CN108052680A - Image data target identification Enhancement Method based on data collection of illustrative plates, Information Atlas and knowledge mapping - Google Patents

Abstract

The present invention be it is a kind of based on data collection of illustrative plates, Information Atlas, knowledge mapping framework image data target identification Enhancement Method.It is mainly used for solving the problem of image recognition for not marking classification in conventional images recognition methods None- identified training set, belongs to Distributed Calculation and Software Engineering technology crossing domain.Key is from the existing image type recognition result based on deep learning method, three layers of collection of illustrative plates are built according to existing image resource, unidentified image category is subjected to characteristic matching in data collection of illustrative plates and obtains initial matching result, identified image category is subjected to relationship match on Information Atlas and obtains intermediate match result, indirect interaction relationship match is finally carried out in knowledge mapping, the confidence level of intermediate match result and sequence are calculated, recommends the highest matching image category of User reliability.

Description

Image data target identification enhancement method based on data map, information map and knowledge map

Technical Field

The invention relates to an image data target identification enhancement method based on a data map, an information map and a knowledge map framework. The method is mainly used for solving the problem that the existing image recognition method cannot recognize the image recognition of the unmarked category in the training set, and belongs to the cross field of distributed computing and software engineering technology.

Background

Knowledge-graph has become a powerful tool for representing knowledge in the form of labeled directed graphs and can impart semantics to textual information. A knowledge graph is a graph constructed by representing items, entities or users in the form of nodes and linking nodes that interact with each other in the form of edges, where the edges between nodes can represent any semantic relationship. The construction of the knowledge graph is divided into three levels of information extraction, knowledge fusion and knowledge processing according to the knowledge acquisition process, the defined knowledge graph is a mesh knowledge base formed by connecting entities with attributes through relations, and the research value of the knowledge graph is that information accumulated in the Internet can be organized into available knowledge at the minimum cost, so that concept retrieval and graphical knowledge display are realized through reasoning. Data is obtained by observing numbers or other basic individual items. Information is conveyed by the context of data and data combinations, suitable for analysis and interpretation. Knowledge is a general understanding and experience derived from the accumulated information from which new context can be inferred.

The most traditional image recognition technology is manual recognition, a marked image can be recognized by a deep learning method, an equation is gradually simulated by a machine in the process of repeatedly recognizing the same type of object, the equation is close to the common characteristic of the recognized object, and the recognition purpose is finally achieved, but the recognition problem of the unmarked image cannot be solved by the deep learning method. The invention mainly identifies the unidentified objects as much as possible by the aid of the data map, the information map and the knowledge map.

Disclosure of Invention

The technical problem is as follows: the invention relates to an image data target identification enhancement method based on a data map, an information map and a knowledge map framework. And introducing a data map, an information map and a knowledge map framework to carry out knowledge reasoning on the unidentified image, thereby realizing automatic and intelligent recognition of the image in the picture or shot by the camera. The structure, characteristics, frequency, interaction relation and semantic relation of the image are respectively placed in the data map, the information map and the knowledge map, and the identification result with the highest credibility is given by means of the framework.

The technical scheme is as follows: the method is a strategic method, can be applied to image target recognition of pictures or cameras, and is beneficial to solving the problem that unidentified images cannot be recognized in current machine learning. The structure, the characteristics, the frequency, the interactive relation and the semantic relation of the image are respectively placed in a data map, an information map and a knowledge map, and then the unidentified image is compared with the data in the data map to obtain an initial matching result; then, the identified images are compared with the interactive relation in the information map, and an intermediate matching result is obtained through calculation; and finally, carrying out knowledge reasoning in the knowledge graph to obtain an indirect interaction relationship, wherein the highest credibility is the final result identified by the people.

The system structure is as follows: the method comprises the steps of firstly establishing a data map, an information map and a knowledge map framework aiming at a recognized image, traversing the data map aiming at An unrecognized image A to obtain a data matching degree R (Ai) with Ai, and selecting initial matching results A1, A2, \ 8230An according to the sequence of R (Ai) from large to small; and (3) carrying out relation matching on the initial matching result and the identified image set { B } in an information map to obtain relation matching degrees Y (Ai) and X (Ai), selecting A1, A2, \8230Akaccording to the sequence of X (Ai) from large to small, and carrying out calculation and inference on Ai' and { B } in the knowledge map to obtain a result with the highest reliability. At present, the identified images can be identified by deep learning, and on the basis, the invention introduces a data map, an information map and a knowledge map framework to carry out knowledge reasoning on the unidentified images, thereby realizing the automatic and intelligent identification of the images in the pictures or the images shot by the camera. Specific explanations for constructing the data map, the information map and the knowledge map are given below:

data mapping: data maps can record basic attributes in an image entity, including color morphology, etc., but they are not meaningful by themselves without contextual context. The data map can be expressed by data structures such as arrays, linked lists, queues, trees, stacks, graphs and the like. The data map may also record the frequency of appearance of structures contained in the image entities, including the frequency of the three levels of structure, time and space. The data map can describe the degree of closeness of the association of different image entities, which is called as density, and can reflect which entities are closely associated and which entities are sparsely associated. However, the data map can only be used for static analysis, and cannot express the interaction relationship between entities. Meanwhile, the data map does not analyze the accuracy of the data, different entities may appear but represent the same meaning, such as tomatoes and tomatoes, and the attribute frequencies of the two image entities are the same, so that data redundancy is generated;

an information map: information is conveyed by context after data and data combination, information that is suitable for analysis and interpretation after concept mapping and correlation combination. The information map may be expressed according to a relational database. And (4) performing data cleaning on the information map to eliminate redundant data. The information map can record the interaction relation between the entities;

knowledge graph: the knowledge graph further improves the semantic relation between the entities according to the data graph and the information graph, the edge density and the node density of the knowledge graph are improved through information reasoning and entity linking, and the unstructured characteristic of the knowledge graph enables the knowledge graph to be seamlessly linked. Information reasoning needs to be supported by relevant rules, and the accuracy Cr of a new relationship obtained through reasoning can be obtained by formula 1. The new relationship between entity 1 and entity 2 may be represented asAnd z represents all of the relationships,representing a new relationship weight, representing a relationship between entity 1 and entity 2, which relationship is considered valid when accuracy exceeds a threshold

（1）。

Has the advantages that: the invention relates to an image data target identification enhancing method based on a data map, an information map and a knowledge map framework, which has the following remarkable advantages:

(1) According to the invention, a three-layer framework of a data map, an information map and a knowledge map is constructed, and the three-layer framework is progressive layer by layer, so that efficient identification is realized;

(2) The method can solve the problem of the image category of the unlabeled category in the training set and recommend the identification result with high reliability to a user;

(3) The invention carries out data matching and relation matching on the image and the map, continuously reduces the identification range and obtains the most accurate result through knowledge reasoning.

Drawings

FIG. 1 is a formal definition of data, information and knowledge maps.

FIG. 2 is a detailed flow diagram of an image recognition process based on a data-graph, information-graph, and knowledge-graph architecture.

Detailed Description

The specific flow of the image identification method based on the data map, the information map and the knowledge map framework is as follows:

step 1) corresponding to 001 in fig. 2, a frame based on a data map, an information map and a knowledge map is established according to the existing image resources;

step 2) acquiring an image to be recognized corresponding to 002 in fig. 2;

step 3) corresponding to 003 in fig. 2, the image to be recognized is divided into two modules, namely a recognized image set { Bi } and an unrecognized image A;

and 4) corresponding to 004 in fig. 2, performing data matching on the unidentified image A obtained in the step 3 and the entity Ai of the data map, and obtaining the matching degree R according to the formula 1. Let the entity attribute (including structure, color, feature, frequency, local structure, etc.) in the data map and Ai match value be，Representing one attribute that Ai can match with an unidentified image, x represents all attributes that can be matched:

（1）；

step 5) corresponding to 005 in fig. 2, setting An initial value n according to experience, and selecting A1, A2, \ 8230An according to the sequence of R (Ai) from large to small;

step 6) corresponding to 006 in fig. 2, A1, A2, \ 8230An is taken as An initial matching result;

step 7) corresponds to 007 in the graph 2, the information map is traversed by combining the unidentified image set { B } obtained in the step 3 based on the initial matching result obtained in the step 6, the corresponding interaction relation is found, and the relation matching degrees Y (A1), Y (A2) \ 8230and Y (An) of A1, A2, \ 8230and { B } are obtained through formulas 2 and 3. The matching degree of the interactive relation p between Bi and Ai in the information map is assumed to be，Where an interaction is represented, q represents all relationships:

（2）；

（3）；

step 8) corresponds to 008 in fig. 2, the intermediate matching degrees X (A1), X (A2) \8230, X (An) are obtained from formula 4, where R (Ai) and Y (Ai) are independent of each other and R1+ R2=1,

（4）；

step 9) corresponding to 009 in fig. 2, setting an initial value k (k < = n) according to experience, and selecting A1, A2, \ 8230and Ak according to the sequence of X (Ai) from large to small;

step 10) corresponds to 010 in fig. 2, based on the intermediate matching results A1, A2, \8230Akobtained in step 8, the indirect interaction relationship between { B } and A1, A2, \8230Akcan be obtained through reasoning in the knowledge graph, and the accuracy Cr of the obtained new relationship can be obtained through formula 5. The new relationship between entity 1 and entity 2 may be represented asAnd z represents all new indirect interactions,the degree of matching of the new relationship is represented,represents an indirect relationship between entity 1 and entity 2, which is considered valid when the accuracy exceeds a threshold:

（5）；

step 11) corresponds to 011 in FIG. 2, computing a Confidence level (Confidence) if the relationship holds, whereinAnd the interaction relationship in the information map is independent of each other, and y1+ y2=1, which is obtained by formula 6:

（6）；

step 12) corresponds to 012 in fig. 2, and the image entity with the highest degree of reliability is the final result of recognition by comparing the degrees of reliability.

Claims (1)

1. The invention mainly introduces a data map, an information map and a knowledge map framework to carry out knowledge reasoning on unidentified images so as to realize automatic and intelligent recognition of images in pictures or images shot by a camera, the invention respectively puts the structure, characteristics, frequency, interactive relation and semantic relation of the images in the data map, the information map and the knowledge map, and gives a recognition result with the highest credibility by means of the framework, and the specific flow is as follows:

step 1) establishing a framework based on a data map, an information map and a knowledge map according to the existing image resources;

step 2) acquiring an image to be identified;

step 3) dividing the image to be identified into two modules, namely an identified image set { Bi } and an unidentified image A;

step 4) carrying out data matching on the unidentified image A obtained in the step 3 and an entity Ai of the data map, and obtaining a matching degree R according to a formula 1;

let the entity attribute (including structure, color, feature, frequency, local structure, etc.) and Ai match values in the data map be，Representing one attribute that Ai can match with an unidentified image, x represents all attributes that can be matched:

（1）；

step 5) setting An initial value n according to experience, and selecting A1, A2, \8230Anaccording to the sequence of R (Ai) from large to small;

step 6), taking A1, A2, \8230andan as initial matching results;

step 7) traversing the information map by combining the unidentified image set { B } obtained in the step 3 based on the initial matching result obtained in the step 6, finding out the corresponding interactive relation, and obtaining the relation matching degrees Y (A1), Y (A2) \ 8230and Y (An) of A1, A2, \ 8230An and { B } by formulas 2 and 3;

the interactive relation p between Bi and Ai in the information map is assumed to be matching degree，Where an interaction is represented, q represents all relationships:

（2）；

（3）；

step 8) obtaining the intermediate matching degrees X (A1), X (A2) \8230andX (An) by a formula 4, wherein R (Ai) and Y (Ai) are mutually independent, and R1+ R2=1,

（4）；

step 9) setting an initial value k (k < = n) according to experience, and selecting A1, A2, \ 8230and Ak according to the sequence of X (Ai) from large to small;

step 10) obtaining an indirect interaction relation between { B } and A1, A2, \ 8230Ak through reasoning in a knowledge graph based on the intermediate matching results A1, A2, \8230Akobtained in the step 8, wherein the accuracy Cr of the obtained new relation can be obtained through a formula 5;

the new relationship between entity 1 and entity 2 mayIs represented byAnd z represents all new indirect interaction relationships,the degree of matching of the new relationship is represented,representing an indirect relationship between entity 1 and entity 2, which is considered valid when the accuracy exceeds a threshold:

（5）；

step 11) if the relationship is true, calculating Confidence level (Confidence), whereinAnd the interactions in the information graph are independent of each other, and y1+ y2=1, which is given by equation 6:

（6）；

and step 12) comparing the credibility, wherein the image entity with the highest credibility is the final result of the recognition.