CN117935021B - Scene fire image analysis model training method and system based on deep learning - Google Patents

Abstract

The invention discloses a scene fire image analysis model training method and system based on deep learning, comprising the following steps: firstly, multi-dimensional fire attribute data corresponding to sample fire image data and preset fire scene element information are obtained, and then the information is used for carrying out fire scene category analysis training on an initial model, so that a target scene fire image analysis model is obtained. By the design, the fire scene type analysis result of the fire scene can be deduced according to the input fire image data, and the accuracy of the fire scene type analysis can be ensured by continuously optimizing and updating the model through training.

Description

Scene fire image analysis model training method and system based on deep learning

Technical Field

The invention relates to the technical field of artificial intelligence, in particular to a scene fire image analysis model training method and system based on deep learning.

Background

Fire detection and analysis is a critical security area in which complex techniques such as image processing and machine learning are involved.

Conventional fire detection methods typically rely on thermal or smoke detectors, but these devices may not be able to detect the source of the fire immediately at the beginning of the fire.

In recent years, with the development of deep learning and computer vision technologies, more researches have begun to utilize these technologies for fire detection.

For example, flames or smoke in an image are identified by training a deep learning model.

However, it remains a challenge to efficiently train such models and utilize the predictions of the models for practical operation.

Disclosure of Invention

The invention aims to provide a scene fire image analysis model training method and system based on deep learning.

In a first aspect, an embodiment of the present invention provides a training method for a scene fire image analysis model based on deep learning, including:

Acquiring first multidimensional fire attribute data corresponding to first sample fire image data and first preset fire scene element information corresponding to the first sample fire image data; the first preset fire scene element information is obtained by performing fire scene category analysis on the first sample fire image data based on a first scene fire image analysis model and the first multidimensional fire attribute data corresponding to the first sample fire image data; the first scene fire image analysis model is obtained by performing scene fire type analysis training on the first initial model based on second multidimensional fire attribute data corresponding to second sample fire image data and second preset scene element information corresponding to the second sample fire image data; the second preset fire scene element information is obtained by performing fire scene category analysis on the second sample fire image data based on a second scene fire image analysis model and a second sample fire image type corresponding to the second sample fire image data;

loading the first multidimensional fire attribute data to a second initial model for fire scene category analysis, and obtaining first inferred fire scene element information corresponding to the first sample fire image data;

And training the second initial model according to the first inferred fire scene element information and the first preset fire scene element information to obtain a target scene fire image analysis model.

In a second aspect, an embodiment of the present invention provides a server system, including a server, where the server is configured to perform the method described in the first aspect.

Compared with the prior art, the invention has the beneficial effects that: by adopting the training method and the training system for the scene fire image analysis model based on deep learning, disclosed by the invention, the object scene fire image analysis model is obtained by acquiring the multidimensional fire attribute data corresponding to the sample fire image data and the preset scene element information, and then performing scene category analysis training on the initial model by utilizing the information.

By the design, the fire scene type analysis result of the fire scene can be deduced according to the input fire image data, and the accuracy of the fire scene type analysis can be ensured by continuously optimizing and updating the model through training.

Drawings

In order to more clearly illustrate the technical solution of the embodiments of the present invention, the drawings that are required to be used in the embodiments will be briefly described.

It is appreciated that the following drawings depict only certain embodiments of the invention and are therefore not to be considered limiting of its scope.

Other associated drawings may also be made by those of ordinary skill in the art without undue burden from these drawings.

FIG. 1 is a schematic flow chart of the training method of the scene fire image analysis model based on deep learning according to the embodiment of the invention;

Fig. 2 is a schematic block diagram of a computer device according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

It will be apparent that the described embodiments are some, but not all, embodiments of the invention.

The components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.

The following describes specific embodiments of the present invention in detail with reference to the drawings.

In order to solve the technical problems in the prior art, fig. 1 is a schematic flow chart of a training method for a scene fire image analysis model based on deep learning according to an embodiment of the disclosure, and the training method for the scene fire image analysis model based on deep learning is described in detail below.

Step S201, first multidimensional fire attribute data corresponding to first sample fire image data and first preset fire scene element information corresponding to the first sample fire image data are obtained; the first preset fire scene element information is obtained by performing fire scene category analysis on the first sample fire image data based on a first scene fire image analysis model and the first multidimensional fire attribute data corresponding to the first sample fire image data; the first scene fire image analysis model is obtained by performing scene fire type analysis training on the first initial model based on second multidimensional fire attribute data corresponding to second sample fire image data and second preset scene element information corresponding to the second sample fire image data; the second preset fire scene element information is obtained by performing fire scene category analysis on the second sample fire image data based on a second scene fire image analysis model and a second sample fire image type corresponding to the second sample fire image data;

Step S202, loading the first multi-dimensional fire attribute data to a second initial model for fire scene category analysis, and obtaining first inferred fire scene element information corresponding to the first sample fire image data;

Step S203, training the second initial model according to the first inferred fire scene element information and the first preset fire scene element information, to obtain a target scene fire image analysis model.

In an exemplary embodiment of the invention, a researcher collects a large amount of fire-related image data, such as a photograph of a building when a fire occurs or a fire image captured by a camera.

At the same time, they are also annotated with multidimensional fire attribute data of these images, such as the color, size, shape, etc. of the flames.

In addition, they also record the fire scene categories to which each image belongs, such as indoor fires, forest fires, etc.

Using the already trained fire image analysis model, researchers input first sample fire image data into the model.

The model can analyze according to the multidimensional fire attribute data of the image and judge the fire scene category to which the image belongs.

For example, when a picture containing indoor flames is input, the model analyzes the multidimensional attribute data corresponding to the image and classifies the multidimensional attribute data as indoor fires.

In the last step, the researchers perform fire scene category analysis on the first sample fire image data through the existing model.

Next, they will further train the first initial model using the second sample fire image data and the corresponding multi-dimensional fire attribute data as a training set, and the second preset fire scene element information corresponding to the second sample fire image data as a tag.

Thus, the first initial model can improve the fire scene classification capacity of the first initial model by learning the characteristics of the fire image data of the second sample and the information of the second preset fire scene elements related to the characteristics of the fire image data of the second sample.

After training, researchers use the second initial model to analyze the fire scene category of the first sample fire image data.

They input the first sample fire image data into the model and extract the corresponding multi-dimensional fire attribute data.

By comparing the multidimensional attribute data of the first sample fire image data with the features learned in the training process, the model can infer fire scene element information such as flame size, fire level and the like to which the image belongs.

Based on the first inferred fire scene element information and the first preset fire scene element information, the researcher further trains the second initial model.

They use sample data containing the first inferred fire scene element information and the first preset fire scene element information as a training set, while using the second initial model as a base model.

By training these samples, the second initial model gradually adjusts its own weights and parameters to more accurately predict and analyze the fire scene elements in the scene fire image.

This process may require multiple iterations until a preset training goal is reached, resulting in a target scene fire image analysis model.

The design is that from the acquisition of fire image data and multidimensional fire attribute data, the use of preset information to conduct fire scene category analysis and training models, the use of inferred results and preset information to conduct further training on the models, a target model is finally obtained, the whole process fully utilizes deep learning technology and related fire attribute data, and accuracy and reliability of the scene fire image analysis model are improved.

In an embodiment of the present invention, the second initial model includes: the system comprises a fire semantic feature extraction model, a fire image feature extraction model, a feature integration model and an element type identification model; the first multi-dimensional fire attribute data includes first sample fire image description data and first sample fire frame data; the aforementioned step S202 may be implemented by performing the following manner.

(1) Loading the first sample fire image description data to the fire semantic feature extraction model to extract fire image description semantic features and obtain fire image description semantic features;

(2) Loading the first sample fire frame data to the fire image feature extraction model to extract fire frame data features and obtain fire frame data features;

(3) Loading the fire image description semantic features and the fire frame data features into the feature integration model to execute feature integration operation, and obtaining first fire image integration features;

(4) And loading the first fire image integration characteristic to the element type recognition model to execute element type recognition operation, and acquiring the first inferred fire scene element information.

In the embodiment of the invention, the first sample fire image description data is loaded to the fire semantic feature extraction model for fire image description semantic feature extraction to obtain fire image description semantic features.

For example, the first sample fire image description data is input into a fire semantic feature extraction model, which is trained to understand and extract semantic information in the fire image.

For example, by inputting image description data of a descriptive sentence "fire inside a building", the fire semantic feature extraction model may extract feature vectors related to the semantics.

And loading the first sample fire frame data to a fire image feature extraction model to extract fire frame data features and obtain fire frame data features.

For example, the first sample fire frame data is input into a fire image feature extraction model that is trained to extract low-level features and visual information of the fire image.

For example, a video frame of flame combustion is input, and the fire image feature extraction model can extract features such as color, texture, shape and the like of the flame.

And loading the fire image description semantic features and the fire frame data features into the feature integration model to execute feature integration operation, and obtaining a first fire image integration feature.

For example, the fire image description semantic features and the fire frame data features are input into a feature integration model that can integrate semantic information and visual information by integrating two different types of features.

For example, a neural network model is used for fusing and weighting the fire image description semantic features and the fire frame data features, and fused first fire image integration features are generated.

And loading the first fire image integration characteristic to the element type recognition model to execute element type recognition operation, and obtaining first inferred fire scene element information.

For example, the first fire image integration feature is input into an element type recognition model that is trained to determine the element type of the fire scene based on the feature.

For example, the element type recognition model may output the element type to which the image belongs, such as a combustion object, smoke, person, etc., by inputting the integrated first fire image integration feature.

By combining the first multidimensional fire attribute data and different modules in the second initial model, the method can extract semantic information and low-level features from the fire image, integrate and fuse the semantic information and the low-level features, and finally infer element information of a fire scene.

The method can improve the accuracy and reliability of fire scene analysis.

In the embodiment of the invention, the first sample fire image description data comprises a first sample fire image type and a first sample fire identification description corresponding to the first sample fire frame data; the fire image description semantic features comprise fire image type features and fire image identification description features; the step of loading the first sample fire image description data into the fire semantic feature extraction model to extract the fire image description semantic features and obtain the fire image description semantic features can be implemented by the following example implementation.

(1) Loading the first sample fire image type to the fire semantic feature extraction model to extract fire image type features and obtain the fire image type features;

(2) And loading the first sample fire identification description to the fire semantic feature extraction model to extract fire image identification description features and obtain the fire image identification description features.

In the embodiment of the invention, the first sample fire image type is input into a fire semantic feature extraction model, and the model is trained to extract the features related to the fire image type.

For example, entering "building fire" as a fire image type, the fire semantic feature extraction model may extract vectors representing building fire features.

And inputting the fire identification description corresponding to the fire frame data of the first sample into a fire semantic feature extraction model, wherein the model can be trained to extract the features related to the fire image identification description.

For example, by inputting "dense smoke roll, bright fire" as the first sample fire identification description, the fire semantic feature extraction model may extract vectors representing fire image features.

By loading the first sample fire image description data into the fire semantic feature extraction model, the method can extract fire image type features and fire image identification description features from the fire images.

These features facilitate more accurate analysis and identification of the fire scene.

In the embodiment of the invention, the first preset fire scene element information corresponding to the first sample fire image data is obtained through the following steps.

(1) Acquiring the second sample fire image type and the second multidimensional fire attribute data;

(2) Loading the second sample fire image type to the second scene fire image analysis model, and performing scene of fire type analysis on the second sample fire image data to obtain second preset scene of fire element information corresponding to the second sample fire image data;

(3) Based on the second multidimensional fire attribute data and the second preset fire scene element information, performing fire scene category analysis training on the first initial model to obtain the first scene fire image analysis model;

(4) And loading the first multidimensional fire attribute data to the first scene fire image analysis model, and carrying out fire scene category analysis on the first sample fire image data to obtain the first preset fire scene element information.

In an embodiment of the present invention, a second sample fire image is obtained from a fire database, and multi-dimensional fire attribute data associated with the fire image, including flame size, smoke density, etc., is obtained.

The second sample fire image type is input into a second scene fire image analysis model that is trained to identify different types of scene elements.

For example, if the second sample fire image type is "building fire," the second scene fire image analysis model may infer that the image contains elements of a building, a flame, etc.

And training the first initial model by using the second multidimensional fire attribute data and the second preset fire scene element information as training data.

In the training process, the model learns how to analyze the category of the fire scene according to the multidimensional fire attribute and the preset element information.

And after training is completed, obtaining a first scene fire image analysis model.

The first multi-dimensional fire attribute data is input into a first scene fire image analysis model which is trained to perform category analysis on the fire scene according to the multi-dimensional fire attribute data.

For example, if the first multi-dimensional fire attribute data includes information of flame size, smoke density, etc., the first scene fire image analysis model may infer preset scene element information of the image, such as larger flames, higher smoke density, etc.

By the design, the second preset fire scene element information can be obtained by acquiring the second sample fire image type and the second multidimensional fire attribute data and then applying the second sample fire image type and the second multidimensional fire attribute data to the second scene fire image analysis model.

Training based on the second preset information and the first multi-dimensional fire attribute data to obtain a first scene fire image analysis model, and finally inputting the first multi-dimensional fire attribute data into the model to obtain first preset fire scene element information.

The method can improve the accuracy and reliability of fire scene analysis.

In the embodiment of the invention, the second multidimensional fire attribute data comprises the second sample fire image type, second sample fire frame data and second sample fire identification description corresponding to the second sample fire frame data; the step of performing fire scene category analysis training on the first initial model based on the second multidimensional fire attribute data and the second preset fire scene element information to obtain the first scene fire image analysis model may be implemented by the following example.

(1) Acquiring a first preset weight parameter corresponding to the second sample fire image type, a second preset weight parameter corresponding to the second sample fire identification description and a third preset weight parameter corresponding to the second sample fire frame data;

(2) Based on the first preset weight parameter, the second preset weight parameter and the third preset weight parameter, performing weight synthesis operation on the second sample fire image type, the second sample fire identification description and the second sample fire frame data to obtain a second fire image integration feature;

(3) Loading the second fire image integration feature to the first initial model for fire scene category analysis, and obtaining second inferred fire scene element information corresponding to the second sample fire image data;

(4) Training the first initial model according to the second inferred fire scene element information and the second preset fire scene element information to obtain the first scene fire image analysis model.

In the embodiment of the invention, for the second sample fire image type, the first preset weight parameter corresponding to the second sample fire image type is obtained according to a preset rule or statistical analysis.

Similarly, according to the second sample fire identification description and the second sample fire frame data, second and third preset weight parameters corresponding to the second sample fire identification description and the second sample fire frame data are acquired.

And weighting the second sample fire image type, the second sample fire identification description and the second sample fire frame data according to the first preset weight parameter, the second preset weight parameter and the third preset weight parameter.

For example, assuming that the first preset weight parameter is 0.4, the second preset weight parameter is 0.3, and the third preset weight parameter is 0.3, the corresponding fire image type, identification description and fire frame data are respectively "building fire", "dense smoke roll" and flame size 20 square meters, the second fire image integration feature can be obtained through the weight synthesis operation.

The second fire image integration feature is input into the first initial model, and the model is trained to perform category analysis on the fire scene according to the input feature.

For example, by inputting the second fire image integration feature, fire scene element information, such as flame size, smoke density, etc., corresponding to the second sample fire image can be deduced.

And training the first initial model by using the second inferred fire scene element information and the second preset fire scene element information as training data.

During the training process, the model learns how to analyze the category of the fire scene according to the inferred and preset fire scene element information.

After training is completed, a first scene fire image analysis model is obtained.

In the embodiment of the present invention, after step S203, the following example is also provided.

(1) Receiving current fire image data, and loading the current fire image data to the target scene fire image analysis model to obtain a current scene fire image analysis result corresponding to the current fire image data;

(2) Inputting the current scene fire image analysis result into a pre-trained fire control strategy output model to obtain a current fire control strategy output result corresponding to the current scene fire image analysis result, and sending the fire control strategy output result to a preset dispatching center.

In an embodiment of the present invention, a camera or other fire monitoring device is used to obtain current fire image data, for example, when a fire occurs.

These data are input into a target scene fire image analysis model which is trained to analyze from the input fire image data.

For example, by loading the current fire image data into the model, the analysis result of the current scene fire image, such as information of flame size, combustion area, etc., can be obtained.

The current scene fire image analysis result is input into a pre-trained fire strategy output model, and the model can judge which fire strategy is adopted according to the characteristics and situation of fire conditions, so that the fire strategy is worth to be explained, wherein the current scene fire image analysis result is determined by the current scene fire element information (namely, the current scene fire image analysis result is composed of the current scene fire element information), and in the scene fire analysis, the scene fire element information refers to various elements and factors related to fire, such as flame, smoke, temperature, illumination and the like.

Such elemental information may be obtained from the scene of a fire by image acquisition, typically using image processing techniques for analysis and extraction.

The fire scene element information is important for fire image analysis.

By analyzing the characteristics of the flame such as size, color and shape, the severity of the fire and the type of burning substance can be determined.

The density, colour and spread of smoke can help determine the extent of fire spread and the effect of smoke on the human body.

The change in temperature can be used to identify the source of the fire and predict the trend of the fire.

The lighting conditions may affect the quality of the image and the observation of the fire scene.

Based on the current fire scene element information, the fire image analysis can be performed.

By analyzing and extracting the fire scene element information, important data and conclusions about fire properties, fire spread, smoke spread, personnel safety and the like can be obtained.

For example, if the current scene fire image analysis results show that the flame area is large and spreading rapidly, the fire strategy output model may suggest immediate dispatch of more fire fighters and increased use of fire extinguishing equipment.

And sending the output result of the current fire control strategy to a preset dispatching center, wherein the center can be a fire control command center or other relevant institutions.

The preset dispatching center can timely know the current fire situation and take corresponding measures by sending the fire control strategy output result.

For example, after receiving the fire policy output, the dispatch center may notify the firefighter to go to the scene, contact other rescue authorities, and so on.

The design is that the analysis result corresponding to the current fire image data is obtained by receiving the current fire image data and loading the current fire image data into the target site fire image analysis model, the current fire strategy output result is obtained through the pre-trained fire strategy output model, and the current fire strategy output result is sent to the preset dispatching center, so that fire strategy decision and dispatching management based on the fire image are realized.

In an embodiment of the present invention, the aforementioned fire strategy output model is obtained in the following manner.

(1) Acquiring a training pool for training the fire control strategy output model, wherein the training pool comprises at least one training data set, and the training data set comprises a database record, a scene fire image analysis result, a scene fire image analysis command and a fire control strategy output result which are mutually related; the fire control strategy output result is an answer corresponding to the scene fire image analysis result obtained by searching from the database record by adopting the scene fire image analysis command;

(2) Generating a training example corresponding to the fire strategy emergency response according to the training data set; the training examples corresponding to the fire control strategy emergency response take the database records and the on-site fire image analysis results as training input data, and target data are determined based on the fire control strategy output results;

(3) Generating a training example corresponding to at least one association rule learning of the fire control strategy emergency response according to the training data set, wherein the association rule learning is a learning task for training the fire control strategy output model in combination with the fire control strategy emergency response;

(4) And training the fire control strategy output model by adopting the training examples corresponding to the fire control strategy emergency response and the training examples corresponding to the at least one association rule learning, so as to obtain the fire control strategy output model after training.

In an exemplary embodiment of the present invention, the training pool may be a database comprising a plurality of data sets, each data set comprising information related to a fire.

For example, the database records may contain information about building structure, fire type, fire size, etc.; the scene fire image analysis result is obtained by analyzing the fire image; the scene fire image analysis command represents a database script in the form of a standard query statement; the fire strategy output results are answers obtained by searching and matching the scene fire image analysis results from the database records.

Based on database record in training data set and analysis result of fire image in scene, they are used as input data of training example.

For example, one training example may include database records and corresponding field fire image analysis results for training a model to generate a corresponding fire strategy emergency response.

And generating at least one training instance corresponding to association rule learning by using the related information in the training data set.

Association rule learning is a data mining technique used to discover associations between different attributes.

Here, the association rule learning and fire strategy emergency response jointly train the fire strategy output model to improve accuracy and reliability of the model.

Training the fire control strategy output model by using training examples corresponding to fire control strategy emergency response and training examples corresponding to at least one association rule learning as training data.

Through the training process, the model learns how to generate corresponding fire control strategy emergency response according to the input database records, the analysis results of the scene fire images, the association rules and other information.

After training is completed, a fire control strategy output model with the training completed is obtained.

In the embodiment of the present invention, the association rule learning includes query conversion rule learning, and the step of generating the training instance corresponding to at least one association rule learning of the fire strategy emergency response according to the training data set may be implemented by the following example execution.

(1) Based on a standard query statement contained in the scene fire image analysis command, carrying out query conversion on the scene fire image analysis result to obtain a first query conversion expression expressed by the standard query statement;

(2) Generating a training example corresponding to the query conversion rule learning according to the database record, the on-site fire image analysis result and the first query conversion expression; the training examples corresponding to the query conversion rule learning take the database records and the on-site fire image analysis results as training input data, and the first query conversion expression as target data.

In the embodiment of the invention, it is assumed that the scene fire image analysis command includes a standard query statement for acquiring corresponding fire information from the database.

By executing this query statement, the analysis result of the scene fire image can be converted into a form expressed by a standard query statement.

For example, if the query statement is "SELECT FROM fire information WHERE floor= '3'", then the first query conversion expression is "floor= '3'".

And generating a training example corresponding to query conversion rule learning according to the database records, the on-site fire image analysis result and the first query conversion expression.

For example, the training examples may include a database record and scene fire image analysis results associated with the record.

The goal of the training model is to convert these input data into a first query conversion expression.

And carrying out query conversion on the analysis result of the scene fire image by using a standard query statement, and generating a training example corresponding to query conversion rule learning by combining the database record and the analysis result of the scene fire image.

Therefore, the accuracy and the adaptability of the fire strategy output model to fire response under different conditions can be improved.

In the embodiment of the invention, the following implementation manner is also provided.

(1) Generating a second query conversion expression corresponding to the scene fire image analysis result according to the database record and the scene fire image analysis result through the fire strategy output model;

(2) And calculating a first price function parameter according to the second query conversion expression and the first query conversion expression, wherein the first price function parameter is used for evaluating the model precision of the fire control strategy output model on the query conversion rule learning.

In the embodiment of the invention, the trained fire control strategy output model is used for inputting database records and on-site fire image analysis results, and a second query conversion expression corresponding to the on-site fire image analysis results is generated through model prediction.

For example, from a given database record and scene fire image analysis results, the model may predict that the second query transformation corresponding to the scene fire image analysis results is expressed as "floor= '3' and fire= 'big'".

And calculating a first cost function parameter as an evaluation index of model accuracy by using the second query conversion expression and the first query conversion expression.

For example, based on the differences between the second query conversion expression and the first query conversion expression, a first cost function parameter may be calculated for measuring accuracy of the fire strategy output model in query conversion rule learning.

A lower first cost function value indicates a higher accuracy of the model.

By adopting the design, the fire control strategy output model is used for generating the second query conversion expression corresponding to the analysis result of the scene fire image, and calculating the first price function parameter according to the second query conversion expression and the first query conversion expression, so that the model precision of the fire control strategy output model on the query conversion rule learning task can be estimated.

This can help optimize the performance and accuracy of the fire strategy output model.

In the embodiment of the invention, the association rule learning comprises target database record generation learning; the foregoing step of generating at least one association rule learning corresponding training instance of the fire strategy emergency response from the training data set may be implemented by the following example execution.

(1) Generating a query converted on-site fire image analysis command according to the on-site fire image analysis result and the on-site fire image analysis command, wherein the query converted on-site fire image analysis command is used for searching complete field information of at least one attribute item associated with the on-site fire image analysis result in the database record;

(2) Extracting complete field information of at least one attribute item associated with the analysis result of the scene fire image from the database record according to the query converted scene fire image analysis command to obtain a first target database record;

(3) Generating a target database record according to the database record, the on-site fire image analysis result and the first target database record to generate a training example corresponding to learning; the target database record generates training examples corresponding to learning, the database record and the analysis result of the scene fire image are used as training input data, and the first target database record is used as target data.

In the embodiment of the present invention, it is exemplarily assumed that the scene fire image analysis command is "SELECT floor, fire FROM fire information WHERE time= '2022-01-01'".

This command may be converted to "SELECT complete field information FROM fire information WHERE time= '2022-01-01'" by query conversion, WHERE complete field information refers to all field contents of at least one attribute item associated with the analysis result of the scene fire image.

And executing query operation in the database by using the query converted scene fire image analysis command, and extracting the complete field information of at least one attribute item associated with the scene fire image analysis result.

For example, if the query converted command is "SELECT complete field information FROM fire information WHERE time= '2022-01-01'", a record conforming to the condition is obtained FROM the database, and the complete field information of the attribute item associated with the analysis result of the scene fire image in these records is extracted.

And generating a target database record to generate a training example corresponding to learning according to the database record, the on-site fire image analysis result and the first target database record.

For example, the training examples may include a database record and the results of the analysis of the scene fire image and the first target database record associated with the record.

The goal of the training model is to convert these input data into a first target database record.

The method comprises the steps of extracting complete field information of at least one attribute item associated with a scene fire image analysis result from a database record by using a scene fire image analysis command after query conversion, and generating a training example corresponding to learning by combining the database record, the scene fire image analysis result and a first target database record.

This can help optimize the performance and accuracy of the fire strategy output model on the target database record generation task.

In the embodiment of the present invention, the following examples are also provided.

(1) Generating a second target database record corresponding to the on-site fire image analysis result according to the database record and the on-site fire image analysis result through the fire control strategy output model;

(2) And calculating a second cost function parameter according to the second target database record and the first target database record, wherein the second cost function parameter is used for evaluating the model accuracy of the fire control strategy output model in the target database record generation learning.

In the embodiment of the invention, the trained fire control strategy output model is used for inputting the database record and the on-site fire image analysis result, and the second target database record corresponding to the on-site fire image analysis result is generated through model prediction.

For example, from a given database record and scene fire image analysis results, the model may predict that the second target database record corresponding to the scene fire image analysis results is "floor= '3' and fire= 'big'".

And calculating a second cost function parameter as an evaluation index of the model precision by using the second target database record and the first target database record.

For example, based on the variability of the second target database record and the first target database record, a second cost function parameter may be calculated for measuring the accuracy of the fire strategy output model in generating the target database record.

A lower second cost function value indicates a higher accuracy of the model.

By adopting the design, the fire control strategy output model is used for generating the second target database record corresponding to the analysis result of the scene fire image, and the second cost function parameter is calculated according to the second target database record and the first target database record, so that the model precision of the fire control strategy output model on the target database record generation learning task can be evaluated.

This can help optimize the performance and accuracy of the fire strategy output model.

In the embodiment of the invention, the association rule learning comprises decision process learning; the foregoing step of generating at least one association rule learning corresponding training instance of the fire strategy emergency response from the training data set may be implemented by the following example execution.

(1) Based on a standard query statement contained in the scene fire image analysis command, carrying out query conversion on the scene fire image analysis result to obtain a first query conversion expression expressed by the standard query statement;

(2) Generating a query converted on-site fire image analysis command according to the on-site fire image analysis result and the on-site fire image analysis command, wherein the query converted on-site fire image analysis command is used for searching complete field information of at least one attribute item associated with the on-site fire image analysis result in the database record;

(3) Extracting complete field information of at least one attribute item associated with the on-site fire image analysis result or the first query conversion expression from the database record according to the query converted on-site fire image analysis command to obtain a first target database record;

(4) Generating a training example corresponding to the decision process learning according to the database record, the on-site fire image analysis result, the fire control strategy output result, the first query conversion expression and the first target database record; the training examples corresponding to the decision process learning take the database records and the on-site fire image analysis results as training input data, and take the first query conversion expression, the first target database records and the fire control strategy output results as target data.

In the embodiment of the present invention, it is exemplarily assumed that the scene fire image analysis command is "SELECT floor, fire FROM fire information WHERE time= '2022-01-01'".

The analysis result of the scene fire image is converted into a first query conversion expression in the form of a standard query statement, such as 'SELECT floor, fire FROM fire information WHERE time=' 2022-01 'limit 1'.

And executing query operation in the database by using the query converted scene fire image analysis command, and extracting the complete field information of at least one attribute item associated with the scene fire image analysis result.

For example, if the query converted command is "SELECT floor, fire FROM fire information WHERE time= '2022-01' limit1", the eligible records are obtained FROM the database, and the complete field information of the attribute item associated with the analysis result of the scene fire image in these records is extracted.

Based on the given database record, the scene fire image analysis result, the fire strategy output result, the first query conversion expression and the first target database record, a training example corresponding to decision process learning is generated.

The training examples may include a database record and scene fire image analysis results, fire strategy output results, a first query conversion expression, and a first target database record associated with the record.

The goal of the training model is to generate a decision process learning result corresponding to the first target database record from the input data.

By means of the design, the on-site fire image analysis command is inquired and converted, complete field information of attribute items associated with on-site fire image analysis results in the database records is extracted, and training examples corresponding to decision process learning are generated by combining fire strategy output results, first inquiry conversion expression and first target database records.

Therefore, the performance and the accuracy of the fire strategy output model on the target database record generation learning task can be helped to be optimized, and the decision making process of the model is further improved.

In the embodiment of the invention, the following implementation manner is also provided.

(1) Generating a second query conversion expression, a second target database record and a predicted fire strategy corresponding to the on-site fire image analysis result according to the database record and the on-site fire image analysis result through the fire strategy output model;

(2) Calculating a third price function parameter according to the deviation between the second query conversion expression and the first query conversion expression, the deviation between the second target database record and the first target database record and the deviation between the predicted fire strategy and the fire strategy output result, wherein the third price function parameter is used for evaluating model precision of the fire strategy output model in the decision process learning.

In the embodiment of the invention, a fire control strategy output model is used for predicting a second query conversion expression, a second target database record and a predicted fire control strategy corresponding to the on-site fire image analysis result according to the given database record and the on-site fire image analysis result by way of example.

For example, based on the trained model, the input database records and the scene fire image analysis results, the model may predict that the second query transformation is expressed as "SELECT floor, the fire FROM fire information WHERE time= '2022-01' limit5", the second target database records as "floor= '3' and fire= 'big'", and the fire strategy is predicted as "evacuation floor 3 and fire extinguisher is maneuvered to put out.

And calculating a third price function parameter as an evaluation index of model accuracy according to the difference between the second query conversion expression and the first query conversion expression, the difference between the second target database record and the first target database record and the difference between the predicted fire strategy and the fire strategy output result.

For example, by calculating the absolute value of these differences or other suitable metrics, third generation tariff parameters may be derived for measuring the accuracy of the fire strategy output model in the decision process learning task.

A lower third cost function value indicates a higher accuracy of the model.

And generating a second query conversion expression, a second target database record and a predicted fire strategy corresponding to the analysis result of the scene fire image through the fire strategy output model, and calculating a third price function parameter according to the difference between the first query conversion expression, the first target database record and the fire strategy output result, so that the model precision of the fire strategy output model on the decision process learning task can be estimated.

Therefore, the performance and the accuracy of the fire strategy output model can be optimized, and the performance of the model in the decision making process can be further improved.

In the embodiment of the present invention, the step of generating the training examples corresponding to the fire strategy emergency response according to the training data set may be implemented by the following example execution.

(1) Query conversion is carried out on the fire control strategy output result to obtain a fire control strategy output result after query conversion, and the fire control strategy output result after query conversion expresses the fire control strategy output result by adopting standard query sentences recorded by the database;

(2) Generating a training example corresponding to the fire control strategy emergency response according to the database record, the on-site fire image analysis result and the query converted fire control strategy output result; the training examples corresponding to the fire control strategy emergency response take the fire control strategy output result after query conversion as target data.

In the present embodiment, it is assumed, by way of example, that the fire strategy output is "evacuate floor 3 and maneuver a large fire extinguisher to extinguish".

Through query conversion, the fire control strategy output result is converted into a standard query statement form recorded by a database to be expressed, for example, a ' SELECT FROM fire control strategy WHERE strategy= ' evacuation floor 3 and large fire extinguishers are mobilized to extinguish '.

Based on the given database record, the on-site fire image analysis result and the fire control strategy output result after query conversion, a training example corresponding to the fire control strategy emergency response is generated.

The training examples may include a database record, a scene fire image analysis result associated with the record, and a fire strategy output result after query conversion as target data.

For example, the training examples may include a database record, fire information obtained by analysis of scene fire images, and query the converted fire strategy output result "SELECT FROM fire strategy WHERE strategy= 'evacuate floor 3 and fire extinguishing'" by moving a large fire extinguisher.

By means of the design, the fire control strategy output result is inquired and converted, and the training examples corresponding to the fire control strategy emergency response are generated by combining the database records and the scene fire image analysis results.

The training examples take the query converted fire control strategy output result as target data, are beneficial to optimizing the performance and accuracy of the fire control strategy emergency response model on training tasks, and further improve the effect of the model in practical application.

In the embodiment of the present invention, the step of training the fire control strategy output model by using the training instance corresponding to the fire control strategy emergency response and the training instance corresponding to the learning of the at least one association rule to obtain the trained fire control strategy output model may be implemented by the following example.

(1) According to the cost function parameters corresponding to the fire strategy emergency response and the at least one association rule, learning the corresponding cost function parameters, and calculating to obtain comprehensive cost function parameters;

(2) And executing optimization operation on parameters of the fire control strategy output model according to the comprehensive cost function parameters to obtain the trained fire control strategy output model.

The cost function parameter obtained by model evaluation of the training example corresponding to the emergency response of the fire strategy is assumed to be 0.6, and the cost function parameter obtained by model evaluation of the training example corresponding to the association rule learning is assumed to be 0.3.

The composite cost function parameters are obtained by calculating a weighted average of the two cost function parameters or other suitable means.

For example, the composite cost function parameter may be calculated as (0.6×0.8) + (0.3×0.2) =0.54.

And (3) carrying out optimization operation on parameters of the output model of the fire protection strategy based on the calculated comprehensive cost function parameters, for example, using gradient descent and other methods.

The model parameters are adjusted so that the overall cost function parameters are minimized or close to the minimized values.

And finally obtaining the fire control strategy output model after finishing training through repeated iteration and optimization operation, wherein the model has better performance and accuracy under the condition of considering fire control strategy emergency response and association rule learning.

According to the design, corresponding training examples are learned according to the fire control strategy emergency response and the association rule, comprehensive cost function parameters are calculated, and the parameters are used for carrying out parameter optimization on the fire control strategy output model, so that the fire control strategy output model with training completed is obtained.

The training process can help to improve the performance and accuracy of the model, so that the model is more suitable for actual fire scenes and can better cope with various emergency situations.

The embodiment of the invention provides a computer device 100, wherein the computer device 100 comprises a processor and a nonvolatile memory storing computer instructions, and when the computer instructions are executed by the processor, the computer device 100 executes the training method of the scene fire image analysis model based on deep learning.

As shown in fig. 2, fig. 2 is a block diagram of a computer device 100 according to an embodiment of the present invention.

The computer device 100 comprises a memory 111, a processor 112 and a communication unit 113.

For data transmission or interaction, the memory 111, the processor 112 and the communication unit 113 are electrically connected to each other directly or indirectly.

For example, the elements may be electrically connected to each other via one or more communication buses or signal lines.

The foregoing description, for purpose of explanation, has been presented with reference to particular embodiments.

The illustrative discussions above are not intended to be exhaustive or to limit the disclosure to the precise forms disclosed.

Many modifications and variations are possible in light of the above teaching.

The embodiments were chosen and described in order to best explain the principles of the disclosure and its practical application, to thereby enable others skilled in the art to best utilize the disclosure and various embodiments with various modifications as are suited to the particular use contemplated.

Claims (8)

1. The on-site fire image analysis model training method based on deep learning is characterized by comprising the following steps of:

Acquiring first multidimensional fire attribute data corresponding to first sample fire image data and first preset fire scene element information corresponding to the first sample fire image data; the first preset fire scene element information is obtained by performing fire scene category analysis on the first sample fire image data based on a first scene fire image analysis model and the first multidimensional fire attribute data corresponding to the first sample fire image data; the first scene fire image analysis model is obtained by performing scene fire type analysis training on the first initial model based on second multidimensional fire attribute data corresponding to second sample fire image data and second preset scene element information corresponding to the second sample fire image data; the second preset fire scene element information is obtained by performing fire scene category analysis on the second sample fire image data based on a second scene fire image analysis model and a second sample fire image type corresponding to the second sample fire image data;

loading the first multidimensional fire attribute data to a second initial model for fire scene category analysis, and obtaining first inferred fire scene element information corresponding to the first sample fire image data;

Training the second initial model according to the first inferred fire scene element information and the first preset fire scene element information to obtain a target scene fire image analysis model;

The second initial model includes: the system comprises a fire semantic feature extraction model, a fire image feature extraction model, a feature integration model and an element type identification model; the first multi-dimensional fire attribute data includes first sample fire image description data and first sample fire frame data; the first sample fire image description data comprises a first sample fire identification description corresponding to a first sample fire image type and the first sample fire frame data; the fire image description semantic features comprise fire image type features and fire image identification description features;

The loading the first multidimensional fire attribute data into a second initial model for fire scene category analysis, and the obtaining of the first inferred fire scene element information corresponding to the first sample fire image data comprises the following steps:

Loading the first sample fire image type to the fire semantic feature extraction model to extract fire image type features and obtain the fire image type features;

loading the first sample fire identification description to the fire semantic feature extraction model to extract fire image identification description features and obtain the fire image identification description features;

loading the first sample fire frame data to the fire image feature extraction model to extract fire frame data features and obtain fire frame data features;

loading the fire image description semantic features and the fire frame data features into the feature integration model to execute feature integration operation, and obtaining first fire image integration features;

loading the first fire image integration feature to the element type recognition model to execute element type recognition operation, and acquiring the first inferred fire scene element information;

The first preset fire scene element information corresponding to the first sample fire image data is obtained through the following steps:

acquiring the second sample fire image type and the second multidimensional fire attribute data;

loading the second sample fire image type to the second scene fire image analysis model, and performing scene of fire type analysis on the second sample fire image data to obtain second preset scene of fire element information corresponding to the second sample fire image data; the second multidimensional fire attribute data comprises the second sample fire image type, second sample fire frame data and second sample fire identification description corresponding to the second sample fire frame data;

Acquiring a first preset weight parameter corresponding to the second sample fire image type, a second preset weight parameter corresponding to the second sample fire identification description and a third preset weight parameter corresponding to the second sample fire frame data;

Based on the first preset weight parameter, the second preset weight parameter and the third preset weight parameter, performing weight synthesis operation on the second sample fire image type, the second sample fire identification description and the second sample fire frame data to obtain a second fire image integration feature;

Loading the second fire image integration feature to the first initial model for fire scene category analysis, and obtaining second inferred fire scene element information corresponding to the second sample fire image data;

training the first initial model according to the second inferred fire scene element information and the second preset fire scene element information to obtain the first scene fire image analysis model;

and loading the first multidimensional fire attribute data to the first scene fire image analysis model, and carrying out fire scene category analysis on the first sample fire image data to obtain the first preset fire scene element information.

2. The method of claim 1, wherein after the acquiring the target site fire image analysis model, the method further comprises:

receiving current fire image data, and loading the current fire image data to the target scene fire image analysis model to obtain a current scene fire image analysis result corresponding to the current fire image data;

Inputting the current scene fire image analysis result into a pre-trained fire control strategy output model to obtain a current fire control strategy output result corresponding to the current scene fire image analysis result, and sending the fire control strategy output result to a preset dispatching center.

3. The method of claim 2, wherein the fire strategy output model is derived by:

Acquiring a training pool for training the fire control strategy output model, wherein the training pool comprises at least one training data set, and the training data set comprises a database record, a scene fire image analysis result, a scene fire image analysis command and a fire control strategy output result which are mutually related; the fire control strategy output result is an answer corresponding to the scene fire image analysis result obtained by searching from the database record by adopting the scene fire image analysis command;

Query conversion is carried out on the fire control strategy output result to obtain a fire control strategy output result after query conversion, and the fire control strategy output result after query conversion expresses the fire control strategy output result by adopting standard query sentences recorded by the database;

Generating a training example corresponding to the fire control strategy emergency response according to the database record, the on-site fire image analysis result and the query converted fire control strategy output result; the training examples corresponding to the fire control strategy emergency response take the fire control strategy output result after query conversion as target data; the training examples corresponding to the fire control strategy emergency response take the database records and the on-site fire image analysis results as training input data, and target data are determined based on the fire control strategy output results;

Generating a training example corresponding to at least one association rule learning of the fire control strategy emergency response according to the training data set, wherein the association rule learning is a learning task for training the fire control strategy output model in combination with the fire control strategy emergency response;

according to the cost function parameters corresponding to the fire strategy emergency response and the at least one association rule, learning the corresponding cost function parameters, and calculating to obtain comprehensive cost function parameters;

and executing optimization operation on the parameters of the fire control strategy output model according to the comprehensive cost function parameters to obtain the trained fire control strategy output model.

4. A method according to claim 3, wherein the association rule learning comprises query conversion rule learning;

generating a training example corresponding to at least one association rule learning of the fire strategy emergency response according to the training data set, wherein the training example comprises the following steps:

Based on a standard query statement contained in the scene fire image analysis command, carrying out query conversion on the scene fire image analysis result to obtain a first query conversion expression expressed by the standard query statement;

Generating a training example corresponding to the query conversion rule learning according to the database record, the on-site fire image analysis result and the first query conversion expression; the training examples corresponding to the query conversion rule learning take the database records and the on-site fire image analysis results as training input data, and the first query conversion expression as target data.

5. The method according to claim 4, wherein the method further comprises:

Generating a second query conversion expression corresponding to the scene fire image analysis result according to the database record and the scene fire image analysis result through the fire strategy output model;

and calculating a first price function parameter according to the second query conversion expression and the first query conversion expression, wherein the first price function parameter is used for evaluating the model precision of the fire control strategy output model on the query conversion rule learning.

6. A method according to claim 3, wherein the association rule learning comprises target database record generation learning;

generating a training example corresponding to at least one association rule learning of the fire strategy emergency response according to the training data set, wherein the training example comprises the following steps:

generating a query converted on-site fire image analysis command according to the on-site fire image analysis result and the on-site fire image analysis command, wherein the query converted on-site fire image analysis command is used for searching complete field information of at least one attribute item associated with the on-site fire image analysis result in the database record;

extracting complete field information of at least one attribute item associated with the analysis result of the scene fire image from the database record according to the query converted scene fire image analysis command to obtain a first target database record;

Generating a target database record according to the database record, the on-site fire image analysis result and the first target database record to generate a training example corresponding to learning; the target database record generates training examples corresponding to learning, the database record and the analysis result of the scene fire image are used as training input data, and the first target database record is used as target data;

The method further comprises the steps of:

Generating a second target database record corresponding to the on-site fire image analysis result according to the database record and the on-site fire image analysis result through the fire control strategy output model;

and calculating a second cost function parameter according to the second target database record and the first target database record, wherein the second cost function parameter is used for evaluating the model accuracy of the fire control strategy output model in the target database record generation learning.

7. A method according to claim 3, wherein the association rule learning comprises decision process learning;

generating a training example corresponding to at least one association rule learning of the fire strategy emergency response according to the training data set, wherein the training example comprises the following steps:

Based on a standard query statement contained in the scene fire image analysis command, carrying out query conversion on the scene fire image analysis result to obtain a first query conversion expression expressed by the standard query statement;

generating a query converted on-site fire image analysis command according to the on-site fire image analysis result and the on-site fire image analysis command, wherein the query converted on-site fire image analysis command is used for searching complete field information of at least one attribute item associated with the on-site fire image analysis result in the database record;

Extracting complete field information of at least one attribute item associated with the on-site fire image analysis result or the first query conversion expression from the database record according to the query converted on-site fire image analysis command to obtain a first target database record;

Generating a training example corresponding to the decision process learning according to the database record, the on-site fire image analysis result, the fire control strategy output result, the first query conversion expression and the first target database record; the training examples corresponding to the decision process learning take the database records and the on-site fire image analysis results as training input data, and take the first query conversion expression, the first target database records and the fire control strategy output results as target data;

The method further comprises the steps of:

Generating a second query conversion expression, a second target database record and a predicted fire strategy corresponding to the on-site fire image analysis result according to the database record and the on-site fire image analysis result through the fire strategy output model;

Calculating a third price function parameter according to the deviation between the second query conversion expression and the first query conversion expression, the deviation between the second target database record and the first target database record and the deviation between the predicted fire strategy and the fire strategy output result, wherein the third price function parameter is used for evaluating model precision of the fire strategy output model in the decision process learning.

8. A server system comprising a server for performing the method of any of claims 1-7.