CN-116522134-B - Deep learning method and device based on symmetrical cross composite training

CN116522134BCN 116522134 BCN116522134 BCN 116522134BCN-116522134-B

Abstract

The invention relates to the technical field of deep learning, in particular to a novel deep learning method and device based on symmetrical cross compound training, wherein the method specifically comprises the steps of obtaining first sample data, dividing the first sample data into a first training set, a first verification set and a first test set, and dividing the first training set into a first data set and a second data set; taking the first data set and the second data set as training sets and verification sets in turn to perform symmetrical cross training to obtain outlier samples and true samples; and carrying out compound training on the outlier sample and the true sample, and simultaneously carrying out output correction classification to obtain a first optimal model. According to the invention, by changing the mode of the traditional deep learning training method and utilizing the function of deep learning feature extraction, the outlier samples in the training set are automatically identified and separated, so that the model performance is further improved.

Inventors

WANG XU
HE ZHAOSHUI
LIN ZHIJIE
TAN JI
SU WENQING
LIANG HAO

Assignees

广东工业大学

Dates

Publication Date: 20260508
Application Date: 20230324

Claims (6)

1. The deep learning method based on symmetrical cross composite training is characterized by comprising the following steps of: Acquiring first sample data, dividing the first sample data into a first training set, a first verification set and a first test set, and sub-dividing the first training set into a first data set and a second data set, wherein the first sample data is a CC-CCII data set or a HUST-19 data set, and the first training set comprises a positive CT image, a negative CT image and an informationless CT image; Performing symmetrical cross training on the first data set and the second data set serving as a training set and a verification set in turn to obtain an outlier sample and a true sample, wherein the outlier sample comprises a first outlier sample and a second outlier sample, and the true sample comprises a first true sample and a second true sample; performing compound training on the outlier sample and the true sample, and performing output correction classification at the same time to obtain a first optimal model, wherein the method specifically comprises the following steps: Combining the outlier sample and the true sample, and then marking and dividing according to a true positive sample, a true negative sample, a false positive sample and a false negative sample to obtain a fourth training set; Training the fourth training set based on a third neural network model, and obtaining a first optimal model according to the first verification set; Evaluating the first optimal model according to the first test set, and carrying out output correction classification through a classification formula; wherein the categorization formula satisfies ; ; Wherein, the The positive result of the test is indicated, The result of the test is indicated as negative, The positive outlier sample result output by the network is represented, And (5) representing a negative outlier sample result output by the network.
2. The method according to claim 1, wherein the symmetrically cross-training the first data set and the second data set as training sets and verification sets in turn, to obtain outlier samples and true samples, specifically comprises: Training the first data set and the second data set as a second training set and a second verification set respectively, and separating a first outlier sample and a first true sample; Training the second data set and the first data set as a third training set and a third verification set respectively, and separating a second outlier sample and a second true sample.
3. The method according to claim 2, wherein training the first data set and the second data set as a second training set and a second validation set, respectively, separates a first outlier sample and a first true sample, specifically comprising: Determining the first data set as a second training set, and determining the second data set as a second verification set; Training the second training set based on the first neural network model, and obtaining a second optimal model according to the second verification set; And applying the second optimal model to the second training set to determine a first outlier sample and a first true sample.
4. The method according to claim 2, wherein training the second dataset and the first dataset as a third training set and a third validation set, respectively, separates out a second outlier sample and a second true sample, comprising: determining the second data set as a third training set and the first data set as a third validation set; training the third training set based on a second neural network model, and obtaining a third optimal model according to the third verification set; and applying the third optimal model to the third training set to determine a second outlier sample and a second true sample.
5. The method of claim 1, wherein the fourth training set comprises ; ; Wherein, the Indicating that the total of the samples is true positive, Indicating that the total of the samples is true negative, Representing a total false positive sample of the sample, Represents a total false negative sample of the sample, Representing a true positive sample in the first true sample, Representing a true negative sample in the first true sample, Representing false positive samples in the first outlier sample, Representing a false negative sample in the first outlier sample, Representing a true positive sample in the second true sample, Representing a true negative sample in the second true sample, Representing false positive samples in the second outlier sample, Representing a false negative sample in the second outlier sample.
6. Deep learning device based on symmetrical cross compound training, which is characterized in that the device specifically comprises: The acquisition module is used for acquiring first sample data, dividing the first sample data into a first training set, a first verification set and a first test set, and subdividing the first training set into a first data set and a second data set, wherein the first sample data is a CC-CCII data set or a HUST-19 data set, and the first training set comprises a positive CT image, a negative CT image and an informationless CT image; The symmetric cross training module is used for performing symmetric cross training on the first data set and the second data set serving as a training set and a verification set in turn to obtain an outlier sample and a true sample, wherein the outlier sample comprises a first outlier sample and a second outlier sample, and the true sample comprises a first true sample and a second true sample; the compound training module is used for carrying out compound training on the outlier sample and the true sample, and simultaneously carrying out output correction and classification to obtain a first optimal model, and specifically comprises the following steps: Combining the outlier sample and the true sample, and then marking and dividing according to a true positive sample, a true negative sample, a false positive sample and a false negative sample to obtain a fourth training set; Training the fourth training set based on a third neural network model, and obtaining a first optimal model according to the first verification set; Evaluating the first optimal model according to the first test set, and carrying out output correction classification through a classification formula; wherein the categorization formula satisfies ; ; Wherein, the The positive result of the test is indicated, The result of the test is indicated as negative, The positive outlier sample result output by the network is represented, And (5) representing a negative outlier sample result output by the network.

Description

Deep learning method and device based on symmetrical cross composite training Technical Field The invention relates to the technical field of deep learning, in particular to a deep learning method and device based on symmetrical cross compound training. Background Deep learning, which benefits from its powerful feature extraction capability, exhibits excellent performance in the field of computer vision. Currently, in deep learning, a proper neural network model is generally obtained according to a traditional deep learning training strategy, and the method comprises the steps of 1) firstly dividing a data set into a training set, a verification set and a test set, 2) then training the deep neural network on the training set, obtaining a parameterized model through gradient optimization, and selecting an optimal model on the verification set, and 3) finally executing specific tasks such as classification and prediction on the test set. However, in the case of a limited dataset, some outlier samples in the training set will cause the neural network to suffer from over-fitting problems, i.e. outlier samples lead to the neural network over-learning, resulting in a reduced generalization ability of the model. For this problem, scholars have proposed various methods such as soft weight sharing, dropout strategy, and stopping training immediately when the performance of the model on the validation set begins to deteriorate. The last method is to use the verification set to avoid the over fitting of the deep learning model, so as to obtain a proper deep learning model, and the method can prevent the performance of the model from being reduced, but does not discuss the interference caused by the over fitting outlier sample on the deep learning network. Disclosure of Invention The invention aims to provide a deep learning method and device based on symmetrical cross compound training, which automatically identify and separate outlier samples in a training set by utilizing the function of deep learning feature extraction by changing the mode of the traditional deep learning training method, thereby further improving the model performance. In one aspect, the invention provides a deep learning method based on symmetrical cross composite training, which specifically comprises the following steps: acquiring first sample data, dividing the first sample data into a first training set, a first verification set and a first test set, and sub-dividing the first training set into a first data set and a second data set; Performing symmetrical cross training on the first data set and the second data set serving as a training set and a verification set in turn to obtain an outlier sample and a true sample, wherein the outlier sample comprises a first outlier sample and a second outlier sample, and the true sample comprises a first true sample and a second true sample; And carrying out compound training on the outlier sample and the true sample, and simultaneously carrying out output correction classification to obtain a first optimal model. Further, the step of performing symmetric cross training by taking the first data set and the second data set as training sets and verification sets in turn to obtain outlier samples and true samples specifically includes: Training the first data set and the second data set as a second training set and a second verification set respectively, and separating a first outlier sample and a first true sample; Training the second data set and the first data set as a third training set and a third verification set respectively, and separating a second outlier sample and a second true sample. Further, the training the first data set and the second data set as a second training set and a second verification set respectively, and separating a first outlier sample and a first true sample specifically includes: Determining the first data set as a second training set, and determining the second data set as a second verification set; Training the second training set based on the first neural network model, and obtaining a second optimal model according to the second verification set; And applying the second optimal model to the second training set to determine a first outlier sample and a first true sample. Further, the training the second dataset and the first dataset as a third training set and a third verification set, respectively, to separate a second outlier sample and a second true sample, specifically includes: determining the second data set as a third training set and the first data set as a third validation set; training the third training set based on a second neural network model, and obtaining a third optimal model according to the third verification set; and applying the third optimal model to the third training set to determine a second outlier sample and a second true sample. Further, the performing composite training on the outlier sample and the true sample, and performing output correction