connected layer, and calculate the subjective evidence for each category as The total evidence for the sample classification is The prediction confidence probability corresponding to each class is obtained as and The final prediction confidence of the sample is

[0051] like The corresponding predicted classification category is the positive sample category; otherwise, the corresponding predicted classification category is the negative sample category.

[0052] The present invention also discloses a fundus image classification method including the image classification method based on uncertainty measurement, which specifically includes the following steps:

[0053] (1) Obtain an existing fundus image dataset;

[0054] (2) using the image classification method based on uncertainty measurement to learn the fundus image dataset obtained in step (1) to obtain a final fundus image classification model;

[0055] (3) Using the fundus image classification model obtained in step (2), classify the actual fundus image and output the confidence of the final classification result.

[0056] The image classification method and fundus image classification method based on uncertainty measurement provided by the present invention continuously overcome the uncertainty of the model through repeated training on the basis of the uncertainty information of the measurement model, so that the classification result of the model is more accurate and reliable. Therefore, the method of the present invention has high reliability and good accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 The figure is a flowchart of the image classification method of the present invention.

[0058] Figure 2 The figure is a schematic diagram of the method flow of the fundus image classification method of the present invention. DETAILED DESCRIPTION

[0059] like Figure 1 The method flow chart of the image classification method of the present invention is shown as follows: The image classification method based on uncertainty measurement provided by the present invention comprises the following steps:

[0060] S1. Obtain existing data sets and classify and mark them;

[0061] S2. Construct a basic classification network, input the data set obtained in step S1 into the basic classification network, and measure the uncertainty information according to the output of the network; specifically, the following steps are included:

[0062] A. Use VGG16 as the basic classification network; specifically include the following steps:

[0063] VGG16 is used as the basic classification network to provide basic classification performance for the model;

[0064] The cross entropy loss is used as the basic loss, and the following formula is used as the loss formula:

[0065]

[0066] In the formula Cross entropy loss value; y i is the true label of the i-th sample, which takes a value of 0 or 1; p i The network predicts the classification probability for the i-th sample; log() represents the logarithm with base 2;

[0067] B. Input the data set obtained in step S1 into the basic classification network, and measure the uncertainty information according to the output of the basic classification network; specifically, convert the output of the fully connected layer of the basic classification network into evidence information, and measure the uncertainty information based on the evidence information; specifically, the following steps are included:

[0068] B1. Use the output value of the network for each category as the support for the prediction: For the binary classification task, the output of the network for each category is The class activation value for the network output classified as a positive sample, The network outputs the class activation value classified as a negative sample; in order to ensure the non-negativity of the evidence corresponding to each category, the support opinion value corresponding to each category is passed through the ReLU activation function to generate the classification subjective evidence corresponding to each category The evidence value for the network output classification as a positive sample, Output the evidence value for the network to classify as a negative sample;

[0069] B2. For the binary classification task, the confidence score corresponding to the positive sample is The confidence score corresponding to the negative sample is The uncertainty fraction of the sample is u i , then the confidence score and uncertainty score of the sample satisfy the relationship The evidence corresponding to the uncertainty of each category of the sample is assigned a fixed value of 1, combined with the subjective evidence corresponding to each category Calculate the total evidence S corresponding to the sample i for

[0070] B3. Calculate the predicted uncertainty score u of the sample i for At the same time, the confidence score corresponding to the positive sample is calculated for Confidence scores corresponding to negative samples for

[0071] C. Perform several samplings on step B to measure deeper uncertainty information; specifically, the following steps are included:

[0072] C1. Share the parameter weights of the convolutional layer of the network to ensure the consistency of the parameters of the convolutional layer, and perform multiple random sampling of the measured uncertainty scores in the fully connected layer; in specific implementation, set dropout in the last two fully connected layers of the network; set the dropout probability to 0.3, which means that each parameter in the fully connected layer has a 0.3 probability of not participating in the calculation;

[0073] C2. Perform several random samplings on the fully connected layer parameters to obtain several sets of network output values Thus, the corresponding uncertainty score u is obtained for each set of sampled output values i ; Finally, through sampling, a set of uncertainty scores are collected for each sample Where N is the number of sampling times, is the uncertainty value of the network prediction for the i-th sample at the j-th sampling;

[0074] C3. In order to estimate the stability of the network's prediction of samples, the uncertainty score obtained by sampling is statistically analyzed by the standard deviation σ(U i ), variance σ 2 (U i ) and mean μ(U i ), where variance σ 2 (U i ) is a measure of network uncertainty, and in is the mean value μ(U i ),and

[0075] S3. According to the uncertainty information measured in step S2, the basic classification network is optimized to obtain the final classification model; specifically, the steps include:

[0076] a. Use evidence information to parameterize the Dirichlet distribution to reduce uncertainty; specifically, the following steps are included:

[0077] a1. Obtain the evidence information corresponding to each category obtained in step S2 and α i is the total evidence that the network classifies the i-th sample as a positive sample, β i is the total evidence that the network classifies the i-th sample as a negative sample; then the obtained evidence information is used to parameterize the Dirichlet distribution, and the parameterized distribution is expressed as in is the probability that the network predicts the sample as a positive sample, α i ' is the total evidence corresponding to the true category of the sample being a positive sample, β i ' is the total evidence corresponding to the true category of the sample being a negative sample;

[0078] a2. Because the classification evidence information of the sample also contains negative evidence information corresponding to uncertainty, and the final optimization goal is to reduce the classification evidence information corresponding to uncertainty as much as possible, the KL divergence is used to measure the difference between the Dirichlet distribution parameterized by evidence and the Dirichlet distribution under 0 evidence; the following formula is used to express the difference information:

[0079]

[0080] Where KL(A|B) is the KL divergence value between distribution A and distribution B; is the Dirichlet distribution under 0 evidence;

[0081] a3. Use the following formula to construct the constraints for network training and overcome the uncertainty information in the network by minimizing the constraints:

[0082]

[0083] Where L i EDL (θ) is the constraint term for network training; θ represents all network parameters; y i is the true label of the i-th sample; p i is the network prediction probability of the i-th sample; || ||2 is the sum of squares; D(x|α i ,β i ) is the i-th sample with respect to the continuous variable x, and the parameter is (α i ,β i ) of the Dirichlet distribution; λ is a dynamically changing coefficient, and t is the current training epoch number; when y i =0<α i ′,β i ′>=<1,β i >, when y i =1<α i ′,β i ′>=<α i ,1>;

[0084] b. Minimize the variance, standard deviation, and mean of the network predictions obtained during the sampling process to reduce the uncertainty of the network; specifically, the following steps are included:

[0085] b1. Set the standard deviation σ(Ui ), variance σ 2 (U i ) and mean μ(U i ) as the optimization target;

[0086] In the specific implementation, considering the prediction stability of the optimization model, the standard deviation σ(U i ) and variance σ 2 (U i ) is used as one of the optimization goals, because the variance and standard deviation indicate that the model has some uncertainty information in the prediction process, which causes large fluctuations in the prediction. Sampling can reflect the uncertainty information in the model. By minimizing the variance and standard deviation, the model can learn and overcome this uncertainty information.

[0087] At the same time, the mean μ(U i ) as one of the optimization goals; this mean value represents the prediction mean after multiple predictions by the model, which can well reflect the uncertainty information of the model's true prediction; at the same time, because the ultimate goal is to improve the reliability of model prediction by reducing the uncertainty of the model, minimizing the mean value can make the model prediction more reliable;

[0088] b2. Combined standard deviation σ(U i ), variance σ 2 (U i ) and mean μ(U i ), and use the following formula as the optimization target formula:

[0089]

[0090] Wherein λ1, λ2 and λ3 are set parameters, preferably 1.5, 0.15 and 0.15; To optimize the target value, the uncertainty in the sampling process is reduced by minimizing the optimized target value;

[0091] c. Construct the final constraint term to optimize the basic classification network; specifically, the following steps are included:

[0092] The final constraint term Loss is constructed using the following formula:

[0093] Loss = L i CE +L i EDL (θ)+L i MS

[0094] Where L i CE is the cross entropy loss corresponding to the i-th sample; L iEDL (θ) is the constraint term for the constructed network training; L i MS The optimization target value for the construction; by minimizing the final constraint term Loss, the reliability of the model confidence prediction is improved;

[0095] S4. Using the classification model obtained in step S3, classify the target image and output the confidence of the corresponding classification result; specifically comprising the following steps:

[0096] Obtain the support value output by the network's fully connected layer, and calculate the subjective evidence for each category as The total evidence for the sample classification is The prediction confidence probability corresponding to each class is obtained as and The final prediction confidence of the sample is

[0097] like The corresponding predicted classification category is the positive sample category; otherwise, the corresponding predicted classification category is the negative sample category.

[0098] like Figure 2 The method flow diagram of the fundus image classification method of the present invention is shown as follows: the fundus image classification method including the image classification method based on uncertainty measurement provided by the present invention specifically includes the following steps:

[0099] (1) Obtain an existing fundus image dataset;

[0100] (2) using the image classification method based on uncertainty measurement to learn the fundus image dataset obtained in step (1) to obtain a final fundus image classification model;

[0101] (3) Using the fundus image classification model obtained in step (2), classify the actual fundus image and output the confidence of the final classification result.

[0102] In addition, the fundus image classification method of the present invention can also be applied to existing fundus image detection instruments to form an imaging method, which specifically includes the following steps:

[0103] 1) Acquire a target fundus image using an existing fundus image detection instrument;

[0104] 2) classifying the target fundus image using the fundus image classification method to obtain a corresponding classification result;

[0105] 3) the classification result obtained in step 2) is directly annotated and imaged on the target fundus image to obtain a fundus image with the classification result;

[0106] 4) The fundus image detection instrument outputs the fundus image with the classification result to complete the corresponding imaging.

[0107] In specific applications, an existing fundus image detection instrument is used to obtain the corresponding target fundus image, and then the target fundus image is directly classified using the above-mentioned fundus image classification method; after the classification is completed, the classification result is directly marked and re-imaged on the target fundus image; finally, the fundus image output by the fundus image detection instrument is the fundus image with the classification result.

[0108] In addition, the image classification method of the present invention can also be applied as follows:

[0109] The present invention can be applied to the classification of character pictures.

[0110] A method for classifying pictures based on the gender of a person, comprising the following steps:

[0111] Ⅰ. Obtain existing pictures with gender of people, classify and label them, and form an image dataset; the pictures are labeled as male pictures or female pictures;

[0112] Ⅱ. Using the image classification method based on uncertainty measurement, the image data set obtained in step Ⅰ is studied to obtain the final image classification model based on the gender of the person;

[0113] III. Use the image classification model based on character gender obtained in step II to classify the actual character images to be classified, and output the confidence of the final classification result.

[0114] The present invention can also be applied to detecting the appearance of produced mechanical parts.

[0115] A method for detecting the appearance of mechanical parts, comprising the following steps:

[0116] ⅰ. Obtaining the appearance pictures of existing mechanical parts (such as the front view pictures of manufactured mechanical parts), and classifying and marking them to form an image data set; marking the pictures as normal parts or abnormal parts;

[0117] ⅱ using the image classification method based on uncertainty measurement, the image data set obtained in step ⅰ is learned to obtain the final mechanical parts appearance detection model;

[0118] ⅲ using the mechanical parts appearance detection model obtained in step ⅱ, the appearance of the mechanical parts to be detected are classified, and the final classification results and the confidence of the classification results are output;

[0119] iv. Based on the classification result of step iii, complete the appearance inspection of the mechanical parts to be inspected.

Claims

1. An image classification method based on uncertainty measurement, comprising the following steps: S1. Obtain existing data sets and classify and mark them; S2. Construct a basic classification network, input the data set obtained in step S1 into the basic classification network, and measure the uncertainty information according to the output of the network; specifically, the following steps are included: A. Use VGG16 as the basic classification network; specifically include the following steps: Use VGG16 as the basic classification network; The cross entropy loss is used as the basic loss, and the following formula is used as the loss formula: In the formula Cross entropy loss value; y i is the true label of the i-th sample; p i The network predicts the classification probability for the i-th sample; log() represents the logarithm with base 2; B. Input the data set obtained in step S1 into the basic classification network, and measure the uncertainty information according to the output of the basic classification network; specifically, convert the output of the fully connected layer of the basic classification network into evidence information, and measure the uncertainty information based on the evidence information; specifically, the following steps are included: B1. Use the output value of the network for each category as the support for the prediction: For the binary classification task, the output of the network for each category is The class activation value for the network output classified as a positive sample, The class activation value for the network output classified as a negative sample; In order to ensure the non-negativity of the evidence corresponding to each category, the support opinion value corresponding to each category is passed through the ReLU activation function to generate the classification subjective evidence corresponding to each category. The evidence value for the network output classification as a positive sample, Output the evidence value for the network to classify as a negative sample; B2. For the binary classification task, the confidence score corresponding to the positive sample is The confidence score corresponding to the negative sample is The uncertainty fraction of the sample is u i , then the confidence score and uncertainty score of the sample satisfy the relationship The evidence corresponding to the uncertainty of each category of the sample is assigned a fixed value of 1, combined with the subjective evidence corresponding to each category Calculate the total evidence S corresponding to the sample i for B3. Calculate the predicted uncertainty score u of the sample i for At the same time, the confidence score corresponding to the positive sample is calculated for Confidence scores corresponding to negative samples for C. Perform several samplings on step B to measure deeper uncertainty information; specifically, the following steps are included: C1. Set dropout in the last two fully connected layers of the network, and set the dropout probability to 0.3; C2. Perform several random samplings on the fully connected layer parameters to obtain several sets of network output values Thus, the corresponding uncertainty score u is obtained for each set of sampled output values i ; Finally, through sampling, a set of uncertainty scores are collected for each sample Where N is the number of sampling times, is the uncertainty value of the network prediction for the i-th sample at the j-th sampling; C3. The uncertainty fraction obtained by sampling, statistical standard deviation σ(U i ), variance σ 2 (U i ) and mean μ(U i );in, Variance σ 2 (U i ) is a measure of network uncertainty, and in is the mean value μ(U i ),and S3. According to the uncertainty information measured in step S2, the basic classification network is optimized to obtain the final classification model; S4. Use the classification model obtained in step S3 to perform image classification of the target image and output the confidence of the corresponding classification result.

2. The image classification method based on uncertainty measurement according to claim 1 is characterized in that The step S3 specifically includes the following steps: a. Use evidence information to parameterize the Dirichlet distribution, thereby reducing uncertainty; b. Minimize the variance, standard deviation, and mean of the network predictions obtained during the sampling process to reduce the uncertainty of the network; c. Construct the final constraints to optimize the basic classification network.

3. The image classification method based on uncertainty measurement according to claim 2 is characterized in that The step a specifically comprises the following steps: a1. Obtain the evidence information corresponding to each category obtained in step S2 and α i is the total evidence that the network classifies the i-th sample as a positive sample, β i is the total evidence that the network classifies the i-th sample as a negative sample; then the obtained evidence information is used to parameterize the Dirichlet distribution, and the parameterized distribution is expressed as in is the probability that the network predicts the sample as a positive sample, α i ' is the total evidence corresponding to the true category of the sample being a positive sample, β i ' is the total evidence corresponding to the true category of the sample being a negative sample; a2. KL divergence is used to measure the difference between the Dirichlet distribution parameterized by evidence and the Dirichlet distribution under 0 evidence. The difference information is expressed by the following formula: Where KL(A|B) is the KL divergence value between distribution A and distribution B; is the Dirichlet distribution under 0 evidence; a3. Use the following formula to construct the constraints for network training and overcome the uncertainty information in the network by minimizing the constraints: In the formula is the constraint item of network training; θ represents all parameters of the network; y i is the true label of the i-th sample; p i is the network prediction probability of the i-th sample; || ||2 is the sum of squares; D(x|α i ,β i ) is the i-th sample with respect to the continuous variable x, and the parameter is (α i ,β i ) of the Dirichlet distribution; λ is a dynamically changing coefficient, and t is the current training epoch number; when y i =0<α′ i ,β′ i >=<1,β i >, when y i =1<α′ i ,β′ i >=<α i ,1>.

4. The image classification method based on uncertainty measurement according to claim 3 is characterized in that The step b specifically comprises the following steps: b1. Set the standard deviation σ(U i ), variance σ 2 (U i ) and mean μ(U i ) as the optimization target; b2. Use the following formula as the optimization target formula: Where λ1, λ2 and λ3 are the set parameters; To optimize the target value, the uncertainty in the sampling process is reduced by minimizing the optimized target value; The step c specifically comprises the following steps: The final constraint term Loss is constructed using the following formula: Loss=L i CE +L i EDL (θ)+L i MS Where L i CE is the cross entropy loss corresponding to the i-th sample; L i EDL (θ) is the constraint term for the constructed network training; L i MS The optimization target value for the construction is minimized, thereby improving the reliability of the network confidence prediction.

5. The image classification method based on uncertainty measurement according to claim 4 is characterized in that The S4 specifically comprises the following steps: Obtain the support value of the network's fully connected layer output, and calculate the subjective evidence for each category as The total evidence for the sample classification is The prediction confidence probability corresponding to each class is obtained as and The final prediction confidence of the sample is like The corresponding predicted classification category is the positive sample category; otherwise, the corresponding predicted classification category is the negative sample category.

6. A fundus image classification method comprising the image classification method based on uncertainty measurement as claimed in any one of claims 1 to 5, characterized in that The specific steps include: (1) Obtain an existing fundus image dataset; (2) using the image classification method based on uncertainty measurement to learn the fundus image dataset obtained in step (1) to obtain a final fundus image classification model; (3) Using the fundus image classification model obtained in step (2), classify the actual fundus image and output the confidence of the final classification result.

Images