Classification method for training sample expansion according to multi-classifier recognition result

Abstract

This invention discloses a classification method for expanding training samples based on multi-classifier recognition results, comprising: selecting multiple classification methods; selecting initial training samples including each category from the dataset to be classified to form an initial training sample set; classifying and recognizing the dataset data using each classification method and the training sample set to obtain the classification results of the dataset to be classified by each classification method; calculating the classification result acceptance rate for each data based on the classification results of the dataset to be classified; comparing the classification result acceptance rate with its preset threshold to obtain new training samples and expand the training sample set; determining and executing the next iteration classification based on the expanded training sample set; and taking the maximum acceptance classification result of the last iteration as the final data classification result of the dataset to be classified. This invention improves classification accuracy by iteratively verifying the results of multiple classifiers and gradually expanding the training samples.

Description

Technical Field

[0001] This invention relates to the field of multi-classifier fusion algorithm technology, and in particular to a classification method for expanding training samples based on the recognition results of multi-classifiers. Background Technology

[0002] A classifier's role is to determine the category of a new observed sample based on labeled training data. The typical task of a classifier is to learn classification rules using given categories and known training data, and then classify (or predict) unknown data. Classifiers are widely used in many fields such as object recognition, weather forecasting, signal processing, and image classification. Currently, researchers have proposed several mature classifiers. The differences in classification performance among different classifiers are related to several factors, including the statistical distribution characteristics of the data they classify, prior knowledge, the size of the training data samples, and the structure of the classifier itself. Scholars have conducted comparative studies on the performance of various classifiers, and the basic conclusion is that no single classifier is a "panacea" (Giacinto and Roli, 1997). Fernández-Delgado (2014) applied 179 classifiers to classify and predict 121 datasets (UCI database), and found that the performance of classifiers is related to the dataset. Some classifiers perform well on specific datasets but have weak generalization ability, while some classifiers perform poorly on certain datasets but perform well on different datasets. Therefore, different classifiers have a certain degree of complementarity. The method of combining multiple classifiers is also called multiple classifier fusion or multiple classifier systems. The purpose is to combine various existing classification algorithms using appropriate methods to improve classification accuracy. Currently, multiple classifier fusion algorithms mainly use weighted or unweighted voting of the results of multiple classifiers to form the final classification result. Compared with single classifier methods, it usually has higher classification accuracy, but it takes longer. Summary of the Invention

[0003] The purpose of this invention is to provide a classification method that expands training samples based on the recognition results of multiple classifiers when it is unclear in the early stages which classification method is optimal. This method gradually expands the training samples by iteratively verifying the results of multiple classifiers and eliminates classifiers with poor recognition performance by comparing the results of multiple classifiers, thereby improving classification speed and accuracy.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] Classification methods that augment training samples based on multi-classifier recognition results include the following:

[0006] Step S10. Preparation procedure: Select the classification method used for multiple classifications; select initial training samples including each category in the dataset to be classified to form the initial training sample set;

[0007] Step S20. Classification and recognition procedure: Classify and recognize the dataset to be classified using various classification methods and training sample sets to obtain the classification results of the dataset to be classified using each classification method;

[0008] Step S30. Acceptance Rate Calculation Procedure: Based on the classification results of the dataset to be classified, calculate the acceptance rate of the classification result for each data point; whereby the acceptance rate refers to the proportion of the number of classification methods that are recognized as the same classification result to the total number of all participating classification methods;

[0009] Step S40. Sample expansion procedure: Compare the classification result consensus rate with its preset threshold to obtain new training samples and expand the training sample set;

[0010] Step S50. Iteration procedure: Based on the expanded training sample set, determine and execute the next iteration classification; if the proportion of training samples does not meet the standard, execute the aforementioned steps S20-S50, otherwise execute the following step S60; where the proportion of samples refers to the proportion of the number of data in the training sample set to the total dataset to be classified.

[0011] Step S60. Output program: Take the maximum consensus classification result of the last iteration as the final data classification result of the dataset to be classified; where the maximum consensus classification result refers to the classification result corresponding to the largest consensus rate in this iteration.

[0012] The specific content of step S40 is as follows: compare the classification result recognition rate with its preset threshold. When the classification result recognition rate of a certain data is greater than or equal to the recognition rate threshold, the data is used as a new training sample and added to the training sample of the classification result recognition category corresponding to the data to expand the training sample set.

[0013] The specific content of step S50 is as follows: Based on the expanded training sample set, obtain the proportion of training samples; when the proportion of training samples is less than its preset threshold, and the difference between the proportions of training samples in two consecutive iterations is greater than the preset threshold, it is determined that the next iteration classification needs to be performed, and the aforementioned S20-S50 operations are repeated using the expanded training sample set and the classification method used in this iteration; otherwise, the following S60 operation is performed.

[0014] By adopting the above technical solution, the present invention has the following beneficial effects:

[0015] 1. The classification method of the present invention, which expands training samples based on the recognition results of multiple classifiers, gradually expands the training samples and improves the classification accuracy when it is unclear which classification method is optimal in the early stage. Finally, it can obtain a classification result that is closest to the optimal method.

[0016] 2. This invention employs multiple classification algorithms simultaneously. The calculation process is more time-consuming than a single classification method. By comparing the results of multiple classifiers, classifiers with poor recognition performance can be eliminated during the iteration process, thereby improving the classification speed and reducing the interference of misclassification on the final classification result. Attached Figure Description

[0017] Figure 1 This is a flowchart of Example 1 of the classification method of the present invention.

[0018] Figure 2 This is a flowchart of Example 2 of the classification method of the present invention.

[0019] Figure 3 This is a data processing flowchart for Example 2 of the classification method of the present invention.

[0020] Figure 4 This is a canopy diagram of the rice to be classified according to the present invention.

[0021] Figure 5 yes Figure 4 The sample region map for each category.

[0022] Figure 6 yes Figure 4 Verification area diagrams for each category.

[0023] Figure 7 yes Figure 4 The final classification result diagram of the classification method of the present invention.

[0024] Figure 8 yes Figure 4 The classification results of the first iteration (independent operation) of other classification methods are shown in the figure; among them, Figure 8 (a) represents the maximum likelihood. Figure 8 (b) is a neural network. Figure 8 (c) is a support vector machine. Figure 8 (d) is a parallelepiped. Figure 8 (e) represents the minimum distance. Figure 8 (f) represents the Mahalanobis distance. Detailed Implementation

[0025] The specific implementation of the invention will be further described below with reference to the accompanying drawings.

[0026] Example 1

[0027] See Figure 1 The classification method for expanding training samples based on the multi-classifier recognition results in this embodiment includes the following:

[0028] Step S10. Preparation procedure: Select the classification methods used for multiple classifications; select initial training samples including each category in the dataset to be classified to form the initial training sample set; the order of preparation of classification methods and training sample sets can be interchanged or prepared simultaneously as needed.

[0029] Step S20. Classification and recognition procedure: Classify and recognize the dataset to be classified using various classification methods and training sample sets to obtain the classification results of the dataset to be classified using each classification method;

[0030] Step S30. Acceptance Rate Calculation Procedure: Based on the classification results of the dataset to be classified, calculate the acceptance rate of the classification result for each data point; whereby the acceptance rate refers to the proportion of the number of classification methods that are recognized as the same classification result to the total number of all participating classification methods;

[0031] Step S40. Sample expansion procedure: Compare the classification result consensus rate with its preset threshold to obtain new training samples and expand the training sample set;

[0032] Step S50. Iteration procedure: Based on the expanded training sample set, determine and execute the next iteration classification; if the proportion of training samples does not meet the standard, execute the aforementioned steps S20-S50 (return to execute step S20), otherwise execute the following step S60; where the proportion of samples refers to the proportion of the number of data in the training sample set to the total dataset to be classified.

[0033] Step S60. Output program: Take the maximum consensus classification result of the last iteration as the final data classification result of the dataset to be classified; where the maximum consensus classification result refers to the classification result corresponding to the largest consensus rate in this iteration.

[0034] The specific content of step S40 is as follows: compare the classification result recognition rate with its preset threshold. When the classification result recognition rate of a certain data is greater than or equal to the recognition rate threshold, the data is used as a new training sample and added to the training sample of the classification result recognition category corresponding to the data to expand the training sample set.

[0035] The specific content of step S50 is as follows: Based on the expanded training sample set, obtain the proportion of training samples; when the proportion of training samples is less than its preset threshold, and the difference between the proportions of training samples in two consecutive iterations is greater than the preset threshold, it is determined that the next iteration classification needs to be performed, and the aforementioned S20-S50 operations are repeated using the expanded training sample set and the classification method used in this iteration; otherwise, the following S60 operation is performed.

[0036] As described above, the classification method of the present invention, which expands training samples based on the recognition results of multiple classifiers, gradually expands the training samples and improves the classification accuracy when it is unclear which classification method is optimal in the early stage. Ultimately, it can obtain a classification result that is closest to the optimal method.

[0037] See Figure 2 As an option, based on the aforementioned example, in one instance, this classification method selects the number of classifiers during the iteration process. In multi-classifier fusion algorithms, the choice of classifiers affects the final classification result, while the number of classifiers relates to the time efficiency of the classification operation. Specifically:

[0038] The classification method further includes the following steps between step S30 and step S50:

[0039] Precision calculation procedure: Based on the consensus rate of classification results, calculate the precision rate of correct classification for each classification method; where, precision rate refers to the proportion of correct classification results of a certain classification method based on the most consensus classification result as the truth value.

[0040] Screening procedure: Based on the accuracy of each classification method, the accuracy is compared with its preset threshold to screen and update the multiple classification methods used in the next iteration of classification.

[0041] The specific content of the screening procedure is as follows: the accuracy of each classification method is compared with its preset threshold. If the accuracy is less than the threshold, the classification method is removed; otherwise, the classification method is selected. The screening and updating process is used to obtain the multiple classification methods to be used in the next iteration of classification.

[0042] The criterion calculation and screening procedures can be placed between steps S30 and S40; or between steps S40 and S50; or as follows: Figure 2 The steps shown are placed between steps S30 and S40, and between steps S40 and S50, respectively, and are defined as steps S32 and S42 at this time.

[0043] As mentioned above, this invention employs multiple classification algorithms simultaneously. Compared to individual classification methods, the computation process is more time-consuming. By comparing the results of multiple classifiers, classifiers with poor recognition performance can be eliminated during the iteration process, thereby improving the classification speed, achieving better time efficiency in classification operations, and reducing the interference of misclassification on the final classification result.

[0044] See Figure 3 As an option, based on the aforementioned example, a specific data processing flow example of this classification method is provided, including the following steps:

[0045] 1) Select multiple classification methods to be used in this classification, such as Figure 3 The classification methods shown are 1 to n.

[0046] 2) Select samples of each category from the dataset to be classified to form a training sample set.

[0047] 3) Classify and identify the data using different classification methods and training sample sets to obtain the classification results for each method, such as... Figure 3 The classification results R1 to Rn are shown.

[0048] 4) Calculate the classification result consensus rate for each data point in the dataset; where consensus rate refers to the proportion of classification methods that are recognized as the same classification result out of the total number of participating classification methods.

[0049] 5) The classification result with the highest recognition rate is formed based on the recognition rate of the classification results; where the highest recognition rate of the classification result refers to the classification result with the highest recognition rate in this case.

[0050] 6) Set a threshold for the recognition rate of classification results (100% in this case). When the recognition rate of a certain data is greater than or equal to the threshold, the classification result of the data is considered to have high credibility. The data is then used as a new training sample and added to the training samples of the category that the classification result recognizes, thus expanding the training sample set.

[0051] 7) Calculate the expanded training sample proportion; where the sample proportion refers to the proportion of the number of data in the training sample set to the total dataset to be classified.

[0052] 8) Calculate the accuracy of each classification method, such as... Figure 3 The accuracy rates C1 to Cn shown are: based on the maximum consensus classification result as the truth value, the proportion (e.g., percentage) of the correct classification results of a certain classification method compared with the maximum consensus classification result, reflecting the degree of matching between the classification results and the maximum consensus classification results.

[0053] 9) Set a precision threshold (0.7 in this case) (the misclassification rate is greater than 0.3). When the precision is less than the precision threshold, the classification method will be eliminated in the next round of classification iteration. This step can remove classification methods with low accuracy and reduce the calculation process. On the other hand, it can reduce the interference of misclassification on the final classification result.

[0054] 10) Set a training sample proportion threshold (i.e., the classification result agreed upon by all methods; in this case, it is set to 0.9). When the training sample proportion is less than the proportion threshold, use the expanded training sample set from step 6) and the classification method determined in step 9), and repeat steps 3)-9) until the training sample proportion is greater than or equal to the set proportion threshold, or the difference between the training sample proportions of two consecutive iterations is less than or equal to its preset difference threshold. For example, if the training sample proportions are the same in two consecutive iterations, it means that the training samples are no longer increasing, and the classification methods cannot find any more agreed-upon classification results. Alternatively, a difference threshold range for judging proximity can be set. When the difference between the training sample proportions of two consecutive iterations is within the difference threshold range, it can be considered that the training samples are no longer increasing.

[0055] 11) Output the classification result with the highest consensus in the last iteration as the final data classification result.

[0056] Example 2

[0057] Based on the aforementioned Example 1, this Example 2 provides an example of rice canopy cover classification, as detailed below:

[0058] Canopy cover is a useful indicator for assessing crop growth and predicting crop yield. It is typically obtained by taking images of the crop canopy with a digital camera, using classification algorithms to categorize and identify different categories of regions within the image, obtaining the extent of the leaf canopy in the image, and calculating its proportion.

[0059] This case study uses images of the rice canopy taken with a digital camera (see...). Figure 4 The image, used as the data to be classified, is 850*650 pixels in size, containing 552,500 pixels and three layers: R, G, and B. Using the aforementioned classification method, the image was classified into four categories: rice canopy leaves, water, soil, and shadows. The category labels are shown below. Figures 5-7 The bottom left corner and Figure 8 The lower corner. First, select training samples of 4 categories from the image, and delineate the training sample regions (see...). Figure 5 To analyze the accuracy of the classification results, a manual visual interpretation method was used to select areas such as leaves, water bodies, soil, and shadows in the images (see...). Figure 6These serve as validation samples, used to compare and verify the classification results. The number of pixels in each category's training and validation samples is shown in Table 1 below.

[0060] Table 1. Number of training and validation samples for each category

[0061]

[0062] In this case, six commonly used supervised classification methods for image classification were selected: parallelepiped method, minimum distance method, maximum likelihood method, Mahalanobis distance method, neural network, and support vector machine. The classification result recognition rate threshold was set to 100%, the accuracy threshold was set to 0.7, and the training sample proportion threshold was set to 0.9. The iteration termination condition was set to the training sample proportion being greater than 0.9 or the training sample proportion being the same in two consecutive iterations.

[0063] The program uses ENVI / IDL software to write code to execute the calculation process of the above-mentioned patented method. The program presets the default parameters for each classification method in ENVI software. The standard deviation thresholds for the parallelepiped method, minimum distance method, maximum likelihood method, and Mahalanobis distance method are all set to the standard deviation of the training sample set by default; no likelihood threshold is set for the maximum likelihood method; and no maximum error distance is set for the Mahalanobis distance method.

[0064] The parameters of the neural network model are set as follows: the activation function is a logarithmic function, the number of hidden layers is 3, the training contribution threshold is 0.9, the weight adjustment speed is 0.2, the training impulse is 0.9, the root mean square (RMS) threshold for determining training termination is 0.1, and the maximum number of iterations is 100.

[0065] The support vector machine parameters are set as follows: the inner product kernel function is a cubic polynomial, the classification probability domain is 0.5, the penalty coefficient is 75, the bias coefficient is 2, and the GAMMA coefficient is taken as the reciprocal of the number of image bands by default, which is 1 / 3 in this case.

[0066] The program first selects the image to be classified when it runs (see...). Figure 4 Then, manually defined training sample sets for each category are selected (see...). Figure 5 Finally, the program selects the output path for the calculation results. The program will output the classification results of each classification method in each iteration, the maximum consensus classification result, the proportion of training samples, the methods involved in the classification, the accuracy of each classification method, the new training sample set, the number of iterations, and other calculation results in the selected output path.

[0067] The changes in precision and sample proportion of each classification method during the iterative calculation are shown in Table 2 below. In the first iteration, the precision of the parallelepiped method, the minimum distance method, and the Mahalanobis distance method were 0.402621, 0.564147, and 0.64891, respectively, which were lower than the set precision threshold of 0.7. Therefore, these three classification methods were eliminated after the first iteration and no longer participated in subsequent iterations. The remaining three classification methods all had precision rates greater than 0.7 in subsequent iterations and continued to participate in the iterations until the entire calculation was terminated.

[0068] Table 2. Changes in Precision and Sample Proportion of Classification Methods

[0069]

[0070] As shown in Table 2, the highest sample percentage did not occur in the last iteration, but rather in the 7th iteration. It decreased in subsequent iterations, indicating that the classification algorithm's recognition rate based on the training samples had reached saturation. Different classification algorithms could no longer generate samples with consistent agreement; the newly added training samples actually reduced the agreement ratio between different classification methods. Therefore, after two more sample adjustments, the number of commonly agreed-upon samples no longer increased. After the 10th iteration, the sample percentage was 0.537689, the same as in the 9th iteration, reaching the iteration termination condition set by the program, thus ending the iteration operation. The classification result with the highest consensus in the last iteration is the final classification result, as shown below. Figure 7 As shown.

[0071] To analyze the accuracy of each classification algorithm, the classification results of different algorithms were compared with the verification sample areas selected by manual visual interpretation. Simultaneously, to analyze the difference in performance between the patented method and individual classification methods, as well as the changes in classification performance during iterative calculations, accuracy analysis was performed on the first iteration result, the iteration result with the highest sample proportion (7th iteration), and the final iteration result (10th iteration) for each classification method.

[0072] In the first iteration, each classification method used manually selected initial training samples, which can be considered as independent calculations based on the initial training samples. To more intuitively demonstrate the difference in performance between each classification method and the method in this patent, the classification results of the first iteration of each classification method (i.e., the classification results of each classification method's independent calculations based on the initial training samples) are listed separately in... Figure 8 In the middle, the classification effects of each method are compared.

[0073] Two parameters, overall precision and Kappa coefficient, are used to evaluate the classification accuracy of each method. Overall precision refers to the percentage of all classes correctly classified in the validation region. The Kappa coefficient is a statistical indicator for evaluating consistency and can be used to measure classification precision. The Kappa coefficient is calculated between -1 and 1. Generally, a Kappa >= 0.75 indicates good consistency; 0.4 <= Kappa < 0.75 indicates moderate consistency; and Kappa < 0.4 indicates poor consistency. The Kappa coefficient is calculated as follows:

[0074]

[0075] Where Po is the sum of the number of correctly classified samples in each class divided by the total number of samples, which is the overall classification accuracy.

[0076] Assume the number of true samples in each class is a1, a2, ..., aC, and the number of predicted samples in each class is b1, b2, ..., bC, with a total number of samples of n. Then:

[0077]

[0078] The results of this precision analysis are shown in Table 3 below.

[0079] Table 3. Accuracy Analysis of Each Classification Method

[0080]

[0081] As shown in Table 3 of the accuracy analysis results, in the first iteration, the overall accuracy of the parallelepiped method was 38.51%, and the Kappa coefficient was only 0.19. (Classification results...) Figure 8 (d) also shows that the classification results deviate significantly from the actual image conditions, with a large number of areas failing to be classified and identified; the overall accuracy and Kappa coefficient of both the minimum distance method and the Mahalanobis distance method are low, and the classification results are... Figure 8 (e) and Figure 8 (f) shows that a significant portion of the leaf canopy was misidentified as shadow or soil, and the minimum distance method also misidentified some leaf canopies as water bodies. The accuracy analysis results indicate that the classification results of these three methods have poor consistency with the validation samples, resulting in low classification accuracy. Therefore, they are not suitable for classifying and recognizing this image, and these three methods need to be eliminated in subsequent iterations. The change in the accuracy rate of each classification method in Table 2 confirms that these three methods were eliminated after the first iteration. Therefore, the screening results also demonstrate that eliminating low-accuracy classification methods based on the accuracy rate threshold is reasonable and feasible.

[0082] In this case, the classification method of this application achieved an overall classification accuracy of 92.13% and a Kappa coefficient of 0.80, demonstrating good classification performance. Its classification accuracy ranked second among all classification results, second only to the first iteration result of the neural network method, which had an overall accuracy of 92.89% and a Kappa coefficient of 0.81. Compared with the results of other classification methods, the classification accuracy of this application's method is not the highest, but it is closest to the highest accuracy classification result. Therefore, when it is unclear which classification method is optimal in the early stages, using the classification method of this application can achieve classification results that are closer to the optimal method.

[0083] Of course, the method of this application uses multiple classification algorithms to perform calculations at the same time, so the calculation process takes longer than that of a single classification method. If the optimal classification method is known, the classification accuracy and efficiency obtained by directly using the optimal classification method are the best.

[0084] It should be noted that the examples of the above embodiments can preferably be combined with one or more of each other according to actual needs, and the accompanying drawings of multiple examples adopt a set of combined technical features, which will not be described in detail here.

[0085] The above description is a detailed explanation and illustration of the preferred embodiments of the present invention. However, these descriptions are not intended to limit the scope of protection claimed by the present invention. All equivalent changes or modifications made under the technical teachings of the present invention should fall within the patent protection scope covered by the present invention.

Claims

1. A classification method of training sample expansion according to multi-classifier identification results, characterized in that, The classification method is used for image classification, crop canopy picture shooting, classification and identification of the picture by using the classification method, division of different categories of regions in the picture, obtaining the leaf canopy range in the picture and calculating the proportion to obtain the canopy coverage of the crop, including the following contents: Step S10. Preparation program: selecting a plurality of classification methods for classification; Selecting initial training samples including each category in the data set to be classified to form an initial training sample set; Step S20. Classification and identification program: using each classification method and the training sample set to classify and identify the data set to be classified respectively to obtain the classification results of the data set to be classified of each classification method; Step S30. Identification rate calculation program: calculating the classification result identification rate of each data according to the classification result of the data set to be classified; wherein the identification rate refers to the proportion of the number of classification methods identified as the same classification result to the total number of all participating classification methods; Step S40. Sample expansion program: comparing the classification result identification rate with its preset threshold to obtain new training samples and expand the training sample set; Step S50. Iteration program: determining and executing the next iteration classification according to the expanded training sample set; if the training sample proportion does not meet the standard, executing the foregoing steps S20-S50 operations, otherwise executing the following step S60 operation; wherein the sample proportion refers to the proportion of the number of data in the training sample set to the total data set to be classified; Step S60. Output program: taking the maximum identification classification result of the last iteration operation as the final data classification result of the data set to be classified; wherein the maximum identification classification result refers to the classification result corresponding to the maximum classification result identification rate of this time; The specific content of step S50 is as follows: obtaining the training sample proportion according to the expanded training sample set; when the training sample proportion is less than its preset threshold, and the difference between the training sample proportions of the last two iterations is greater than the difference preset threshold, it is determined that the next iteration classification needs to be executed, and the expanded training sample set and the classification method used in this iteration are used to repeat the steps S20-S50 operations, otherwise the step S60 operation is executed. 2.The method of claim 1, wherein the method further comprises: The specific content of step S40 is as follows: comparing the classification result identification rate with its preset threshold, when the classification result identification rate of a certain data is greater than or equal to the identification rate preset threshold, the data is taken as a new training sample and added to the training sample of the classification result identification category corresponding to the data to expand the training sample set. 3.The method of claim 1, wherein the method further comprises: Further comprising the following steps: Accuracy calculation program: calculating the accuracy rate of the correct classification of each classification method based on the classification result identification rate; wherein the accuracy rate refers to the proportion of the correct classification of the classification result of a certain classification method based on the maximum identification classification result as the true value; Screening program: comparing the accuracy rate with its preset threshold based on the accuracy rate of each classification method to screen and update the plurality of classification methods used in the next iteration classification; The accuracy calculation program is between step S30 and step S40; the screening program is between step S40 and step S50.

4. The classification method of claim 3, wherein the method further comprises: determining a plurality of classifiers; and training the plurality of classifiers based on the plurality of training samples. The screening procedure is as follows: the hit rate of each classification method is compared with its preset threshold value, and if the hit rate is less than the threshold value, the classification method is eliminated, otherwise the classification method is selected, and the multiple classification methods used in the next iteration classification are obtained by screening and updating.

Images