Image label calibration method based on active learning and semi-supervised learning algorithms

Abstract

The invention discloses an image label calibration method based on an active learning and semi-supervised learning algorithm, and relates to the technical field of semi-supervised image label calibration, which comprises the steps of data preprocessing, data division, data enhancement, label-free sample label prediction based on a Transductive-VAT model, and calculation of uncertainty of a label-free sample. According to the method, samples with large uncertainty are calibrated on the basis of active learning, all original processing and current processing schemes are recorded, and the effect value of the processing scheme is obtained.The method has the advantages that high accuracy is shown under all given training sample numbers through the Active-T-VAT model, the accuracy is synchronously improved along with increasing of the training sample numbers, and the effect value of the processing scheme is obtained. The accuracy of the Activity-T-VAT model is stably improved, meanwhile, the accuracy fluctuation of the Activity-T-VAT model is more stable, related unlabeled data of small sample data can be effectively learned, the accuracy and performance of simulation recognition of small samples in an underwater scene are improved, label calibration is carried out on the samples, and the difficulty degree and cost of obtaining labeled samples are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of semi-supervised image label calibration, and in particular to an image label calibration method based on active learning and semi-supervised learning algorithms. Background Art

[0002] Unlabeled samples usually have distribution differences with small sample data due to factors such as time. In semi-supervised learning, it is impossible to provide effective data expansion for small samples. Effective unlabeled sample selection and label calibration have a certain improvement effect on the performance of small sample models. Such tasks can be completed based on active learning. On this basis, combined with semi-supervised learning algorithms, the label calibration of small sample underwater images can be effectively solved.

[0003] In the underwater image classification problem, affected by the complex marine environment, the collected images are easily polluted by noise, and the contrast between the target and the background is low, which increases the difficulty of classification. Although it is relatively easy to obtain underwater target samples under certain conditions, the samples obtained are mostly a large amount of unlabeled data, while it is more difficult and costly to obtain labeled samples, resulting in a small sample state. Therefore, the present invention proposes an image label calibration method based on active learning and semi-supervised learning algorithms. Summary of the invention

[0004] The purpose of the present invention is to provide an image label calibration method based on active learning and semi-supervised learning algorithms.

[0005] In order to solve the problems raised in the above background technology, the present invention provides the following technical solution: an image label calibration method based on active learning and semi-supervised learning algorithms, comprising the following steps:

[0006] Step 1: Data preprocessing: collecting underwater image data and preprocessing it, including image Gaussian mixture denoising and normalization processing;

[0007] Step 2: Data division: divide the data into training set L and test set U, and the total samples N = L + U;

[0008] Step 3: Data enhancement: Use projection transformation, affine transformation and random cropping to increase the number of training set samples.

[0009] As a further solution of the present invention: in the step one, when collecting underwater image data, it can be obtained by combining professional underwater cameras and sonar imaging technology, and better quality underwater image data can be obtained by using professional underwater cameras as the main and sonar imaging technology as the auxiliary.

[0010] As a further solution of the present invention: in the step three, when randomly cropping the underwater image data, a single underwater image data cropping method is randomly formulated, and then the image is cropped, thereby increasing the diversity of the underwater image data.

[0011] As a further solution of the present invention: based on active learning image label calibration, the following steps are included:

[0012] S1, output unlabeled test set label probability based on Transductive-VAT model;

[0013] S2. Calculate the uncertainty entropy of unlabeled samples and calculate the average of all entropies;

[0014] S3, select samples with small uncertainty according to the average value of entropy for active learning to complete label calibration, and add the labeled samples to the training set L`;

[0015] S4, repeat steps S1-S3 until the model converges, and use Transductive-VAT to label the remaining test set U';

[0016] S5. Record all original and current processing solutions, and analyze the operating costs of the two processing solutions respectively;

[0017] S6. Calculate the effects of the current treatment plan and the original treatment plan, check the difference between the two treatment effects, and obtain the effectiveness value of the treatment plan.

[0018] As a further solution of the present invention: the specific training process of the Transductive-VAT is as follows:

[0019] ① For a given data set, an initial model is trained using the cross entropy loss function;

[0020] ② For each sample, generate a virtual sample through random perturbation;

[0021] ③ Calculate a difference metric on the virtual sample, where the difference metric reflects the sensitivity of the model to the virtual sample;

[0022] ④ According to the difference metric, calculate an adversarial perturbation, which can change the prediction result of the corresponding label sample of the model and has a certain smoothness property;

[0023] ⑤ Recalculate the difference metric on the perturbed samples and back-propagate to update the model parameters;

[0024] ⑥Calculate the uncertainty of unlabeled samples;

[0025] ⑦ Repeat the above process until the model converges.

[0026] As a further solution of the present invention: the uncertainty entropy of the unlabeled sample is calculated as follows:

[0027] Use the softmax function to convert the prediction results predictions into probability distribution probabilities. The softmax function is defined as follows, where x is the vector of prediction results;

[0028]

[0029] Use the log function to calculate the logarithmic probability log_probabilities. The log function is defined as follows, where x is the vector of probability distribution;

[0030] log(x)i=log(x;)

[0031] The uncertainty entropy of the sample is calculated using the definition of entropy, which is defined as follows, where x is a vector of probability distribution;

[0032]

[0033] The uncertainty can be expressed as:

[0034]

[0035] The above processing scheme can combine active learning, sample uncertainty and VAT model to improve the performance and efficiency of semi-supervised label calibration of small sample images.

[0036] As a further solution of the present invention: in S5, when recording the original treatment plan and the current treatment plan, a temporary database can be established in advance, and the information of the original treatment plan and the current treatment plan can be recorded in the temporary database. Then a calculation model is established, and the calculation model is used to calculate the operating costs of the two plans. At this time, a cost cycle function is established, and the user can type the cycle required by the user in the cost cycle function. At this time, the cost cycle function can calculate all the costs required within the cycle.

[0037] As a further solution of the present invention: in said S6, after analyzing the effect information of the current treatment plan and the original treatment plan, the effects of the current treatment plan and the original treatment plan can be evaluated, so as to obtain the quality evaluation values ​​of the current treatment plan and the original treatment plan, and then calculate the numerical difference between the two sets of quality evaluation values, and then obtain the operating costs of the original treatment plan and the current treatment plan, and calculate the effectiveness value of the original treatment plan according to the operating cost of the original treatment plan and the quality evaluation value of the original treatment plan, and then calculate the effectiveness value of the current treatment plan according to the operating cost of the current treatment plan and the quality evaluation value of the current treatment plan, and let the operating cost be YC , let the quality assessment value be Y P , let the effect value be Y C ;

[0038]

[0039] According to the above formula, the effectiveness values ​​of the original solution and the current solution can be calculated, and the user can select the best solution based on the effectiveness values.

[0040] By adopting the above technical solution, compared with the prior art, the beneficial effects of the present invention are:

[0041] 1. The present invention shows high accuracy under all given numbers of training samples through the Active-T-VAT model, and the accuracy increases synchronously with the increase of the number of training samples, and its accuracy rate steadily increases. At the same time, the accuracy rate fluctuation of the Active-T-VAT model is more stable. The Active-T-VAT model can also reduce the sensitivity to a small number of training samples, and can effectively learn the unlabeled data related to small sample data, thereby improving the accuracy and performance of simulated recognition of small samples in underwater scenes, and can improve computing efficiency, label the samples, and reduce the difficulty and cost of obtaining labeled samples;

[0042] 2. Through the setting of the Active-T-VAT model, the accuracy of the Active-T-VAT model can remain relatively stable under different sample numbers, thereby showing good model stability, and the difference between the accuracy of the Active-T-VAT model and other models is more than 5%, indicating that the Active-T-VAT model is more accurate than other models, thereby further improving the accuracy of simulation recognition of small samples in underwater scenes;

[0043] 3. The present invention can provide users with a scientific and objective way to select the best treatment plan through the calculation of effectiveness value, avoiding the deviation and uncertainty of subjective judgment. At the same time, it can also more comprehensively measure the advantages and disadvantages of the plans, which helps to ensure the quality and effect of the plans while controlling costs. It can also enable users to understand the performance of each plan more clearly. The cost cycle function can meet the personalized needs of different users for the cost calculation cycle in different scenarios, which is helpful for comprehensive cost planning and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a schematic diagram of the image label calibration process based on active learning and semi-supervised learning algorithms in an embodiment of the present invention;

[0045] Figure 2Schematic diagram of data preprocessing, division and enhancement steps in an embodiment of the present invention;

[0046] Figure 3 A schematic diagram of the steps of an image label calibration method based on active learning in an embodiment of the present invention;

[0047] Figure 4 Schematic diagram of specific training steps of Transductive-VAT in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0049] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0050] Please see attached Figure 1 -Attached Figure 4 The image label calibration method based on active learning and semi-supervised learning algorithms of the present invention includes the following steps:

[0051] Step 1: Data preprocessing: collecting underwater image data and preprocessing it, including image Gaussian mixture denoising and normalization processing;

[0052] Step 2: Data division: divide the data into training set L and test set U, and the total samples N = L + U;

[0053] Step 3: Data enhancement: Use projection transformation, affine transformation and random cropping to increase the number of training set samples.

[0054] In one embodiment of the present invention: In step one, when collecting underwater image data, it can be obtained by combining a professional underwater camera and sonar imaging technology, and better quality underwater image data can be obtained by using a professional underwater camera as the main and sonar imaging technology as the auxiliary.

[0055] In one embodiment of the present invention: in step three, when randomly cropping underwater image data, a single underwater image data cropping method is randomly formulated, and then the image is cropped, thereby increasing the diversity of the underwater image data.

[0056] In one embodiment of the present invention: based on active learning image label calibration, the following steps are included:

[0057] S1, output unlabeled test set label probability based on Transductive-VAT model;

[0058] S2. Calculate the uncertainty entropy of unlabeled samples and calculate the average of all entropies;

[0059] S3, select samples with small uncertainty according to the average value of entropy for active learning to complete label calibration, and add the labeled samples to the training set L`;

[0060] S4, repeat steps S1-S3 until the model converges, and use Transductive-VAT to label the remaining test set U';

[0061] S5. Record all original and current processing solutions, and analyze the operating costs of the two processing solutions respectively;

[0062] S6. Calculate the effects of the current treatment plan and the original treatment plan, check the difference between the two treatment effects, and obtain the effectiveness value of the treatment plan.

[0063] In one embodiment of the present invention: the specific training process of Transductive-VAT is as follows:

[0064] ① For a given data set, an initial model is trained using the cross entropy loss function;

[0065] ② For each sample, generate a virtual sample through random perturbation;

[0066] ③ Calculate a difference metric on the virtual sample, where the difference metric reflects the sensitivity of the model to the virtual sample;

[0067] ④ According to the difference metric, calculate an adversarial perturbation, which can change the prediction result of the corresponding label sample of the model and has a certain smoothness property;

[0068] ⑤ Recalculate the difference metric on the perturbed samples and back-propagate to update the model parameters;

[0069] ⑥Calculate the uncertainty of unlabeled samples;

[0070] ⑦ Repeat the above process until the model converges.

[0071] In one embodiment of the present invention, the uncertainty entropy of unlabeled samples is calculated as follows:

[0072] Use the softmax function to convert the prediction results predictions into probability distribution probabilities. The softmax function is defined as follows, where x is the vector of prediction results;

[0073]

[0074] Use the log function to calculate the logarithmic probability log_probabilities. The log function is defined as follows, where x is the vector of probability distribution;

[0075] log(x)i=log(x;)

[0076] The uncertainty entropy of the sample is calculated using the definition of entropy, which is defined as follows, where x is a vector of probability distribution;

[0077]

[0078] The uncertainty can be expressed as:

[0079]

[0080] The above processing scheme can combine active learning, sample uncertainty and VAT model to improve the performance and efficiency of semi-supervised label calibration of small sample images.

[0081] In one embodiment of the present invention: In S5, when recording the original treatment plan and the current treatment plan, a temporary database can be established in advance, and the information of the original treatment plan and the current treatment plan can be recorded in the temporary database. Then a calculation model is established, and the calculation model is used to calculate the operating costs of the two plans. At this time, a cost cycle function is established, and the user can type the cycle required by the user in the cost cycle function. At this time, the cost cycle function can calculate all the costs required within the cycle.

[0082] In one embodiment of the present invention: In S6, after analyzing the effect information of the current treatment plan and the original treatment plan, the effects of the current treatment plan and the original treatment plan can be evaluated, thereby obtaining the quality evaluation values ​​of the current treatment plan and the original treatment plan, and then calculating the numerical difference between the two sets of quality evaluation values, and then obtaining the operating costs of the original treatment plan and the current treatment plan, and calculating the effectiveness value of the original treatment plan according to the operating cost of the original treatment plan and the quality evaluation value of the original treatment plan, and then calculating the effectiveness value of the current treatment plan according to the operating cost of the current treatment plan and the quality evaluation value of the current treatment plan, and assuming that the operating cost is Y C , let the quality assessment value be Y P , let the effect value be Y C ;

[0083]

[0084] According to the above formula, the effectiveness values ​​of the original solution and the current solution can be calculated, and the user can select the best solution based on the effectiveness values.

[0085] Example 1, please refer to the attached Figure 1 -Attached Figure 4 ,In order to meet the needs of small samples, three groups of experiments are set up. The specific experimental data are shown in Table 1:

[0086]

[0087] Table 1 Experimental data and then get the model accuracy results. The specific model accuracy results are shown in Table 2:

[0088]

[0089]

[0090] Table 2 Model accuracy results (%)

[0091] Example 2, please refer to the attached Figure 1 -Attached Figure 4 When obtaining underwater image data with better quality by using a professional underwater camera as the main method and sonar imaging technology as the auxiliary method, the underwater image data taken by the professional underwater camera can be used as a sample, and the sonar imaging technology can be used to obtain the details of the underwater image information. Then, the details of the underwater image data obtained by the sonar imaging technology are supplemented to the underwater image data taken by the professional camera, thereby obtaining underwater image data with better quality.

[0092] Example 3, please refer to the attached Figure 1 -Attached Figure 4 When calculating the operating costs of the original treatment plan and the current treatment plan, you can use spreadsheet software, use Excel or similar spreadsheet tools to record the detailed information of the original treatment plan and the current treatment plan, create a separate worksheet for each plan, including the details of each cost, and then use formulas and functions to calculate the total cost of each cycle. Users can automatically obtain the corresponding cost results by entering specific cycle values.

[0093] Specifically, the Active-T-VAT model shows high accuracy under all given numbers of training samples, and the accuracy increases synchronously with the increase in the number of training samples. Its accuracy rate steadily improves, and at the same time, the accuracy fluctuation of the Active-T-VAT model is more stable. The Active-T-VAT model can also reduce the sensitivity to a small number of training samples, and can effectively learn the related unlabeled data of small sample data, thereby improving the accuracy and performance of simulated recognition of small samples in underwater scenes, improving computing efficiency, calibrating labels for samples, and reducing the difficulty and cost of obtaining labeled samples.

[0094] Specifically, through the setting of the Active-T-VAT model, the accuracy of the Active-T-VAT model can remain relatively stable under different numbers of samples, thus showing good model stability, and the differences between the accuracy of the Active-T-VAT model and other models are all above 5%, indicating that the Active-T-VAT model is more accurate than other models, thereby further improving the accuracy of simulated recognition of small samples in underwater scenes.

[0095] Specifically, the calculation of effectiveness value can provide users with a scientific and objective way to select the best treatment plan, avoiding the bias and uncertainty of subjective judgment. At the same time, it can more comprehensively measure the pros and cons of the plans, helping to ensure the quality and effectiveness of the plans while controlling costs. It can also enable users to understand the performance of each plan more clearly. The cost cycle function can meet the personalized needs of different users for the cost calculation cycle in different scenarios, which is helpful for comprehensive cost planning and control.

[0096] Working principle:

[0097] Step 1: During the model training process, the data will be preprocessed in real time before being input into the model. For the sonar image data, operations including conversion to Tensor, normalization and denoising are performed. Specifically, the image will be randomly cropped to a size of 224*224, each channel will be standardized, and the image will be denoised using a Gaussian mixture model. According to the requirements of small samples, the data will be divided into a training set L and a test set U, and all samples N = L+U. The training set L will be enhanced to make up for the problem of insufficient data volume. Specifically, the number of training set samples is increased by using methods such as projection transformation, affine transformation and random cropping.

[0098] Step 2: Input the labeled training set and the unlabeled test set into the Transductive-VAT model at the same time, output the unlabeled test set label probability predictions, calculate the uncertainty entropy of the unlabeled samples, and calculate the average value of all uncertainties avg_entropy.

[0099] Step 3: Select samples with small uncertainty according to the average value of entropy for active learning to complete label calibration, and add the labeled samples to the training set L`. Specifically, compare the uncertainty entropy of each sample with the average value avg_entropy. If the entropy of the sample is less than avg_entropy, the sample is actively learned for label calibration. If it is greater than avg_entropy, it remains in the test set. Through this method of sample selection, it can be ensured that active learning has learned samples whose labels are difficult for the Transductive-VAT model to determine. Repeat steps 4-6 until the model converges, and use Transductive-VAT to calibrate the labels of the remaining test set U`;

[0100] Step 4: Record all the original and current treatment plans, analyze the operating costs of the two treatment plans respectively, calculate the effects of the current treatment plan and the original treatment plan, evaluate the effects of the current treatment plan and the original treatment plan, check the gap between the two treatment effects, and obtain the effectiveness value of the treatment plan. At this point, the entire workflow is completed.

[0101] At the same time, the present application uses specific words to describe the embodiments of the present application. For example, "one embodiment", "an embodiment", and / or "some embodiments" refer to a certain feature, structure or characteristic related to at least one embodiment of the present application. Therefore, it should be emphasized and noted that "one embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or more in different positions in this specification does not necessarily refer to the same embodiment. In addition, certain features, structures or characteristics in one or more embodiments of the present application can be appropriately combined.

[0102] Some aspects of the present application may be performed entirely by hardware, entirely by software (including firmware, resident software, microcode, etc.), or by a combination of hardware and software. The above hardware or software may be referred to as "data blocks", "modules", "engines", "units", "components" or "systems". The processor may be one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DAPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, or combinations thereof. In addition, various aspects of the present application may be expressed as computer products located in one or more computer-readable media, which include computer-readable program codes. For example, computer-readable media may include, but are not limited to, magnetic storage devices (e.g., hard disks, floppy disks, tapes ...), optical disks (e.g., compact disks CDs, digital versatile disks DVDs ...), smart cards, and flash memory devices (e.g., cards, sticks, key drives ...).

[0103] A computer-readable medium may include a propagated data signal containing computer program code, such as in baseband or as part of a carrier wave. The propagated signal may have a variety of manifestations, including electromagnetic, optical, etc., or a suitable combination. A computer-readable medium may be any computer-readable medium other than a computer-readable storage medium, which may be connected to an instruction execution system, device or apparatus to communicate, propagate or transmit a program for use. The program code on the computer-readable medium may be propagated through any suitable medium, including radio, cable, fiber optic cable, radio frequency signal, or similar medium, or a combination of any of the above mediums.

[0104] Similarly, it should be noted that in order to simplify the description disclosed in this application, so as to help understand one or more embodiments of the invention, in the above description of the embodiments of the present application, sometimes multiple features are merged into one embodiment, drawings or descriptions thereof. However, this disclosure method does not mean that the features required by the object of the present application are more than the features mentioned in the claims. In fact, the features of the embodiment are less than all the features of the single embodiment disclosed above. In some embodiments, numbers describing the number of components and attributes are used. It should be understood that such numbers used in the description of the embodiments are modified by the modifiers "approximately", "approximately" or "substantially" in some examples. Unless otherwise specified, "approximately", "approximately" or "substantially" indicate that the number allows a variation of ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, which may change according to the characteristics required by individual embodiments. In some embodiments, the numerical parameters should take into account the specified significant digits and adopt the general method of retaining the digits. Although the numerical domains and parameters used to confirm the breadth of their range in some embodiments of the present application are approximate values, in specific embodiments, such numerical settings are as accurate as possible within the feasible range.

[0105] Although the present invention is disclosed as above in terms of preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection defined by the claims of the present invention.

Claims

1. An image label calibration method based on active learning and semi-supervised learning algorithms, characterized in that: The following steps are involved: Step 1: Data preprocessing: collecting underwater image data and preprocessing it, including image Gaussian mixture denoising and normalization processing; Step 2: Data division: divide the data into training set L and test set U, and the total samples N = L + U; Step 3: Data enhancement: Use projection transformation, affine transformation and random cropping to increase the number of training set samples.

2. The image label calibration method based on active learning and semi-supervised learning algorithm according to claim 1 is characterized in that: In the step 1, when collecting underwater image data, it can be obtained by combining professional underwater cameras and sonar imaging technology, and better quality underwater image data can be obtained by using professional underwater cameras as the main technology and sonar imaging technology as the auxiliary technology.

3. The image label calibration method based on active learning and semi-supervised learning algorithm according to claim 2 is characterized in that: In the step three, when randomly cropping the underwater image data, a single underwater image data cropping method is randomly formulated, and then the image is cropped, thereby increasing the diversity of the underwater image data.

4. The image label calibration method based on active learning and semi-supervised learning algorithm according to claim 3 is characterized in that: Active learning-based image label calibration includes the following steps: S1, output unlabeled test set label probability based on Transductive-VAT model; S2. Calculate the uncertainty entropy of unlabeled samples and calculate the average of all entropies; S3, select samples with small uncertainty according to the average value of entropy for active learning to complete label calibration, and add the labeled samples to the training set L`; S4, repeat steps S1-S3 until the model converges, and use Transductive-VAT to label the remaining test set U'; S5. Record all original and current processing solutions, and analyze the operating costs of the two processing solutions respectively; S6. Calculate the effects of the current treatment plan and the original treatment plan, check the difference between the two treatment effects, and obtain the effectiveness value of the treatment plan.

5. The image label calibration method based on active learning and semi-supervised learning algorithm according to claim 4 is characterized in that: The specific training process of Transductive-VAT is as follows: ① For a given data set, an initial model is trained using the cross entropy loss function; ② For each sample, generate a virtual sample through random perturbation; ③ Calculate a difference metric on the virtual sample, where the difference metric reflects the sensitivity of the model to the virtual sample; ④ According to the difference metric, calculate an adversarial perturbation, which can change the prediction result of the corresponding label sample of the model and has a certain smoothness property; ⑤ Recalculate the difference metric on the perturbed samples and back-propagate to update the model parameters; ⑥Calculate the uncertainty of unlabeled samples; ⑦ Repeat the above process until the model converges.

6. The image label calibration method based on active learning and semi-supervised learning algorithm according to claim 4 is characterized in that: The uncertainty entropy of the unlabeled sample is calculated as follows: Use the softmax function to convert the prediction results predictions into probability distribution probabilities. The softmax function is defined as follows, where x is the vector of prediction results; Use the log function to calculate the logarithmic probability log_probabilities. The log function is defined as follows, where x is the vector of probability distribution; log(x)i=log(x;) The uncertainty entropy of the sample is calculated using the definition of entropy, which is defined as follows, where x is a vector of probability distribution; The uncertainty can be expressed as: The above processing scheme can combine active learning, sample uncertainty and VAT model to improve the performance and efficiency of semi-supervised label calibration of small sample images.

7. The image label calibration method based on active learning and semi-supervised learning algorithm according to claim 4 is characterized in that: In S5, when recording the original treatment plan and the current treatment plan, a temporary database can be established in advance, and the information of the original treatment plan and the current treatment plan can be recorded in the temporary database. Then a calculation model is established, and the calculation model is used to calculate the operating costs of the two plans. At this time, a cost cycle function is established, and the user can type the cycle required by the user in the cost cycle function. At this time, the cost cycle function can calculate all the costs required within the cycle.

8. The image label calibration method based on active learning and semi-supervised learning algorithm according to claim 5, characterized in that: In S6, after analyzing the effect information of the current treatment plan and the original treatment plan, the effects of the current treatment plan and the original treatment plan can be evaluated, thereby obtaining the quality evaluation values ​​of the current treatment plan and the original treatment plan, and then calculating the numerical difference between the two sets of quality evaluation values, and then obtaining the operating costs of the original treatment plan and the current treatment plan, and calculating the effectiveness value of the original treatment plan according to the operating cost of the original treatment plan and the quality evaluation value of the original treatment plan, and then calculating the effectiveness value of the current treatment plan according to the operating cost of the current treatment plan and the quality evaluation value of the current treatment plan, and assuming that the operating cost is Y C , let the quality assessment value be Y P , let the effect value be Y C ; According to the above formula, the effectiveness values ​​of the original solution and the current solution can be calculated, and the user can select the best solution based on the effectiveness values.

Images