KR20200064044A - Method And Apparatus for Managing Automatic Learning - Google Patents

Abstract

Disclosed is an automatic learning apparatus and method for image recognition. The automatic learning method for image recognition of the present invention requests a learning request signal of a learning model for image recognition according to a learning instruction from a user terminal to a learning server, It receives the learning result corresponding to the learning request signal from the learning server, registers the input learning result, requests the evaluation of the learned learning model to the build server, and provides the evaluation result from the build server Includes learning means for requesting the received evaluation, registering the evaluation result, and outputting a selection request signal and a composite image generation signal for insufficient data according to the evaluation result to operate the learning model repeatedly. It is characterized by.

Description

Method and Apparatus for Managing Automatic Learning

The present invention relates to an automatic learning management apparatus and method, and more specifically, by recognizing data according to the evaluation results of the learning process, and generating synthetic data to reinforce the database, to improve image recognition to automatically improve the performance of the algorithm. For, it relates to an automatic learning management apparatus and method.

The contents described in this section merely provide background information for this embodiment, and do not constitute a prior art.

In general, in order to obtain information about an object from an image, the image must first be analyzed. Conventional image processing techniques have regarded images as two-dimensional signals and have performed two-dimensional signal processing. This method was also used to recognize objects in the image. An example is a vehicle license plate recognition system.

However, the conventional image processing technique has been insufficient to be used for real-time object recognition in that it takes a lot of extraction time through several stages of processing and that original image information may be lost.

To solve this problem, research using deep learning has been recently conducted. Deep learning refers to a deep neural network (DNN) algorithm and a method of learning it. Initially, it took a long time and had low utilization due to the disadvantage of overfitting with learning data. However, with the advent of graphics processing units (GPUs) capable of parallel computing and the emergence of new algorithms, object recognition technology in images has also been significantly improved. At present, the error rate is very low, which is very close to the human error rate.

On the other hand, in order to generate a DNN model having such a low error rate, it is necessary to collect a large number of images corresponding to big data, and a lot of reinforcement learning needs to be performed based on the collected images, which consumes a lot of time and resources. Consumption of such time and resources has been a burden for recognizing an object using a DNN model on the on board.

Background of the invention is disclosed in Republic of Korea Patent Publication No. 2018-0058175 (2018.05.10. public, DNN-based object learning and recognition system).

For reinforcement learning, the developer manually selects a vast amount of data to learn, waits until the learning is completed, and checks the evaluation results. Then, based on the evaluation results, the developer selects the data necessary for learning from the vast amount of data and repeats the above process.

If the data of the object to be recognized is insufficient based on the evaluation result, there is a hassle that the developer must perform all of these processes individually, such as collecting data and manually making a correct answer sheet. There was a problem in that a large amount of work was required for the process and the data logging task.

The present invention has been devised to improve the above problems, and the object of the present invention according to an aspect is to automatically enter a learning instruction and automatically progress a learning process to select data according to evaluation results and log insufficient data. It is to provide an automatic learning apparatus and method for image recognition that automatically fosters the performance of an algorithm by generating synthetic data as well as actual data to reinforce the database.

According to an aspect of the present invention, in an automatic learning management apparatus included in an automatic learning apparatus for image recognition, a learning request for a learning model for image recognition according to a learning instruction from a user terminal Request means for requesting a signal to a learning server; Registration means for receiving a learning result corresponding to the learning request signal from the learning server and registering the input learning result; Evaluation request means for requesting an evaluation of the learned learning model to a build server and receiving evaluation results from the build server; And learning means for registering the evaluation result and outputting a selection request signal and a composite image generation signal for insufficient data according to the evaluation result to operate the learning model repeatedly. It provides an automatic learning management device characterized by.

According to another aspect of the present invention, in a method of managing automatic learning using an automatic learning management apparatus for image recognition, the learning model for image recognition according to a learning instruction from a user terminal A learning request process for requesting a learning request signal to a learning server; A learning result registration process for receiving a learning result corresponding to the learning request signal from the learning server and registering the input learning result; An evaluation request process for requesting an evaluation of the learned learning model to a build server and receiving evaluation results from the build server; And a learning process of registering the evaluation result and outputting a selection request signal and a composite image generation signal for insufficient data according to the evaluation result to operate the learning model repeatedly. Provides an automatic learning management method characterized by this.

The automatic learning management apparatus and method according to an aspect of the present invention automatically inputs a learning instruction to automatically progress a learning process, selects data according to evaluation results, and insufficient data is generated by generating synthetic data as well as logged actual data. By reinforcing, the performance of the algorithm can be improved automatically, and the learning process can be automated to reduce work efficiency and development time.

1 is a block diagram showing an automatic learning apparatus for image recognition according to an embodiment of the present invention.
2 is a view for explaining the image synthesis process in the automatic learning device for image recognition according to an embodiment of the present invention.
3 is a flowchart illustrating an automatic learning method for image recognition according to an embodiment of the present invention.

Hereinafter, an automatic learning apparatus and method for image recognition according to the present invention will be described with reference to the accompanying drawings. In this process, the thickness of the lines or the size of components shown in the drawings may be exaggerated for clarity and convenience. In addition, terms to be described later are terms defined in consideration of functions in the present invention, which may vary according to a user's or operator's intention or practice. Therefore, the definition of these terms should be made based on the contents throughout the present specification.

1 is a block diagram illustrating an automatic learning apparatus for image recognition according to an embodiment of the present invention, and FIG. 2 illustrates an image synthesis process in an automatic learning apparatus for image recognition according to an embodiment of the present invention. It is for drawing.

As shown in Figure 1, an automatic learning device for image recognition according to an embodiment of the present invention, a configuration management device (configuration management device 50), a learning server (learning server, 40), a build server (build server, 60) and an automatic learning management device (20).

The configuration management device 50 includes learning scripts for learning images for learning images provided by the learning server 40 and the build server 60, and source code and binary files of the learned learning models ( Binary file) is stored and managed, and provided to the learning server 40 and the build server 60 when performing learning or building and learning model to evaluate.

The learning server 40 downloads a learning script from the shape management device 50 based on data stored in the learning data DB (DataBase, 30) and sets a learning environment to perform learning for image recognition, In addition to transmitting the learning results to the automatic learning management apparatus 20, the learned learning model is uploaded to the shape management apparatus 50.

The build server 60 downloads the learning model and source code learned from the configuration management device 50, sets the build environment to evaluate the learning model, builds the source code as a binary file, performs evaluation, and evaluates the results. To the automatic learning management device 20 as well as uploading the binary file and evaluation results to the configuration management device 50.

The automatic learning management device 20 requests the learning server 40 to learn the learning model for image recognition according to the learning command of the user terminal 10, registers to receive the learning result, and performs storage, etc. Request the evaluation of the learning model to the build server (60) and register by receiving the evaluation result.

In addition, when the learning result is registered from the learning server 40 or the evaluation result is registered from the build server 60, the automatic learning management device 20 transmits the learning result and the evaluation result to the user terminal 10 through email. To allow the user to confirm.

On the other hand, the automatic learning management apparatus 20 may select the insufficient data from the learning data DB 30 again according to the evaluation result, and repeatedly request the learning of the learning model to automatically nurture the performance of the algorithm.

The image synthesis device 70 generates a composite image of insufficient data according to the evaluation result requested from the automatic learning management apparatus 20 and stores it in the learning data DB 50 to store the insufficient data according to the performance evaluation result. You can also reinforce yourself to perform enhanced learning.

As shown in FIG. 2, the image synthesizing device 70 renders insufficient data by reflecting traffic such as urban environment, traffic volume, and pedestrians based on a world model as shown in (a). Data images can be synthesized, and as shown in (B), a large amount of composite images can be generated by reflecting time and weather. With respect to the generated composite image, GT (Ground Truth) can be extracted and annotated as shown in (C), stored in the learning data DB 30, and insufficient data can be reinforced.

As described above, according to the automatic learning apparatus for image recognition according to an embodiment of the present invention, the learning process is automatically received by receiving a learning command to select data according to the evaluation result, and the insufficient data is logged. By generating synthetic data as well as synthetic data and reinforcing the database, the performance of the algorithm can be automatically nurtured, and the learning process can be automated to improve work efficiency and shorten development time.

3 is a flowchart illustrating an automatic learning method for image recognition according to an embodiment of the present invention.

As shown in FIG. 3, in the automatic learning method for image recognition according to an embodiment of the present invention, first, the automatic learning management apparatus 20 receives a learning command from the user terminal 10 and projects for automatic learning ( project) is prepared (S10).

After preparing the project in step S10, the automatic learning management apparatus 20 requests the learning server 40 to learn the learning model for image recognition (S12).

If the learning model learning is requested in step S12, the learning server 40 downloads the learning script of the learning model from the shape management device 50 (S14).

After downloading the learning script in step S14, the learning server 40 sets a learning environment for performing learning (S16).

After setting the learning environment in step S16, the learning server 40 performs learning on the learning model based on the data stored in the learning data DB 30 (S18).

After performing the learning in step S18, the learning server 40 transmits the learning results (S20).

Here, the learning server 40 not only uploads the learned learning model to the shape management device 50, but also transmits the learning results to the automatic learning management device 20.

The automatic learning management apparatus 20 that receives the learning result transmitted in step S20 registers the input learning result (S22).

After registering the learning results in step S22, the automatic learning management apparatus 20 transmits the learning results to the user terminal 10 via email so that the user can check (S24).

After sending to the user terminal 10 via e-mail in step S24, the automatic learning management apparatus 20 requests the build server 60 to evaluate the learned learning model (S26).

When requesting a build for evaluation of the learning model learned in step S26, the build server 60 downloads the learning model and source code from the configuration management device 50 (S28).

After downloading the learning model and source code in step S28, the build server 60 sets the build environment (S28).

After setting the build environment in step S28, the build server 60 builds the source code as a binary file and performs evaluation (S30).

After performing the evaluation in step S30, the build server 60 transmits the evaluation result (S34).

Here, the build server 60 not only uploads the binary file and the evaluation result to the shape management device 50, but also transmits the evaluation result to the automatic learning management device 20.

The automatic learning management device 20 receiving the evaluation result transmitted in step S34 registers the input evaluation result (S36).

After registering the evaluation result in step S36, the automatic learning management device 20 transmits the evaluation result to the user terminal 10 via e-mail so that it can be checked (S38).

In addition, the automatic learning management device 20 selects the insufficient data according to the evaluation result from the learning data DB 30 again, and requests the video synthesis device 70 to generate a synthetic image for the insufficient data according to the evaluation result. The learning of the learning model is repeatedly requested (S40).

As described above, according to the automatic learning management method and apparatus according to the embodiment of the present invention, the learning process is automatically processed by receiving a learning command to select the data according to the evaluation result, and the insufficient data is not only logged actual data. In addition, by creating synthetic data and reinforcing the database, the performance of the algorithm can be automatically nurtured, and the learning process can be automated to reduce work efficiency and development time.

The present invention has been described with reference to the embodiment shown in the drawings, but this is only exemplary, and those skilled in the art to which the art belongs can various modifications and equivalent other embodiments from this. Will understand.

Therefore, the true technical protection scope of the present invention should be defined by the claims below.

10: user terminal 20: automatic learning management device
30: learning data DB 40: learning server
50: Configuration management device 60: Build server
70: video synthesis device

Claims (11)

Request means for requesting a learning request signal of a learning model for image recognition to a learning server according to a learning command from a user terminal;
Registration means for receiving a learning result corresponding to the learning request signal from the learning server and registering the input learning result;
Evaluation request means for requesting an evaluation of the learned learning model to a build server and receiving evaluation results from the build server; And
Learning means for registering the evaluation result and outputting a selection request signal and a composite image generation signal for insufficient data according to the evaluation result to operate the learning model repeatedly.
Automatic learning management device comprising a. According to claim 1,
The learning means,
Re-selection of learning data regarding the insufficient data in the learning data DB (DataBase) according to the selection request signal
Automatic learning management device, characterized in that. According to claim 1,
An automatic learning management apparatus further comprising image synthesis means for generating a composite image of the insufficient data according to the composite image generation signal and storing the generated composite image in a learning data DB. According to claim 1,
Automatic learning management device further comprises a transmission means for transmitting the registered learning results and the registered evaluation results to the user terminal. A learning request process for requesting a learning request signal of a learning model for image recognition according to a learning command from a user terminal to a learning server;
A learning result registration process for receiving a learning result corresponding to the learning request signal from the learning server and registering the input learning result;
An evaluation request process for requesting an evaluation of the learned learning model to a build server and receiving evaluation results from the build server; And
A learning process that registers the evaluation result and outputs a selection request signal and a composite image generation signal for insufficient data according to the evaluation result to operate the learning model repeatedly.
Automatic learning management method comprising a. The method of claim 5,
The learning process,
An automatic learning management method characterized in that the learning server downloads a learning script from a configuration management device, sets a learning environment, and the learning server repeatedly performs learning of the learning model based on data stored in the learning data DB. . The method of claim 6,
Automatic learning management method, characterized in that it further comprises a process of outputting the learning result after the learning model repeatedly performs learning. The method of claim 7,
Automatic learning management method, characterized in that it further comprises a process of uploading the learned learning model to the shape management device. The method of claim 5,
The learning result registration process,
Automatic learning management method, characterized in that for transmitting the learning results to the user terminal via email. The method of claim 5,
The evaluation request process,
The build server downloads the learning model and source code learned from the configuration management device, sets the build environment, and the build server builds the source code into a binary file and performs evaluation to obtain evaluation results. Automatic learning management method. The method of claim 10,
And an upload process of uploading the binary file and the evaluation result to the configuration management device after the build server performs the evaluation.

Images