Method for generating a knowledge graph library based on a combination of a scenario graph and a concept net

Abstract

The retrieval knowledge graph library generation method based on the combination of scene graph and concept net comprises the following steps: 1) model pre-training: pre-training input data on a neural network to detect the object types and their positions appearing in a picture; 2) scene graph training: performing unbiased training on the pre-training result of the model, and finally outputting a file related to scene graph information by applying a neural network model to predict the relationship between different types in the picture; 3) automatic expansion of the knowledge graph; 4) testing the trained scene graph model; 5) extracting and processing the file information related to the scene graph and the concept net, and then importing the file information into a retrieval database to finally form the retrieval database; merging the nodes and relationships with high similarity in the scene graph, and fusing the corresponding knowledge graph libraries of the scene graph and the concept net, and finally merging the scene graph and the concept net into a knowledge graph library containing all information; and 6) database regular update.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of computer vision and natural language processing, and particularly relates to a knowledge graph library generation method based on the combination of scene graphs and concept networks. BACKGROUND

[0002] The knowledge graph library technology is an important part of artificial intelligence technology, is a semantic network revealing the relationship between entities, is a graph-based data structure, and can formally describe the things and their mutual relationships in the real world. At present, it has been widely used in many fields such as intelligent search, intelligent question answering, education, and has very broad development prospects.

[0003] At present, the main method of constructing a knowledge graph is the top-down and bottom-up construction method. The top-down construction method is to determine the data model first, then fill in the specific data according to the model, and finally form the knowledge graph library; while the bottom-up construction method is to collect data in the form of three original groups, and then extract the data model according to the data content. However, in the process of constructing the knowledge graph, there are problems such as how to obtain high-quality knowledge and incomplete retrieval information.

[0004] At present, the main method of constructing a knowledge graph is the top-down and bottom-up construction method. The top-down construction method is to determine the data model first, then fill in the specific data according to the model, and finally form the knowledge graph library; while the bottom-up construction method is to collect data in the form of three original groups, and then extract the data model according to the data content. However, in the process of constructing the knowledge graph, there are problems such as how to obtain high-quality knowledge and incomplete retrieval information. SUMMARY

[0005] The present application overcomes the above-mentioned shortcomings of the prior art, and proposes a retrieval knowledge graph library generation method based on the combination of scene graphs and concept networks to improve the retrieval ability of the knowledge graph library.

[0006] In order to obtain high-quality knowledge, extract complete information from the database, and finally improve the retrieval ability of the knowledge graph, the scene graph and the concept net are combined. By obtaining the scene graph information on the image, and combining various knowledge on the concept net, a comprehensive knowledge graph library is generated. The scene graph is a general data structure, which is represented as a graph of an image. The target in the image corresponds to the graph node, and the relationship between the targets corresponds to the graph edge. For the practicality of the scene graph, the bias of the training needs to be adjusted. For the subject-predicate-object triplets, the predicate is mainly improved, and a more accurate verb is used to represent it, so as to solve the problem of inaccurate relationship detection in the scene graph. At the same time, the knowledge graph on the concept net is expanded, the nodes and relationships with high similarity in the scene graph are merged, and the corresponding knowledge graph library of the scene graph and the concept net is fused. The related knowledge of ontology is used to expand the keywords of user problem recognition, increase the search range of the problem, and improve the search completion rate of the question and answer system.

[0007] The technical scheme of the present application is as follows:

[0008] The scene graph and the concept net are combined to generate a knowledge graph library, which comprises the following steps:

[0009] 1) Model pre-training process:

[0010] 1.1) Data set processing stage, the MS-COCO data set is processed, and the pictures with specific categories are screened out, and the redundant pictures are filtered out;

[0011] 1.2) Pre-training on the classic neural network (Faster-CNN). First, the convolutional layer is used to extract the features of the input image, the region proposal network generates the region proposal, and then the region of interest is aligned according to the coordinates [x, y, w, h] provided by the feature map and the region proposal, and finally the fixed size feature map is generated, and finally the specific class classification is performed by using softmax;

[0012] 2) Training process of the scene graph:

[0013] 2.1) Based on the existing scene graph training method, the scene graph generation is unbiasedly trained using the neural network Neural-MOTIFS model. The process of scene graph unbiased training: using the traditional scene graph training method and then debiasing. First, use the fact-based causal graph training method, that is, the training framework of the normal model. X is the target feature, and the target label Z is predicted first. Finally, the predicate verb Y is predicted by the image I, the target feature X, and the target label Z. The form of predicting the predicate verb is (I, X, Z)→Y. Then, using the same model, different methods are used. Train using the original intervened causal graph method. Unlike the previous method, I→X is removed, that is, the target feature X is not affected by the picture I, and the label of the target label Z is also not determined. A virtual value is assigned to X, and then the predicate is inferred. According to the two predicate verbs Y obtained, the results of the two times are subtracted. The influence of bias can be eliminated by relying on the difference between the observed result Y(u) and its counterfactual alternative;

[0014] 2.2) Output the results to get the corresponding json file. For each image, the scene graph information is saved as a target containing the score of the target, the target label, the relationship between the two categories, the relationship label, the score of the relationship, and the probability of each object matching all 51 predicates;

[0015] 3) Automatic expansion method of knowledge graph:

[0016] 3.1) Obtain related files on the ConceptNet website;

[0017] 3.2) An automatic expansion method of knowledge graph is designed, and the specific steps are as follows: First, learn new words from websites such as Baidu Encyclopedia and Baidu Library, and keep up with the meaning represented by the new words in real time. Then, use the BERT model to extract the relationship between concepts and concepts. Position the position of the sentence and the two entities, extract the semantic features of the sentence and the features of the entities, and then concatenate the three features for relationship classification, and then use the softmax layer for classification. Finally, add the new concept to the original knowledge graph;

[0018] 4) Model testing process:

[0019] For the trained scene graph, accuracy (Accuracy), precision (Precision), recall (Recall), and F1_score are used as the measurement standard;

[0020] 5) Generation of retrieval database process:

[0021] 5.1) Process the file: After the above processing, we will get the corresponding json file, extract the content we want, and process it into a scene graph and a concept net corresponding node and relationship csv file;

[0022] 5.2) Import into the retrieval database: Import the csv file of the corresponding nodes and relationships of the scene graph and concept net into the retrieval database;

[0023] 5.3) Design similarity merging algorithm: In the knowledge graph library of the scene graph, merge nodes and relationships with high similarity. In the scene graph, there may be multiple nodes corresponding to the same object. First, count the frequency of each node appearing in the scene graph; second, the inverse document frequency represents the universality of the keyword, the larger the value, the better the discrimination ability for the node; then, the text word vector corresponding to the node can be calculated; finally, the cosine similarity is used to calculate the correlation between nodes. Merge nodes with high similarity. Based on this method, after merging nodes with high similarity, the same method can be used to merge relationships with high similarity.

[0024] 5.4) Design data connection component: In the retrieval database, fuse the corresponding knowledge graph library of the scene graph and concept net. Connect the nodes that represent the same meaning, at this time the similarity of the nodes determines whether two nodes can be connected. First, use the Canopy clustering algorithm to get the value of k. Then, use the K-means clustering algorithm, which takes the number of clusters k and all nodes in the knowledge graph as input, and finally separates nodes with high similarity. Connect the corresponding nodes in the scene graph and concept net, use the relationships in the ontology to expand the keywords of user question recognition, increase the search range of the question, improve the search completion rate of the question and answer system, and form the final database. When we search for the actual situation between a certain category and another category, we will extend the category to various small categories under the category, which can provide us with more detailed information;

[0025] 6) Database timing update process:

[0026] This database can be updated online using the incremental learning method. In this database, the incremental learning method based on playback is used. Part of the data of the old task needs to be retained, and the old data is used to train the model together with the new data.

[0027] The technical concept of the present application is that when constructing a knowledge graph, there are problems of not obtaining high-quality data and incomplete retrieval data. Based on this situation, the present application uses unbiased training of a scene graph, and simultaneously automatically expands the knowledge graph on a concept network, combining the knowledge of the scene graph and the concept network. First, the model pre-training is used to detect the object type and its position in the picture, and then the scene graph is unbiasedly trained to obtain subject-predicate-object triple information. Second, the knowledge graph obtained from the concept network is automatically expanded, and then the scene graph and the concept network files are imported into the retrieval database, the nodes and relationships with high similarity in the scene graph are merged, and the corresponding knowledge graph library of the scene graph and the concept network is fused. Finally, the database can be updated regularly to enhance the applicability of the knowledge graph.

[0028] The beneficial effects of the present application mainly include: 1) using a classic neural network (Faster-CNN) for model pre-training to effectively detect the object type and its position; 2) using unbiased training to train the scene graph in the scene graph training process to accurately predict the relationship between two targets; 3) in the automatic expansion process of the knowledge graph, learning new words from the website, using the BERT model to extract the relationship between concepts, and adding to the original knowledge graph; 4) in the process of generating a retrieval knowledge graph library, merging the nodes and relationships with high similarity in the scene graph, and fusing the corresponding knowledge graph library of the scene graph and the concept network to make the content retrieval more comprehensive; 5) in the process of regularly updating the database, an incremental learning method based on playback is used. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 It is a schematic diagram of the method of the present application.

[0030] Figure 2 It is a schematic diagram of the fact-based causal graph training method based on unbiased training of the present application.

[0031] Figure 3 It is a schematic diagram of the intervened original causal graph training method based on unbiased training of the present application.

[0032] Figure 4 It is a flowchart of the knowledge graph automatic expansion method of the method of the present application.

[0033] Figure 5 It is a flowchart of the step of generating a retrieval database of the method of the present application. DETAILED DESCRIPTION

[0034] The specific embodiments of the present application are described in further detail below in conjunction with the drawings of the specification.

[0035] REFERENCE Figures 1-5, the retrieval knowledge graph library generation method based on the combination of scene graph and concept net, the steps are as follows:

[0036] 1) Model pre-training: data comes from the publicly available MS-COCO dataset, which is a large-scale object detection, segmentation, key point detection and captioning dataset. After inputting the dataset, pre-training is performed on the classic neural network (Faster-CNN) to detect the object categories and their positions appearing in the picture, and the algorithm steps are as follows:

[0037] 1.1) Data set processing stage, pre-processing the MS-COCO data set, screening pictures with specific categories, and filtering out redundant pictures;

[0038] 1.2) Pre-training on the classic neural network (Faster-CNN). First, the convolutional layer is used to extract the features of the input image, the region proposal network generates region proposals, and then the region of interest is aligned according to the coordinates [x, y, w, h] provided by the feature map and the region proposal, and finally the fixed size feature map is generated. Finally, use softmax to classify specific categories.

[0039] 2) Scene graph training: unbiased training is performed on the results of model pre-training, and a file related to the scene graph is finally output by applying the neural network model to predict the relationship between different categories in the picture, and the training process steps are as follows:

[0040] 2.1) Based on the existing scene graph training method, use the neural network Neural-MOTIFS model to perform unbiased training on the scene graph generation. The process of scene graph unbiased training: use the traditional scene graph training method and then remove the bias. First, use the fact-based causal graph training method, that is, the normal model training framework. X is the target feature, first predict the target label Z, and finally predict the predicate Y from the image I, the target feature X, and the target label Z. The form of predicting the predicate is (I, X, Z)→Y. The formula of the training loss is as follows:

[0041]

[0042] Among them, the cross-entropy loss of the target label Z and the predicate label Y is used for training.

[0043] Then, using the same model, use different methods. Use the original causal graph method with intervention to train. Different from the previous method, I→X is removed, that is, the target feature X is not affected by the picture I, nor does it determine the label of the target label Z. Assign a virtual value to X, and then infer what the predicate is. According to the two predicate verbs Y obtained, subtract the two results. It can rely on the observed results Y(u) and its counterfactual substitute Y x,z(u)The difference between the two is used to eliminate the influence of bias, and the formula is as follows:

[0044] TDE=Y (u) -Y x,z(u) (2)

[0045] Where TDE will be the final predicate score of the unbiased scene graph, Y (u) , Y x,z(u) The distribution is the first and second predicate verb;

[0046] 2.2) Output results, get the corresponding json file. For each image, the scene graph information is saved as a target that contains the score of the target, the target label, the relationship between the two categories, the relationship label, the score of the relationship, and the probability of each object matching all 51 predicates.

[0047] 3) Automatic expansion of knowledge graph: based on the concept network knowledge graph, it can be expanded. New knowledge can be obtained through new words and added to the knowledge graph, and the expansion process steps are as follows:

[0048] 3.1) Obtain relevant files from the concept network website;

[0049] 3.2) Design an automatic expansion method for the knowledge graph, and the specific steps are as follows: First, learn new words from websites such as Baidu Encyclopedia and Baidu Library, and keep up with the meaning represented by the new words in real time. Then, use the BERT model to extract the relationship between concepts and concepts. Position the position of the sentence and the two entities, extract the semantic features of the sentence and the features of the entities, and then concatenate the three features for relationship classification. The concatenation and classification formula is as follows:

[0050] h”=W[concat(H'0,H'1,H'2)]+b (3)

[0051] Where h” is a vector obtained finally, H'0, H'1, H'2 represent a sentence and two entity vectors respectively, concat is concatenation, W is the weight coefficient, and b is the bias coefficient.

[0052] Then use the softmax layer for classification, the formula is as follows:

[0053] p=softmax(h”) (4)

[0054] Where h” is the vector obtained from the above formula, and p is the final output relationship classification result.

[0055] Finally, add the new concept to the original knowledge graph.

[0056] 4) Model testing: For the trained scene graph model, use accuracy, precision, recall and F1_score to evaluate the performance of the proposed method, and the test method steps are as follows:

[0057] For the trained scene graph, use accuracy, precision, recall and F1_score as the measurement standard:

[0058]

[0059]

[0060]

[0061]

[0062] Where TP i , TN i , FP i , FN i respectively represent the true positive, true negative, false positive and false negative of the ith label in the label set.

[0063] 5) Generation of retrieval database: Extract and process the file information related to scene graph and concept net, and then import it into the retrieval database, and finally form the retrieval database. Merge the nodes and relationships with high similarity in the scene graph, and fuse the knowledge graph library corresponding to the scene graph and concept net, and finally merge the scene graph and concept net into a knowledge graph library containing all information, and the generation process is as follows:

[0064] 5.1) Processing files: After the above processing, we will get the corresponding json file, extract the content we want, and process it into csv files of scene graph and concept net corresponding nodes and relationships;

[0065] 5.2) Import to retrieval database: Import the csv files of scene graph and concept net corresponding nodes and relationships into the retrieval database;

[0066] 5.3) Design similarity merging algorithm: In the knowledge graph library of scene graph, merge the nodes and relationships with high similarity. In the scene graph, there may be multiple nodes corresponding to the same object. First, count the frequency of each node appearing in the scene graph, and the formula for counting the frequency of each node is as follows:

[0067]

[0068] Where, n iis the number of occurrences of a node, Σ k n k is the total number of nodes, TF i is the frequency of occurrence of a node.

[0069] Inverse document frequency represents the universality of a keyword, the larger the value, the better the distinguishing ability for the node, the formula is as follows:

[0070]

[0071] Where N represents the number of pictures, |j:n i ∈N j | represents the number of pictures including the node n, 1 is used to prevent the case that the number of pictures is 0.

[0072] The text word vector corresponding to a node can be calculated by the above two formulas, the formula is as follows:

[0073] TF-IDF = TF·IDF (11)

[0074] Finally, the cosine similarity is used to calculate the correlation between nodes, the formula is as follows:

[0075]

[0076] Finally, the nodes with high similarity are merged. Based on this method, after merging the nodes with high similarity, the same method can be used to merge the relationships with high similarity;

[0077] 5.3) Design data connection components: In the retrieval database, the knowledge graph library corresponding to the scene graph and the concept net is fused. The nodes representing the same meaning are connected, and the similarity of the nodes determines whether two nodes can be connected. First, the k value is obtained by using the Canopy clustering algorithm. Then, the K-means clustering algorithm is used, which takes the number of clustering clusters k and all nodes X = {x1, x2,..., x n} in the knowledge graph as input, the formula is as follows:

[0078]

[0079] Where x i represents an input node, C = {c1, c2,..., c k} is the output cluster, u j represents the sample mean of the jth cluster c j . This algorithm will select the appropriate u j to minimize the within-cluster sum of squares, and the highest similarity J is obtained.

[0080] Finally, the nodes with high similarity are divided, and the corresponding nodes in the scene graph and the concept network are connected. The keywords for user question recognition are expanded using the relationships in the ontology, the search range of the question is increased, the search completion rate of the question and answer system is improved, and the final database is formed. When we retrieve the actual situation of a certain category and another category having a certain specific relationship, we will extend the category to various subcategories under the category, which can provide us with more detailed information.

[0081] 6) Database timing update: The incremental learning method is used to realize the online update of the database, and the information of the database can be increased. The update process is as follows:

[0082] The database can be updated online using the incremental learning method. The incremental learning method based on playback is used in the database. Part of the data of the old task needs to be retained, and the old data is used to train the model together with the new data.

[0083] Specific experimental example:

[0084] This example can obtain the specific scene picture that the user wants. First, the user inputs the scene graph content that he wants to query in the form of text, such as “people looking at book”. This example splits the triplets, identifies the subject people, the predicate looking at, and the object book. Therefore, the query statement of the database can be constructed as “MATCH (n: SG {name: "people"})-[r: looking at]-> (n1: SG {name: "book"}) return n, r, n1”. By running this statement in the knowledge graph retrieval database, the scene graph and the concept network are combined, and the similarity merging algorithm is designed. When the input query node is people, the node will be expanded, and nodes with names such as children, kid, man, woman, and person related to people will be retrieved. At the same time, when the input query relationship is looking at, the relationship will be expanded, and relationships with names such as watching and visit similar to looking at will be retrieved. Similarly, book is the same. Therefore, more comprehensive retrieval results can be obtained, and various specific information of the picture can be obtained through retrieval. Finally, the picture corresponding to the scene can be obtained.

[0085] The embodiments of the present specification are merely a list of implementation forms of the inventive concept, and the protection scope of the present application should not be regarded as being limited to the specific forms stated in the embodiments, and the protection scope of the present application also encompasses equivalent technical means that can be thought of by those skilled in the art according to the inventive concept.

Claims

1. A method for generating a knowledge graph library based on the combination of a scene graph and a concept net, characterized in that, Comprise the following steps: 1) Model pre-training: data from publicly available MS-COCO dataset, which is a large object detection, segmentation, key point detection and captioning dataset; after inputting the dataset, pre-training on the classic neural network Faster-CNN, detecting the object category and its location appearing in the picture; 2) Scene graph training: unbiased training of the results of model pre-training, and applying a neural network model to finally output a file related to scene graph information, predicting the relationship between different categories in the picture; Specifically comprising: 2.1) Based on the existing scene graph training method, the Neural-MOTIFS model is used for unbiased training of scene graph generation; the process of scene graph unbiased training: using the traditional scene graph training method and then debiasing; first, using the fact-based causal graph training method, that is, the training framework of the normal model; X is the target feature, first predict the target label Z, and finally predict the predicate Y from the image I, the target feature X, the target label Z; the form of predicting the predicate is (I, X, Z)→Y; the formula of training loss is as follows: （1） Where, the cross-entropy loss of target label Z and predicate label Y is used for training; Then, using the same model, a different method; using the intervened original causal graph method to train; unlike the previous method, I→X is removed, that is, the target feature X is not affected by the picture I, nor does it determine the label of the target label Z, a virtual value is assigned to X, and then the predicate is inferred; according to the two predicate verbs Y obtained, the results of the two times are subtracted; the influence of bias can be eliminated by relying on the difference between the observed results and its counterfactual alternative , the formula is as follows: （2） wherein, the final predicate score as an unbiased scene graph, , the distribution is the first and second predicate verbs; 2.2) Output results, get the corresponding json file; for each image, the scene graph information is saved as containing the target, the score of the target, the target label, the relationship between the two categories, the relationship label, the score of the relationship, and the probability of each object matching all 51 predicates; 3) Automatic expansion of knowledge graph: based on the concept net knowledge graph, it can be expanded; new knowledge can be obtained through new words and added to the knowledge graph; specifically comprising: 3.1) Obtain related files on the concept net website; 3.2) An automatic expansion method of knowledge graph is designed, and the specific steps are as follows: first, learn new words from Baidu Baike and Baidu Wenku website, and follow up the meaning represented by the new words in real time, then use BERT model to extract the relationship between concepts and concepts; locate the position of the sentence and the two entities, extract the semantic features of the sentence and the features of the entities, and then splice the three features for relationship classification, the splicing and classification formula is as follows: （3） wherein, is a vector obtained finally, respectively represent a vector of 1 sentence and 2 entities, is spliced, is a weight coefficient, is a bias coefficient; Then the data is classified using the layers, with the formula as follows: （4） wherein, is the final output relationship classification result; Finally, add the new concept to the original knowledge graph; 4) Model testing: test the trained scene graph model, use accuracy (Accuracy), precision (Precision), recall (Recall) and F1_score to evaluate the performance of the proposed method; 5) Generate retrieval database: extract and process the file information related to scene graph and concept net, and then import it into the retrieval database, and finally form the retrieval database; merge the nodes and relationships with high similarity in the scene graph, and fuse the knowledge graph library corresponding to the scene graph and the concept net, and finally merge the scene graph and the concept net into a knowledge graph library containing all information; 6) Database is updated regularly: using the incremental learning method can realize the online updating of the database, and can increase the information of the database.

2. The method of claim 1, wherein the method is based on a combination of a scene graph and a concept net. Step 1) The model pre-training includes: 1.1) Data set processing stage, processing MS-COCO data set, screening out pictures with specific categories, filtering out redundant pictures; 1.2) Pre-training on classic neural network Faster-CNN; first, extract the features of the input image with convolutional layers, the region proposal network generates region proposals, and then the region of interest alignment is performed according to the coordinates [x, y, w, h] provided by the feature map and the region proposal, to generate a fixed-size feature map, and finally the specific class classification is performed using the last fully connected layer.

3. The method of claim 1, wherein the method further comprises: generating a knowledge graph library by combining the scene graph and the concept net. Step 4) The model test includes the following process: ​ For the trained scene graph, use accuracy (Accuracy), precision (Precision), recall (Recall) and F1_score as the measurement standard: （5） （6） （7） （8） wherein respectively denote the true positive, true negative, false positive and false negative of the i-th label in the label set.

4. The method of claim 1, wherein the method further comprises: generating a knowledge graph library by combining the scene graph and the concept net. Step 5) The generation of retrieval database includes the following process: ​ 5.1) Process the file: after the above processing, we will get the corresponding json file, extract the content we want, process into scene graph and concept net corresponding node and relationship csv file; 5.2) Import to the retrieval database: import the scene graph and concept net corresponding node and relationship csv file into the retrieval database; 5.3) Design similarity merging algorithm: in the knowledge graph library of scene graph, merge nodes and relationships with high similarity; In the scene graph, there may be multiple nodes corresponding to the same object. First, count the frequency of each node appearing in the scene graph. The formula for calculating the frequency of each node is as follows: （9） wherein, is the number of occurrences of a node, is the total number of nodes, is the frequency of occurrence of a node; Inverse document frequency represents the universality of keywords. The larger the value, the better the discrimination ability for the node. The formula is as follows: （10） wherein, represents the number of pictures, represents the number of pictures including the node 1 is used to prevent the case where the number of pictures is 0; The text word vector corresponding to a node can be calculated by the above two formulas, as follows: （11） Finally, use cosine similarity to calculate the relevance between nodes. The formula is as follows: （12） Finally, merge nodes with high similarity; Based on this method, after merging nodes with high similarity, the same method can be used to merge relationships with high similarity; 5.4) Design data connection components: In the retrieval database, the scene graph and the concept net corresponding knowledge graph library are fused; the nodes representing the same meaning are connected, and the similarity of the nodes determines whether the two nodes can be connected; first, the k value is obtained by using the Canopy clustering algorithm; then, the K-means clustering algorithm is used, which takes the number of clustering clusters k and all nodes in the knowledge graph As input, the formula is as follows: （13） where, represents a certain node input, is the output cluster, represents the sample mean of the first cluster ; the algorithm will select the appropriate so that the sum of squares within the cluster is minimum, resulting in the highest similarity J; Finally, divide the nodes with high similarity, connect the corresponding nodes in the scene graph and concept net, use the relationship in ontology to expand the keywords of user question recognition, increase the search range of the question, improve the search completion rate of the question and answer system, and form the final database; When we search for a specific relationship between a category and another category, we will extend the category to various subcategories under the category, so that we can get more detailed information.

5. The method of claim 1, wherein the method further comprises: generating a knowledge graph library by combining the scene graph and the concept net. Step 6) The database is updated regularly, including the following process: ​ This database uses the method of incremental learning to update online; The method of incremental learning based on playback is used in this database; Part of the data of the old task needs to be retained, and the old data is used to train the model together with the new data.

Images