WO2019051701A1

WO2019051701A1 - Photographic terminal, and photographic parameter setting method therefor based on long short-term memory neural network

Info

Publication number: WO2019051701A1
Application number: PCT/CN2017/101687
Authority: WO
Inventors: 雷文
Original assignee: 深圳传音通讯有限公司
Priority date: 2017-09-14
Filing date: 2017-09-14
Publication date: 2019-03-21

Abstract

The present application relates to a photographic terminal, and a photographic parameter setting method therefor based on a long short-term memory neural network. The method comprises: when starting a photography function, a photographic terminal acquiring a predicted photographic parameter setting value from a long short-term memory (LSTM) neural network; setting a photographic parameter of the photographic terminal according to the acquired photographic parameter setting value; and using the photographic parameter for photography. The present application can use an LSTM to predict a photographic parameter setting value desired by a user, automatically sets a photographic parameter according to this, can achieve a better photographic effect, and better complies with the desired effect of the user. The present application does not require a user to manually set a photographic parameter, thereby reducing the difficulty of use for the user, and does not require a user to carry out complicated operation settings, thereby improving the user experience.

Description

Photographing terminal and photographing parameter setting method based on long-term and short-term memory neural network

Technical field

The present invention relates to the field of photographing technology, and in particular to a photographing terminal and a photographing parameter setting method based on LSTM (Long Short Term Memory).

Background technique

With the development of electronic products and information storage technology, digital cameras/cameras have been widely used by ordinary users because of their practicality, convenience, low price and friendly operation interface. In order to enable the majority of users, especially non-professional photographers to shoot excellent camera works, each camera business also promotes the camera, but also introduces their own built-in parameter automatic setting software, such as auto focus, flash off light automatically closed and so on. These built-in features provide users with a more convenient interface.

However, for non-professional ordinary users, due to technical, experience and other reasons, and do not know how to better set the camera parameters, this will undoubtedly increase the difficulty of the user. In addition, for professional users, if different objects are photographed in different shooting environments, it is necessary to repeatedly set various parameters, which brings inconvenience to the user to some extent; and for some scenes with similar shooting factors, The user may not be able to set a similar parameter, resulting in a poor shooting result.

technical problem

It is not difficult to understand that because there are too many factors affecting the shooting, and there are many various camera parameters that need to be set, for non-professional users, the parameters cannot be set and the better shooting results are not obtained. For professional users, on the one hand, repeated manual settings will bring complicated operation procedures to users, on the other hand, for similar scenarios, it is difficult to set them to better parameters again, or they cannot be set to better parameters in time. It led to missed some shooting opportunities. All in all, the setting method of the camera parameters of the prior art has not kept pace with the development of the technology, and cannot meet the diversity requirements of the user, and the user experience is poor.

Technical solution

In view of the above technical problems, the present application provides a camera terminal and a camera parameter setting method based on a long-term and short-term memory neural network to meet the diversity requirements of users and improve the user experience.

In order to solve the above technical problem, the present application provides a photographing parameter based on a long-term and short-term memory neural network. a setting method, wherein the photographing parameter setting method based on the long-term and short-term memory neural network comprises:

When the photographing terminal starts the photographing function, the photographed parameter setting value is obtained from the long-term and short-term memory neural network LSTM;

Setting a photographing parameter of the photographing terminal according to the obtained photographing parameter setting value;

The photographing parameters are used for photographing.

The LSTM predicts the setting value of the camera parameter of the camera terminal, and specifically includes:

Obtain historical data of camera parameters;

The historical data is cleaned and normalized;

The historical data after normalization of the cleaning is divided into a training data set and a test data set according to time;

Performing off-line model training on the training data of the training data set to separately train multiple neural network models of the LSTM;

Obtaining a list of predicted values output by the training data for the plurality of neural network models after the training, comparing the predicted value list with an actual photographing setting value, and calculating a weight value occupied by the plurality of neural network models as the combined model;

The prediction data is evaluated by using the test data of the test data set for the plurality of neural network models in the combined model, and the weight values occupied by the plurality of neural network models as the combined model are adjusted according to the prediction effect.

Wherein, after adjusting the weight value of the plurality of neural network models as the combined model according to the prediction effect, the method further includes:

The camera parameter setting value is predicted by using a scroll time window.

The method for predicting the setting value of the photographing parameter by using the rolling time window includes:

Converting the added and subtracted values predicted by the combined model into predicted values of the predicted time, and then filling the currently predicted predicted values into the time window of the next predicted time, and alternately cycling according to this;

When the actual value actually set by the photographing parameter is obtained, the predicted value is compared with the actual value, and the actual value is used as a new set of training data according to the comparison result, and substituted into the model to update the model parameter.

The obtaining historical data of the photographing parameter includes: acquiring contrast, sensory degree, aperture, shutter, ISO, focus, metering, and white balance to comprehensively obtain historical data of the photographing parameter.

The historical data obtained by the comprehensively obtaining the photographing parameters specifically includes:

According to the data distribution characteristics of the photographing parameters, using the accept-reject sampling method, similar scenes, similar objects, similar geographical locations and related data of the same photographer are selected, and the data of the photographing parameters are combined with the data of the photographing parameters to form the original place. State data.

The historical data after the normalization of the cleaning is divided into a training data set and a test data set according to time, including:

The early data in the historical data before the specified time is divided into training data sets, and the late data in the historical data after the specified time is divided into test data sets.

Before performing the offline model training on the training data of the training data set, the method further includes: using the training data of the training data set to generate time series data of different spans; wherein, for each time span (t ₀ , t ₁ , t ₂ , , , t _n ), using (t ₀ , t ₁ , t ₂ , , , t _n-1 ) as input values, using the difference value between t _n-1 and t _n to discretize them After being converted, it is converted into the unique heat code data as the supervisory value;

Performing offline model training on the training data of the training data set, where the correspondence includes:

Each neural network model of the LSTM is trained separately for each time series data in time series data of different spans.

The performing offline model training on the training data of the training data set includes:

Training data of the training data set is trained by using a memory-based distributed training method, wherein the training data is distributed to each node and the initial model parameters of the neural network model are broadcast to each node, and each node is based on the current The model parameters and the training data of a certain scale obtain the current gradient and the update amount of the model parameters, update the model parameters by summarizing the model parameter update amount fed back by each node, and broadcast the updated model parameters to each node, and then iteratively repeat To complete the training of a single LSTM neural network model as required.

Wherein, the calculating the weight value of the plurality of neural network models as the combined model comprises: using the training data of multiple time periods, using a linear regression method to obtain the output of each LSTM neural network model in the final combined model The weight value in .

In order to solve the above technical problem, the present application further provides a camera terminal, wherein the camera terminal includes a processor, and the processor is configured to execute program data, and the implemented steps include:

The photographing parameters are used for photographing.

The processor is further configured to use the LSTM to predict a photo parameter setting value of the photographing terminal, which specifically includes:

Obtain historical data of camera parameters;

The historical data is cleaned and normalized;

Performing off-line model training on the training data of the training data set to separately train multiple neural network models of the long-term and short-term memory neural network LSTM;

The processor is further configured to predict a photographing parameter setting value by using a rolling time window after adjusting the weight values occupied by the plurality of neural network models as the combined model according to the prediction effect.

The processor is specifically configured to convert the added and subtracted value predicted by the combined model into a predicted value of the predicted time, and then fill the currently predicted predicted value into a time window of the next predicted time, and alternately cycle according to the same And when the actual value actually set by the photographing parameter is obtained, the processor is configured to compare the predicted value with the actual value, and substitute the actual value as a new set of training data according to the comparison result, and substitute the model into the model to update the model parameter.

The processor acquires historical data of the photographing parameter, including: acquiring contrast, sensory degree, aperture, shutter, ISO, focus, metering, and white balance to comprehensively obtain historical data of the photographing parameter.

The processor, which comprehensively obtains historical data of the photographing parameter, specifically includes:

The processor uses the accept-reject sampling method according to the data distribution characteristics of the photographing parameters, and selects similar scenes, similar objects, similar geographical locations, and related data of the same photographer with similarly photographed parameters, together with the data of the photographing parameters. The original historical data is constructed.

The processor divides the cleaned normalized historical data into a training data set and a test data set according to time, including:

The processor divides the early data in the historical data before the specified time into a training data set, and divides the late data in the historical data after the specified time into the test data set.

The processor, before performing offline model training on the training data of the training data set, further includes: the processor generating time series data of different spans by using training data of the training data set, where, for each time Span (t ₀ , t ₁ , t ₂ , ... t _n ), using (t ₀ , t ₁ , t ₂ , ... t _n-1 ) as input values, using t _n-1 and t _n The difference value between them is discretized and converted into the unique heat code data as the supervisory value;

The processor performs offline model training on the training data of the training data set, and the corresponding comprises: the processor training each of the plurality of neural network models of the LSTM by using each time series data in the time series data of different spans.

The processor performs offline model training on the training data of the training data set, and specifically includes:

The processor performs training on the training data of the training data set by using a distributed training method based on memory calculation, wherein the training data is distributed to each node and the initial model parameters of the neural network model are broadcast to each node, Each node obtains the current gradient and the model parameter update amount according to the current model parameters and the training data of a certain scale, and updates the model parameters by summarizing the model parameter update amount fed back by each node, and broadcasts the updated model parameters to each node. According to this iteration, the training of a single LSTM neural network model is completed as required.

The processor calculates a weight value of a plurality of neural network models as a combined model, and specifically includes: the processor uses a plurality of time period training data, and uses a linear regression method to obtain each LSTM neural network. The weight value of the model in the final combined model output.

Beneficial effect

This application uses LSTM to predict the camera parameters, uses LSTM to divide the historical data into training data sets and test data sets according to time, and performs offline model training on the training data of the training data sets to train multiple neural networks of LSTM respectively. a model, and then, obtaining a list of predicted values of the training data for the plurality of trained neural network models, and comparing the predicted value list with the actual photographing setting values to calculate a plurality of neural network models as the combined model The weight value occupied, finally, using the test data of the test data set to evaluate the prediction effect of the plurality of neural network models in the combined model, and adjusting the weight values of the plurality of neural network models as the combined model according to the prediction effect, Finally, multiple neural network models are used to predict the camera parameters that the camera terminal may need to set.

In addition, the LSTM of the present application can improve the accuracy of prediction by combining models, and greatly reduce the error of prediction.

DRAWINGS

1 is a flow chart of an embodiment of a method for setting a photographing parameter based on a long-term and short-term memory neural network according to the present application.

FIG. 2 is a schematic flow chart of setting values of photographing parameters of the LSTM predictive photographing terminal of the present application.

3 is a block diagram of a module of an embodiment of a photographing terminal of the present application.

Embodiments of the invention

Referring to FIG. 1 , FIG. 1 is a flowchart of an embodiment of a method for setting a photographing parameter based on a long-short-term memory neural network according to the present application. The method for setting a photographing parameter based on an LSTM according to the embodiment includes, but is not limited to, the following steps.

In step S10, when the photographing terminal starts the photographing function, the photographing terminal obtains the predicted photographing parameter setting value based on the LSTM.

Step S20, setting a photographing parameter of the photographing terminal according to the obtained photographing parameter setting value.

In step S30, the photographing parameter is used for photographing.

Please refer to FIG. 2 . FIG. 2 is a schematic flowchart of the camera parameter setting value of the LSTM predictive camera terminal of the present application. The LSTM predictive camera terminal setting value of the camera terminal includes, but is not limited to, the following steps.

Step S100: Obtain historical data of the photographing parameter.

In step S100, the acquiring the historical data of the photographing parameter, the embodiment may specifically include: acquiring contrast, sensory degree, aperture, shutter, ISO, focus, metering, and white balance to comprehensively obtain the history of the photographing parameter. data.

Further, the historical data obtained by synthesizing the photographing parameters may include the following process: according to the data distribution characteristics of the photographing parameters, using the accept-reject sampling method to select similarly similar photographing parameters. Scenes, similar objects, similar geographical locations, and related data of the same photographer, together with the data of the photographing parameters, constitute the original said historical data.

It is not difficult to see that, in view of the existing actual situation, similar scenes, similar objects, similar geographical locations, and related data of the same photographer are selected, and the data of the photographing parameters are combined with the data of the photographing parameters to constitute the original historical data. Etc. As relevant historical data, the problem that the amount of data of the existing historical data is too small can be effectively solved.

Step S101: Perform data cleaning and normalization on the historical data.

Step S102: The cleaned normalized historical data is divided into a training data set and a test data set according to time.

It should be noted that the historical data that is normalized by the cleaning is divided into the training data set and the test data set according to the time. In this embodiment, the data may be specifically included: the early data before the specified time in the historical data. Divided into a training data set, the late data in the historical data after the specified time is divided into test data sets.

Step S103: Perform offline model training on the training data of the training data set to train multiple neural network models of the long-term and short-term memory neural network LSTM, respectively.

It is noted that, before performing the offline model training on the training data of the training data set in step S103, the method may further include: using the training data of the training data set to generate time series data of different spans; wherein, for each time span ( t ₀ , t ₁ , t ₂ , , , t _n ), using (t ₀ , t ₁ , t ₂ , , , t _n-1 ) as input values, using the difference value between t _n-1 and t _n After discretizing it, it is converted into the unique heat code data as the supervisory value.

The performing offline model training on the training data of the training data set in step S103 may include: training each of the plurality of neural network models of the LSTM by using each time series data in the time series data of different spans. It should be noted that the combined model of this embodiment may be a distributed training mode on a Spark (Distributed Memory Computing) platform.

In the step S103, the training data of the training data set is subjected to offline model training, and the embodiment may specifically include: training the training data of the training data set by using a distributed training method based on memory computing, wherein The training data is distributed to each node and the initial model parameters of the neural network model are broadcasted to each node, and each node obtains the current gradient and the model parameter update amount according to the current model parameters and the training data of a certain scale, by summarizing each The model parameter update quantity fed back by the node updates the model parameters, and broadcasts the updated model parameters to each node, and iterates repeatedly to complete the training of the single LSTM neural network model according to requirements.

It is not difficult to see that the LSTM model is used for the low precision problem of the regression problem, and the regressive problem of the camera parameter trend prediction is converted into the classification problem by the discretization means, which can effectively improve the prediction accuracy. In addition, the training based on the memory computing-based distributed training method of the present embodiment can effectively speed up the training.

Step S104: Obtain a predicted value column of the training data for the output of the plurality of neural network models after the training. The table compares the predicted value list with the actual photographing setting value, and calculates a weight value occupied by the plurality of neural network models as the combined model.

In the step S104, the calculating the weight value of the plurality of neural network models as the combined model, the embodiment may specifically include: using the training data of multiple time periods, using a linear regression method to obtain each LSTM neural network The weight value of the model in the final combined model output.

Step S105: Using the test data of the test data set to evaluate the prediction effect on the plurality of neural network models in the combined model, and adjusting the weight values occupied by the plurality of neural network models as the combined model according to the prediction effect.

It is not difficult to see that the present embodiment can propose a method of model combination for the problem that the prediction accuracy of a single LSTM is not high, and the accuracy of the prediction can be improved.

Specifically, in the specific application, the specific application may use the time window of different lengths to generate sequence data for the characteristics of the predicted time series data, train the LSTM model with different sequence data, and then combine them and determine by linear regression method. The weight of each model to improve prediction accuracy.

Step S106: predicting the camera parameter setting value by using a rolling time window.

In the step S106, the photographing parameter setting value is predicted by using the rolling window. The embodiment may specifically include: converting the added and subtracted value predicted by the combined model into the predicted value of the predicted time, and then predicting the current predicted The predicted value is filled in the time window of the next predicted time, and alternately cycles; when the actual value actually set by the photographing parameter is obtained, the predicted value is compared with the actual value, and the actual value is taken as a group according to the comparison result. New training data is substituted into the model to update the model parameters.

It should be noted that the scrolling window may be a single time window rolling cycle, or a plurality of groups of time windows may be scrolled together, which is not limited herein.

In this embodiment, the supervisory value is also the target value, which is a supervised learning concept involving machine learning in the technical field. In supervised learning, the algorithm in the calculation process of the present application passes the difference between the predicted value and the supervised value. The value is calculated, and then the model parameters are updated according to the loss, iteratively iteratively, and the training process in machine learning is realized, and finally the predicted value is the same as the supervised value.

By combining the models, the present invention avoids the problem that the simple prediction method of the single LSTM model has large error and low practicability, and further improves the prediction accuracy by calculating and adjusting the weight value of the combined model.

For example, the present application may include specific application examples described below.

(1) First, obtain historical data of photographing parameters;

(2) The historical data is cleaned and normalized, and then the cleaned historical data is divided into training data sets and test data sets according to time. For example, the earlier data is divided into training data sets, and the later data is Divided into test data sets. The data of the training data set is used to generate time series data x ₁ , x ₂ , , x _{n of} different spans, wherein the time spans (1<x ₁ <x ₂ , , , <x _n ). For each time span (t ₀ , t ₁ , t ₂ , , t _n ), use (t ₀ , t ₁ , t ₂ , , , t _n-1 ) as the input value x, using t _n-1 and The difference value between t _n is discretized and converted into the unique heat code data as the supervisory value Y.

(3) After the training data is generated, the offline model is trained, and the time series data x ₁ , x ₂ , , x _n generated by the training data set are used to train the LSTM neural network for each time series data. Model M ₁ , M ₂ , , , M _n .

(4) In view of the fact that the deep learning training speed is relatively slow, and the present embodiment needs to train multiple models, the present embodiment is based on a distributed training method of memory computing, first distributing data to each node, and then broadcasting the initial model parameters to Each node, each node obtains the current gradient and the model parameter update amount according to the current model parameters and the training data of a certain scale, and then updates the model parameters by summarizing the update amount fed back by each node, and then broadcasts the updated model parameters. Go out, repeat iteratively, and finally complete the training process of a single LSTM neural network model as required.

(5) randomly extracting a number of time periods in the training data set, and outputting a predicted value for the time (for example, t ₀ period) for each trained LSTM neural network model M _n .

Get a list of predicted values for each model

Then, using the real photo setting value y ⁰ as a reference, the training data for the t ₀ period is obtained.

Through the training data of multiple time periods, the weighting values of the respective LSTM neural network models in the final combined model output are obtained using the linear regression method.

(6) For the trained combination model, the test data set is used to evaluate the prediction effect, and the hyperparameters of the combined model are adjusted according to the evaluation result.

(7) In the actual prediction process, using the combined model, the rolling window is used to predict the setting values of the camera parameters, for example, converting the added and subtracted values predicted by the combined model into the predicted specific values, and then predicting the current prediction. The specific value is filled in the time window of the next moment, so that the cycle is alternated.

(8) When the actual setting data is obtained, the prediction result is compared with the actual result, and as a new set of training data, the model is substituted and the model parameters are updated.

In this way, the present application has higher accuracy and robustness for predicting camera parameters.

Please refer to FIG. 3. FIG. 3 is a block diagram of an embodiment of a camera terminal of the present application.

The present embodiment provides a photographing terminal, which may include a memory 20 for storing program data, a processor 21 for executing program data, and steps of implementing, including but not Limited to the following:

The photographing parameters are used for photographing.

The processor 21 is further configured to use the LSTM to predict a camera parameter setting value of the camera terminal, and specifically includes the following process.

Obtain historical data of camera parameters;

The historical data is cleaned and normalized;

The processor 21 is further configured to predict a photographing parameter setting value by using a rolling time window after adjusting the weight values occupied by the plurality of neural network models as the combined model according to the prediction effect.

The processor 21 is specifically configured to convert the added and subtracted value predicted by the combined model into a predicted value of the predicted time, and then fill the currently predicted predicted value into the time window of the next predicted time, and The alternating cycle; and when the actual value actually set by the photographing parameter is obtained, the processor 21 is configured to compare the predicted value with the actual value, and substitute the actual value as a new set of training data according to the comparison result, and substitute the model into Update model parameters.

The processor 21 acquires historical data of the photographing parameter, including: acquiring contrast, sensory degree, aperture, shutter, ISO, focus, metering, and white balance to comprehensively obtain historical data of the photographing parameter.

The processor 21, when comprehensively obtaining the historical data of the photographing parameter, specifically includes: the processor 21 uses the accept-reject sampling method to select a similar scene with similarly distributed photographing parameters according to the data distribution characteristics of the photographing parameters. Similar data, similar geographical locations, and related data of the same photographer, together with the data of the photographing parameters constitute the original historical data.

The processor 21 divides the cleaned normalized historical data into a training data set and a test data set according to time, and includes: the processor 21: the early data before the specified time in the historical data is located. Divided into a training data set, the late data in the historical data after the specified time is divided into test data sets.

The processor 21, before performing the offline model training on the training data of the training data set, further includes: the processor 21 uses the training data of the training data set to generate time series data of different spans, where, for each Time spans (t ₀ , t ₁ , t ₂ , ... t _n ), using (t ₀ , t ₁ , t ₂ , ... t _n-1 ) as input values, using t _n-1 and t The difference value between _n is discretized and converted into the unique heat code data as the supervisory value; the processor 21 performs offline model training on the training data of the training data set, and the corresponding includes: the processing The device 21 trains each of the neural network models of the LSTM using each of the time series data in different spans of time series data.

The processor 21 performs offline model training on the training data of the training data set, and specifically includes: the processor 21, using the distributed training method based on the memory calculation for the training data of the training data set Training, wherein the training data is distributed to each node and the initial model parameters of the neural network model are broadcast to each node, and each node obtains the current gradient and the model parameter update amount according to the current model parameters and the training data of a certain scale, The model parameters are updated by summarizing the model parameter update amounts fed back by the respective nodes, and the updated model parameters are broadcasted to the respective nodes, and the iteration is repeated to complete the training of the single LSTM neural network model according to the requirements.

The processor 21 calculates a weight value of a plurality of neural network models as a combined model, and specifically includes: the processor 21 obtains each LSTM by using a plurality of time period training data and using a linear regression method. The weight value of the neural network model in the final combined model output.

It should be noted that the camera terminal of the present application may be a mobile phone, a tablet computer, a wearable device, or a special camera. The wearable device may be a virtual reality helmet or the like, which is not limited herein.

Industrial applicability

Claims

A photographing parameter setting method based on a long-term and short-term memory neural network, wherein the photographing parameter setting method based on the long-term and short-term memory neural network comprises:

When the photographing terminal starts the photographing function, the photographed parameter setting value is obtained from the long-term and short-term memory neural network LSTM;

Setting a photographing parameter of the photographing terminal according to the obtained photographing parameter setting value;

The photographing parameters are used for photographing.
The photographing parameter setting method based on the long-short-term memory neural network according to claim 1, wherein the LSTM predicts a photographing parameter setting value of the photographing terminal, specifically comprising:

Obtain historical data of camera parameters;

The historical data is cleaned and normalized;

The historical data after normalization of the cleaning is divided into a training data set and a test data set according to time;

Performing off-line model training on the training data of the training data set to separately train multiple neural network models of the LSTM;

Obtaining a list of predicted values output by the training data for the plurality of neural network models after the training, comparing the predicted value list with an actual photographing setting value, and calculating a weight value occupied by the plurality of neural network models as the combined model;

The prediction data is evaluated by using the test data of the test data set for the plurality of neural network models in the combined model, and the weight values occupied by the plurality of neural network models as the combined model are adjusted according to the prediction effect.
The photographing parameter setting method based on long-short-term memory neural network according to claim 2, wherein:

After adjusting the weight values occupied by the plurality of neural network models as the combined model according to the prediction effect, the method further includes: predicting the set values of the photographing parameters by using a rolling time window;

The method for predicting the camera parameter setting value by using the rolling time window includes: converting the added and subtracted value predicted by the combined model into a predicted value of the predicted time, and filling the current predicted predicted value into the next predicted The time window of the moment, and alternately cycles according to this; when the actual value actually set by the photographing parameter is obtained, the predicted value is compared with the actual value, and the actual value is used as a new set of training data according to the comparison result, and the model is updated to be updated. Model parameters.
a photographing parameter setting method based on a long-term and short-term memory neural network according to claim 2, The obtaining historical data of the photographing parameter includes: acquiring contrast, sensory degree, aperture, shutter, ISO, focus, metering, and white balance to comprehensively obtain historical data of the photographing parameter;

The comprehensively obtaining the historical data of the photographing parameter includes:

According to the data distribution characteristics of the photographing parameters, using the accept-reject sampling method, similar scenes, similar objects, similar geographical locations and related data of the same photographer are selected, and the data of the photographing parameters are combined with the data of the photographing parameters to form the original place. State data.
The photographing parameter setting method based on long-short-term memory neural network according to claim 2, wherein:

The dividing the normalized historical data into the training data set and the test data set according to time, including: dividing the early data in the historical data before the specified time into the training data set, and using the historical data The late data after the specified time is divided into test data sets;

Before performing the offline model training on the training data of the training data set, the method further includes: generating time series data of different spans by using training data of the training data set; wherein, for each time span (t 0 , t 1 , t 2 , , , t n ), using (t 0 , t 1 , t 2 , , , t n-1 ) as the input value, using the difference value between t n-1 and t n to discretize it , converted to single heat code data as a supervisory value;

Performing offline model training on the training data of the training data set, the corresponding method comprises: training each of the plurality of neural network models of the LSTM by using each time series data in the time series data of different spans;

Performing offline model training on the training data of the training data set specifically includes: training training data of the training data set by using a distributed training method based on memory computing, wherein the training data is distributed to each node. The initial model parameters of the neural network model are broadcasted to each node, and each node obtains the current gradient and the model parameter update amount according to the current model parameters and the training data of a certain scale, and updates by updating the model parameter update amount fed back by each node. Model parameters, and broadcast the updated model parameters to each node, and iteratively iteratively, to complete the training of a single LSTM neural network model according to requirements;

The calculation calculates the weight values of the plurality of neural network models as the combined model, specifically including the training data through multiple time periods, and uses the linear regression method to obtain the weight of each LSTM neural network model in the final combined model output. value.
A camera terminal, wherein the camera terminal includes a processor, and the processor is configured to execute program data, and the implemented steps include:

When the photographing terminal starts the photographing function, the photographed parameter setting value is obtained from the long-term and short-term memory neural network LSTM;

Setting a photographing parameter of the photographing terminal according to the obtained photographing parameter setting value;

The photographing parameters are used for photographing.
The photographing terminal according to claim 6, wherein the processor is further configured to use the LSTM to predict a photographing parameter setting value of the photographing terminal, which specifically includes:

Obtain historical data of camera parameters;

The historical data is cleaned and normalized;

The historical data after normalization of the cleaning is divided into a training data set and a test data set according to time;

Performing off-line model training on the training data of the training data set to separately train multiple neural network models of the long-term and short-term memory neural network LSTM;

Obtaining a list of predicted values output by the training data for the plurality of neural network models after the training, comparing the predicted value list with an actual photographing setting value, and calculating a weight value occupied by the plurality of neural network models as the combined model;

The prediction data is evaluated by using the test data of the test data set for the plurality of neural network models in the combined model, and the weight values occupied by the plurality of neural network models as the combined model are adjusted according to the prediction effect.
The photographing terminal according to claim 7, wherein:

The processor is further configured to: after adjusting the weight values occupied by the plurality of neural network models as the combined model according to the prediction effect, predicting the set values of the photographing parameters by using a rolling time window;

The processor is specifically configured to convert the added and subtracted value predicted by the combined model into a predicted value of the predicted time, and then fill the currently predicted predicted value into a time window of the next predicted time, and alternately cycle according to the same And when the actual value actually set by the photographing parameter is obtained, the processor is configured to compare the predicted value with the actual value, and substitute the actual value as a new set of training data according to the comparison result, and substitute the model into the model to update the model parameter.
The photographing terminal according to claim 7, wherein the processor acquires historical data of photographing parameters, including: obtaining contrast, sensory degree, aperture, shutter, ISO, focus, metering, and white balance, to obtain a comprehensive The historical data of the photographing parameter; the processor, which comprehensively obtains the historical data of the photographing parameter, specifically includes:

The processor uses the accept-reject sampling method according to the data distribution characteristics of the photographing parameters, and selects similar scenes, similar objects, similar geographical locations, and related photographers with similarly distributed photographing parameters. The data, together with the data of the photographing parameters, constitutes the original said historical data.
The photographing terminal according to claim 7, wherein:

The processor divides the cleaned normalized historical data into a training data set and a test data set according to time, and the processor includes: dividing, by the processor, early data before the specified time in the historical data into training data. a set, the late data in the historical data after the specified time is divided into test data sets;

The processor, before performing offline model training on the training data of the training data set, further includes: the processor generating time series data of different spans using the training data of the training data set, where, for each time span ( t 0 , t 1 , t 2 , ... t n ), using (t 0 , t 1 , t 2 , ... t n-1 ) as the input value, using between t n-1 and t n The difference value, after discretizing it, is converted into the unique heat code data as the supervisory value;

The processor performs offline model training on the training data of the training data set, and the corresponding comprises: the processor training each of the plurality of neural network models of the LSTM by using each time series data in the time series data of different spans;

The processor performs offline model training on the training data of the training data set, and specifically includes: the processor, training the training data of the training data set by using a distributed training method based on memory computing, where The training data is distributed to each node and the initial model parameters of the neural network model are broadcasted to each node, and each node obtains the current gradient and the model parameter update amount according to the current model parameters and the training data of a certain scale, by summarizing the nodes. The updated model parameter update quantity is used to update the model parameters, and the updated model parameters are broadcasted to each node, and iteratively iteratively, to complete the training of a single LSTM neural network model according to requirements;

The processor calculates a weight value of the plurality of neural network models as a combined model, and specifically includes: the processor uses a plurality of time period training data, and uses a linear regression method to obtain each LSTM neural network model. The weight value in the final combined model output.