KR102240422B1 - Method for analysis of blind ends in tunnels using artificial intelligence and parameter optimization - Google Patents

Abstract

Provided is a method for automatically outputting analysis information on image information of an inputted tunnel face through a tunnel face automatic analysis system using learning that uses tunnel face image information as an input layer, and uses analysis information digital face-mapped to the tunnel face image information as an output layer. Optimal parameters required for learning were determined by converting a learning rate and a training step for a preset number of times, thereby providing a system excellent in accuracy and speed.

Description

Method for analysis of blind ends in tunnels using artificial intelligence and parameter optimization}

The present invention provides a method for automatically analyzing the scene of a tunnel using artificial intelligence and parameter optimization.

The construction of tunnel construction is performed in the order of charging and blasting, ventilation, burlock treatment, blind end surface observation and analysis, support and excavation pattern determination, support material installation, etc. It returns to the blasting step and repeats a number of times during the specified period.

Here, the accuracy of the observation and analysis of the scene is very important. Safe and effective follow-up work, that is, the next stage of excavation inside the curtain, is possible only when the analysis of the presence of carcinoma, rock quality, joints, and groundwater identified at the curtain surface is accurately performed.

The speed of observation and analysis of the scene is also important. In the case of blasting more than twice a day in blasting excavation, the analysis of the scene must be completed within 15 minutes. This is because there is a lot of time required for charging, blasting, ventilation, burlock treatment, and installation of support materials.

Various methods for this are proposed in the prior art.

Japanese Patent Publication No. 2019-023392A proposes a system for evaluating the dead end using machine learning. The prior art does not present a specific machine learning method, but when using a general machine learning method (Softmax classification, SVM, etc.), the accuracy of the final evaluation result is 30 to 40%, and thus there is a big problem that the accuracy is low. In addition, it takes a considerable amount of time to build a system based on machine learning and to infer the output layer based on the analysis of the scene image in the field, so the possibility of application in the actual field is very low.

Japanese Patent No. 4094770B2 proposes a rock mass determination system. Although artificial intelligence such as fuzzy logic is used, there is a problem that rock sample extraction is required and a rock analysis equipment must be provided at the tunnel construction site because it uses real rock data confirmed by an excavator rather than image analysis.

European Patent No. 1176393B1 proposes a last joint prediction system using machine learning. There is a problem that a 3D camera needs to be provided for each site, and carcinoma and rock quality other than joints cannot be determined.

As described above, there have been attempts in the prior art to automatically analyze the scene using artificial intelligence, but there are problems such as low accuracy, relatively long inference time, restrictions on the type of data to be identified, and expensive equipment required in the field. At the tunnel construction site, the analysis of the scene using artificial intelligence is not being applied.

A program that does not use artificial intelligence and supports manual analysis by manually dividing an area on an image captured by a camera is often used even today. However, this is only about a simple program in which the correction function is added after the image is loaded, and when the non-expert uses the program, the accuracy is lowered, so it is difficult to ensure quality and the analysis time is also increased.

(Patent Document 1) JP 2019-023392A

(Patent Document 2) JP 2018-071165A

(Patent Document 3) KR 10-2009-0020359A

(Patent Document 4) JP 4094770B2

(Patent Document 5) EP 1176393B1

(Patent Document 6) JP 2019-065648A

The present invention was devised to solve the above problems.

The important thing in analyzing the image information of the scene is the accuracy, speed, and possibility of modification by the user. According to the prior art, in the method using artificial intelligence, the accuracy was low and there was a problem in speed. You can speed up the inference time, but in this case the accuracy will be low. In some cases, a user interface that allows user modification was provided, but it was difficult for non-experts to approach. An object of the present invention is to solve this problem and propose a method in which accuracy, speed, and possibility of user modification can all be provided.

According to an embodiment of the present invention for solving the above problems, learning is performed using the last scene image information as an input layer and analysis information digital face mapping (DFM) to the last scene image information as an output layer. As a method of automatically outputting analysis information on the input film image information through the used film scene automatic analysis system,-Here, the analysis information includes information on carcinoma, information on rock quality, and information on joints. -(A1) The learning curtain image information input unit 110 loads the learning curtain image information prepared in advance for learning, and the learning analysis information input unit 120 loads the learning analysis information corresponding to the curtain surface image information, Setting learning information by mapping the loaded learning curtain image information and learning analysis information; (A2) the learning information dividing unit 130 generating a plurality of divided learning information for each piece of learning information by cropping the learning information as a preset unit; (A3) the format conversion unit 210 loading the plurality of divided learning information and converting the information into a learning format; (A4) The learning performing unit 230 uses the plurality of divided learning information converted in the format, but converts the learning rate and the training step for a predetermined number of times, while a plurality of meetings for each divided learning information Performing learning; (A5) By performing all of the steps (A1) to (A4) on a plurality of learning curtain image information, the learning parameter setting unit 220 determines the parameter so that the amount of calculation is minimized, and accordingly Building an analysis system; And (B) when the last scene image information is input to the input unit 310, the reasoning unit 330 infers the analysis information using the automatic analysis system for the last scene, and the output unit 350 outputs it. To do, provide a way.

In addition, it is preferable that the parameters determined in step (A5) are an optimizer parameter and a learning parameter.

In addition, it is preferable that the optimization parameter includes an optimizer, and the optimizer parameter determined in step (A5) is RMSProp.

In addition, the optimization parameters include weight_decay, Optimizer, opt_epsilon, rmsprop_momentum, and rmsprop_decay, and the learning parameters preferably include learning_rate, max_number_of_steps, learning_rate_decay_type, num_epochs_per_decay, replicas_to_aggregate, and moving_average_decay.

In addition, after the step (B), (C) the user modification module 400 correcting the analysis information output in the step (B); And (D) the learning information adding module 500 transmits the analysis information modified in step (C) and the corresponding scene image information to the learning analysis information input unit 120 and the learning curtain image information input unit 110, respectively. It is preferable to further include the step of inputting.

The present invention has the following advantages.

First, it is possible to check analysis information with high accuracy even by non-experts. Through learning, the last information is analyzed with high accuracy. As will be described later, the accuracy according to the prior art using artificial intelligence was only 30 to 40%, but an accuracy of 70% or more was verified in an embodiment actually implemented according to the present invention. A number of training sets verified by experts may be used, and the verification result may be reused as learning information when using an actual user, thereby further increasing the accuracy.

Second, the analysis is quick. As described above, the analysis of the scene should be completed within 15 minutes to proceed with the next stage of charging, blasting, and ventilation. One of the reasons for using artificial intelligence is for such rapid analysis, but in the prior art, it has been difficult to implement due to the large amount of computation of artificial intelligence and inference itself. The present invention has been implemented to enable rapid analysis in various ways. By segmenting and learning the image, using the watershed method, modifying and using the inception algorithm, and setting the optimal parameters, we implemented rapid computation.

Third, user modification is free. The results of the primary analysis are provided by artificial intelligence, but experts in the field can freely modify them while checking them. By adopting a node method for free movement of carcinomas represented by polygons and joints represented by rock formations and line segments, users can intuitively modify them, which further reduces the time required for deriving the final analysis result.

In addition, RQD (rock quality index) is automatically calculated, report format output is possible, and a user interface is provided to input joint strike slope and joint correction points.

There is no need to prepare separate equipment on site that is large, expensive, or difficult to maintain. If only one information processing device (for example, a mobile phone, a tablet, a special book, etc.) having a camera and a communication function is provided in the field, it is possible to sufficiently perform the method according to the present invention.

1 schematically shows a system for carrying out the method according to the invention.
2 shows a flow chart for carrying out the method according to the invention.
3 shows an example of a curtain surface image information for learning and analysis information for learning.
4 and 5 illustrate examples of a method of dividing learning information.
6 exemplarily shows the network of the learning method used in the present invention.
7 shows an example of a method of correcting image information of a film scene in the present invention.
8 shows an example of a watershed technique used to correct image information of a film scene in the present invention.
9A to 9C illustrate examples of canny edge detection techniques used in the present invention.
10 shows an example in which analysis information, which is a result inferred according to the present invention, is output.
11 shows an example of various RQD radiation center correction methods used in the present invention.
12 shows an example of a joint correction method by a user in the present invention.
13 shows an example in which a joint strike slope and a joint correction score are input by a user in the present invention.
14 shows an example of a method for modifying a polygon by a user in the present invention.

In the following, "learning" is to build an algorithm that automatically infers output information corresponding to the input information through a predetermined training when an information processing device such as a computer applies input information to the input layer and outputs it to the output layer. It refers to the data processing technique performed and is a kind of artificial intelligence. Machine learning may be included. Learning consists of training using a plurality of training sets in which the input layer and the output layer are determined. Learning is performed by the information processing device as the subject using a predetermined algorithm, and when the algorithm is established by performing a number of learning, it is possible to automatically perform inference in a predetermined manner according to the information input by the user. do. The algorithms used here can vary. Below, inception is proposed as an example of a learning algorithm to be used.

Hereinafter, "digital face mapping (DFM)" means mapping digitized information on a photographed image. By the DFM according to the present invention, information on carcinoma, rock quality, joints, etc. is digitally mapped on a film scene image taken by a user with a camera or the like. In other words, when the user inputs the last scene image using the method according to the present invention, the analysis information is automatically inferred and superimposed on the image to be output, and the analysis result is the DFM result. As shown in the upper part of FIG. 10, a joint may be shown as a line segment, or as shown in the lower part of FIG. 10, a zone of carcinoma (eg, granite, etc.) and rock quality (eg, severe weathering) may be divided.

Description of the system

The present method includes analysis information on the input image information of the last scene through an automatic analysis system using learning that uses the image information of the last scene as an input layer and the analysis information digitally mapped to the image information of the last scene as an output layer. It relates to a method of automatically outputting. Here, the analysis information includes information on carcinoma, information on cancer, and information on joints.

A system for automatically analyzing a film scene for performing the method according to the present invention will be described with reference to FIG. 1.

The system in which the method according to the present invention is performed includes: a learning information generation module 100 for generating learning information; A learning module 200 for constructing an automatic analysis system for the last scene by performing learning based on the learning information; A DFM module 300 for automatically inferring and outputting information such as carcinoma, rock quality, joints, etc. by analyzing the image of the film input by the user based on the built automatic analysis system for the film; A user modification module 400 that provides a function that can be modified by a user based on the inferred analysis information; And a learning information adding module 500 that converts the finally modified analysis information and the last scene image information into learning information to enable learning.

The learning information generation module 100 includes an image information input unit 110 for learning the last scene, an analysis information input unit 120 for learning, a learning information division unit 130 and a learning information database 190.

The learning module 200 includes a format conversion unit 210, a learning parameter setting unit 220, a learning execution unit 230, and a validity verification unit 240.

The DFM module 300 includes an input unit 310 and an image correction unit 320. An inference unit 330, a joint extraction unit 340, an output unit 350, an RQD calculation unit 360, and a DFM database 390 are included.

Each function is described in more detail below.

Preparing learning information

A method of preparing learning information to perform the method according to the present invention will be described with further reference to FIG. 2.

Prepare the input layer, the learning scene image information, and the output layer, the analysis information, for learning. It is preferable that the analysis information for learning is information verified by an expert. The learning screen image information and the learning analysis information are mapped to each other.

If there are N pieces of image information for the learning scene and N pieces of analysis information for learning corresponding thereto, some of them (A pieces) are used for learning and the rest (N-A pieces) are used for verification (N>A).

In the actual implementation system, 900 of the 1000 learning screen image information were used for learning, and the remaining 100 were used for verification.

The learning curtain image information input unit 110 loads the learning curtain image information prepared in advance for learning, and the learning analysis information input unit 120 loads the learning analysis information corresponding to the last scene image information. The image information of the curtain for learning and the analysis information for learning prepared in this way are referred to as learning information (FIG. 3).

It is important not to learn it as it is, but to learn it by dividing it. It is to quickly learn a lot of learning information by reducing the amount of computation. To this end, the learning information dividing unit 130 generates a plurality of divided learning information for each of the last scene image information by cropping the prepared learning information as a preset unit.

In the actual implementation system, the pixel unit was divided into 30×30 and stored (FIG. 4). As learning information, a plurality of scene image information and analysis information corresponding thereto are obtained by mapping, and the divided information is stored in the learning information database 190 as individual files (FIG. 5).

Learning method

The learning module 200 performs learning by using the divided learning information prepared by the learning information generating module 100.

First, the format conversion unit 210 loads a plurality of divided learning information and converts it into a learning format.

In an actual implementation embodiment, learning information was converted to .tfrecord so that learning can proceed using TF-Slim of tensorflow, including the width, height, depth, and label of the image identified in the last scene image information. (label), image raw, etc.

Next, the learning performing unit 230 uses the format-converted plurality of divided learning information, but converts the learning rate and training step for a predetermined number of times, while learning a plurality of times for each divided learning information. Perform. In this process, the learning parameter setting unit 220 may set an optimal parameter so that the amount of calculation is minimized. Optimal parameters will be described later.

Through this process, several times of learning are performed for each piece of divided learning information, and several times of learning are also performed on each of the learning information in which a plurality of pieces of divided learning information are collected.

In an actual implementation embodiment, the learning method used by the learning execution unit 230 is Inception. A deep learning network has better performance as the deeper and wider the layer is, but the computational amount also increases exponentially, so it is not suitable for the learning process for the analysis of the last scene as in the present invention. Inception reduces the connection between nodes and makes the matrix operation denser to reduce the amount of computation. The inception algorithm was adopted in consideration of the peculiarity of analyzing the film scene image as in the present invention (the size of the image itself is wide, the boundary between carcinoma and cancer is important, and line segment classification such as joints is important).

Inception is a method of reducing the number of parameters of the 3X3 and 5X5 convolution layers by reducing the image channel through a 1X1 convolution operation. Accordingly, while configuring a deeper network than CNN (Convolutional Neural Networks), the number of parameters is reduced and the computational amount is reduced.

Specifically, in general inception, (1) 1X1 convolution, (2) 1X1 convolution and 3X3 convolution, (3) 1X1 convolution and 5X5 convolution, and (4) 3X3 max pooling for the previous layer. And 1X1 convolutions, four types of operations are performed, and they are combined into one to proceed with learning.

In the present invention, while using inception, the need to perform a 7X7 convolution operation in the first layer was replaced with a 3x3 convolution operation three times, and the optimizer was set to RMSProp, and the last layer (fully connected) layer) was applied batch normalization (BN). In addition to this, the kernel was improved by applying factorization to further reduce the amount of computation and time by improving the kernel, performing 2 3X3 convolution operations instead of 1 5X5 convolution operation, and 2 3X1 convolution operations instead of 1 3X3 convolution. Improved. As a result, it was confirmed that the amount of computation was reduced by about 2.78 times. 6 schematically shows a network used in this way.

Meanwhile, the learning parameter setting unit 220 sets the final parameter according to the result of performing the learning while converting the learning rate and the training step. The parameters set here can be classified into an optimization parameter and a learning parameter.

Examples of optimization parameters include weight_decay, optimizer, opt_epsilon, rmsprop_momentum, and rmsprop_decay.

weight_decay means a weight for optimizing the model's loss function. Here, it was set to 0.00004 according to the optimization result.

opt_epsilon is the smallest value to prevent a case where the calculated value becomes 0. Here, it was set to 1 according to the optimization result.

rmsprop_momentum is the momentum value of RMSPropOptimizer. Here, it was set to 0.9 according to the optimization result.

rmsprop_decay is the rate of decrease of the probability applied for each update. Here, it was set to 0.9 according to the optimization result.

Learning parameters include learning_rate, max_number_of_steps, learning_rate_decay_type, num_epochs_per_decay, replicas_to_aggregate, moving_average_decay, and the like.

learning_rate is the initial learning rate. Here, it was set to 0.0001 according to the optimization result.

max_number_of_steps is the maximum number of training steps. Here, it was set to 1000 according to the optimization result.

Here, learning_rate and max_number_of_steps are the main parameters. In order to find an appropriate learning speed and learning step, learning was performed by changing the learning_rate to 0.0001, 0.001, 0.01, and 0.1, and in the case of max_number_of_steps, learning was conducted while changing 1,000 to 10,000 (increases by 1,000).

learning_rate_decay_type is a parameter that specifies how to decrease the learning rate. Examples include fixed, exponential, and polynomial. Here, it was set to fixed according to the optimization result.

num_epochs_per_decay is the number of epochs after the learning rate decreases. Here, it was set to 2 according to the optimization result.

replicas_to_aggregate is the number of gradients to collect before updating parameters. Here, it was set to 1 according to the optimization result.

moving_average_decay is the decay rate used for the moving average. Here, it was set to None according to the optimization result.

In this way, learning is performed several times for each divided learning information in one learning curtain image information, and this learning is performed on a plurality of learning curtain image information, thereby constructing an automatic analysis system for the curtain.

The validation unit 240 checks the constructed system by validating the validity of the prepared learning information using other unused last scene image information.

As a result of checking in the actual implementation example, it was confirmed that the accuracy was relatively high by confirming an accuracy of 70% or more. It showed a big difference compared to the case of using a learning algorithm other than inception.

Learning algorithm accuracy Softmax classification 30.6% SVM 47.8% Inception 66.8% Inception (actual application) 71.8%

Acquisition of analysis information using DFM

Now, users can automatically acquire analysis information such as carcinoma, rock quality, joints, etc. by entering the image information of the curtain using the established automatic analysis system for the curtain.

First, image information of the last scene is input to the input unit 310. While using an information processing device equipped with a communication function and a camera, the user may directly shoot with a camera to obtain and input the last scene image information, or select an image stored in the memory in the information processing device to input the input unit 310. You can also enter it through. In the field, it is enough to have an information processing device, and no separate equipment is required.

The image correction unit 320 corrects the image to be input in several ways.

The input image taken directly by the user further includes information on boundary objects such as the surrounding wall as well as the curtain surface (refer to the lower left corner of Fig. 7), and the present invention provides a correction function for extracting the image on the curtain surface. . That is, an image is input (top of Fig. 7) and a tunnel boundary node can be input on the last scene image information input by the image correction unit 320 (bottom left of Fig. 7), and the input tunnel boundary node is connected. And, by extracting only the connected area (bottom right of Fig. 7), the image information of the curtain is corrected. In the prior art, most of the images select a method of modifying the image rather than a method of selecting a part of the image. In the present invention, an area extraction method by nodes is selected.

In addition, as shown in FIG. 8, the image correction unit 320 may divide the inputted scene image information into a plurality of regions using a watershed technique. This reduces the amount of computation when inputting the last scene image information input as an input layer, thereby enabling rapid computation and inference in the inference unit 330 to be described later. As described in the prior art, speed is very important in analyzing the film scene image.

Next, the reasoning unit 330 infers the analysis information using a system constructed for each divided area, and infers the digital face-mapped analysis information for the inputted scene image information by integrating the analysis information. That is, the inference unit 330 inputs the last scene image information as an input layer of the constructed automatic analysis system for the last scene, and when the output layer by artificial intelligence is checked, it infers this as analysis information.

Meanwhile, in another embodiment of the present invention, joint extraction may be performed separately. The joint extraction unit 340 extracts a line segment from the input film image information using a canny edge detection technique (FIG. 9A). Next, the joint extraction unit 340 may remove noise by performing a predetermined test for each extracted line segment (FIG. 9B), and if the length of the line segment is less than a preset pixel value, it will be removed from the extraction result. It may be possible (Fig. 9c). Now, the remaining line segments will be output as joints.

When all the analysis information is arranged through this process, the output unit 350 outputs the analysis information as a result (FIG. 10). The result of this method may be stored in the DFM database 390, may undergo a user modification process to be described later, and may be reused as learning information to increase accuracy. It may be output in a report format as shown in the lower right corner of FIG. 10.

In another embodiment of the present invention, rock quality designation (RQD) may be calculated and further output. RQD calculation can be performed by any widely known method, but the important thing is to automatically set the center of the RQD radiation line. That is, the RQD calculation unit 360 includes a step of setting the center of the RQD radiation line, and a step of calculating and outputting the RQD by using the center of the RQD radiation line.

As a method for the RQD calculation unit 360 to set the center of the RQD radiation, the present invention provides four methods. First, the center of the RQD radiation line may be set as the center of the points constituting the joint in the analysis information (upper left in FIG. 11). Second, it is possible to set the center of the RQD radiation line as the center of the center points of the line segments constituting the joint among the analysis information (top right of FIG. 11). In other words, it is set around the dense area. Third, the center of the RQD radiation line may be set as the center of the longest line segment among the line segments constituting the joint of the analysis information (lower left of FIG. 11). Fourth, it is possible to set the center of the RQD radiation line as the center of the horizontal radiation line that intersects the joint most among the plurality of horizontal radiation lines set on the film image information (bottom right of FIG. 11).

Each of these methods may be necessary depending on the conditions of the scene and the site situation. Accordingly, in the present invention, all of these methods may be performed automatically or may be performed by a user selecting a specific method.

User Modification Module

The present invention provides a function that allows an expert who checks the analysis information to be automatically outputted by learning to correct carcinoma, cancer, joints, and the like. In particular, it adopts an intuitive method with high convenience so that experts can quickly make corrections and confirm analysis information. To this end, the user modification module 400 provides a function that the user can modify based on the analysis information automatically inferred by the output unit 350.

The output unit 350 outputs the joint as a line segment, and the user modification module 400 may add and output a joint node on the joint. The output joint can be added, moved, and deleted by the user correction module 400, and the output joint node can also be added, moved, and deleted by the user correction module 400. When the joint node is moved, the joint shape including the moved joint node is changed accordingly (FIG. 12). Joint strike slope (FIG. 13(A)) and joint correction score may be additionally input (FIG. 13(B)).

In addition, the output unit 350 outputs the area where carcinoma and rock quality are divided into polygons and, if necessary, divides the groundwater into polygons and outputs it. The user modification module 400 can add and output the polygonal corners as polygon nodes. have. The output polygon node can be added, moved, and deleted by the user modification module 400, and when the polygon node is moved, the polygon shape including the moved polygon node is changed accordingly (FIG. 14).

Add learning information

The analysis information that has been more reliably modified by an expert through the user modification process can be used as learning information together with information on the learning scene image used at the time of initial learning. In particular, since it has undergone a user modification process, a training set with excellent data accuracy is used. This increases the accuracy of the analysis.

The learning information addition module 500 may input the analysis information modified by the user and the corresponding scene image information into the analysis information input unit 120 for learning and the image information input unit 110 for learning, respectively. The information is used as learning information through the above-described steps.

In the above, the present specification has been described with reference to the embodiments shown in the drawings so that those skilled in the art can easily understand and reproduce the present invention, but these are only exemplary, and those skilled in the art can use various modifications and equivalents from the embodiments of the present invention. It will be appreciated that examples are possible. Therefore, the scope of protection of the present invention should be determined by the claims.

100: learning information generation module
110: learning screen image information input unit
120: learning analysis information input unit
130: learning information division
190: learning information database
200: learning module
210: format conversion unit
220: learning parameter setting unit
230: Learning Execution Department
240: validation unit
300: DFM module
310: input
320: image correction
330: reasoning unit
340: joint extraction unit
350: output
360: RQD calculation unit
390: DFM database
400: user modification module
500: add learning information module

Claims (5)

Through the automatic analysis system of the last scene using learning in which the image information of the last scene is used as an input layer and analysis information digital face mapping (DFM) is used as the output layer, the input image information of the last scene is As a method of automatically outputting the analysis information about -Here, the analysis information includes information on carcinoma, information on cancer, and information on joints-
(A1) The learning curtain image information input unit 110 loads the learning curtain image information prepared in advance for learning, the learning analysis information input unit 120 loads the learning analysis information corresponding to the curtain surface image information, and the Setting learning information by mapping the loaded learning curtain image information and learning analysis information;
(A2) generating, by the learning information dividing unit 130, a plurality of divided learning information for each piece of learning information by cropping the learning information as a preset unit;
(A3) the format conversion unit 210 loading the plurality of divided learning information and converting the information into a learning format;
(A4) The learning performing unit 230 uses the plurality of divided learning information converted in the format, but converts the learning rate and the training step for a predetermined number of times, while a plurality of meetings for each divided learning information Performing learning;
(A5) By performing all of the steps (A1) to (A4) on a plurality of learning last scene image information, the learning parameter setting unit 220 determines the parameter so that the amount of calculation is minimized, and accordingly, the last scene is automatically Building an analysis system; And
(B) when the last scene image information is input to the input unit 310, the reasoning unit 330 infers the analysis information using the automatic analysis system for the last scene, and the output unit 350 outputs it, ,
The step (B),
Inputting the image information of the last scene to the input unit 310;
Dividing the input film scene image information into a plurality of regions using a watershed technique by the image correction unit 320;
Inferring, by the reasoning unit 330, analysis information for each region through the automatic analysis system for the last scene, using each of the plurality of regions of the inputted last scene image information as an input layer; And
The inference unit 330 infers the digital face-mapped analysis information for the input film scene image information by integrating the analysis information for each region, and outputting the digital face-mapped analysis information by the output unit 350,
Way.
The method of claim 1,
The parameters determined in step (A5) are an optimizer parameter and a learning parameter,
Way.
The method of claim 2,
The optimization parameter includes an optimizer,
The optimizer parameter determined in step (A5) is RMSProp,
Way.
The method of claim 2,
The optimization parameters include weight_decay, Optimizer, opt_epsilon, rmsprop_momentum, and rmsprop_decay, and the learning parameters include learning_rate, max_number_of_steps, learning_rate_decay_type, num_epochs_per_decay, replicas_to_aggregate, and moving_average_decay.
Way.
The method of claim 1,
After step (B),
(C) modifying, by the user modification module 400, the analysis information output in step (B); And
(D) The learning information adding module 500 inputs the analysis information modified in step (C) and the corresponding scene image information into the learning analysis information input unit 120 and the learning curtain image information input unit 110, respectively. Further comprising the step of,
Way.

Images