KR20190143527A - System and Method for Recognitioning Image Pattern using Machine Learning - Google Patents

Abstract

The present invention relates to an apparatus to recognize an image through machine learning, capable of determining an option for obtaining a model parameter applying a reinforcement learning-based exploration technique to option exploration and calculating the best accuracy, and a method thereof. The apparatus includes: a data preprocessing part analyzing the distribution of the positions of objects in an image to be detected, enabling a user to select settings for a net structure or the like, and determining a net structure through analysis on the average size of the objects to be detected; a source data processing part processing learning, verification and test data; a binary file creation part outputting a file made by encoding a data set and information; an option determining part determining an option for obtaining a model parameter calculating the highest accuracy by applying a reinforcement learning-based exploration technique to option exploration; and a learning execution part evaluating model performance through learning using the option determined by the option determining part, while monitoring a learning process and an evaluation process and storing a completely learned model.

Description

System and Method for Recognitioning Image Pattern using Machine Learning

TECHNICAL FIELD The present invention relates to image recognition. Specifically, the present invention relates to image recognition through machine learning to apply a reinforcement learning-based search technique to option searching and to determine an option for obtaining model parameters that yield the highest accuracy. An apparatus and method are provided.

Artificial Intelligence (AI) mimics the human brain and neuronal neural networks, someday allowing computers and robots to think and act like humans.

For example, we can distinguish between dogs and cats very easily with pictures, but not computers.

The machine learning (ML) technique was devised to input a lot of data into a computer and classify similar things. When a picture similar to a stored dog's picture is input, it is called a dog picture. To classify.

Depending on how the data is to be sorted, many machine learning algorithms emerge, including Decision Trees, Bayesian networks, support vector machines (SVMs), and artificial neural networks. did.

Among them, deep learning (DL), which is derived from an artificial neural network algorithm, is a technique used to cluster or classify data using an artificial neural network.

Artificial neural networks in machine learning and cognitive science are statistical learning algorithms inspired by biological neural networks (the central nervous system of animals).

Artificial neural network refers to the overall model that has a problem solving ability by artificial neurons (node) formed by the combination of synapses to change the strength of the synapse through learning.

This machine learning (ML) technique is widely used in various fields such as image recognition, voice recognition, data mining, and intelligent robots.

In particular, in the field of image recognition, the numerical value exceeding the recognition accuracy of human is achieved.

As such, machine learning (ML) techniques provide high maturity for the task, but also require high computational complexity.

For example, given the problem of object location and class determination, it is necessary to determine the various options used in learning to solve it through effective learning.

There are so many options (more than a dozen) to choose from, and every single choice affects learning and performance.

This option selection occurs and is affected by various factors such as the type of problem or application field, data type, type of deep net to be used, learning rate, and whether data enhancement is applied.

The person who learns through the experience of deep learning for a long time analyzes a given problem and selects the optimal option, and also selects the option by examining the basic examples and cases provided by the developer.

This process requires a long time for the user and requires skilled data analysis skills.

However, even with this process, it is practically impossible to select the optimal option to effectively solve the problem to be solved.

For example, suppose you have 10 options that have only applicability (two states), and you test the effect of the applicability of each option. The number of cases for the priority of each option is 9! * There are a number of ¹⁰ cases, and it is impossible to do the learning and analyze and test the results for each of these cases.

Thus, there is a need for developing a new technique for determining options to obtain model parameters that yield the highest accuracy.

Korea Patent Registration No. 10-1850286 Korean Patent Publication No. 10-2016-0122452 Korean Patent Publication No. 10-2010-0129783

The present invention is to solve the problem of the image recognition technology using the conventional machine learning, the option decision to apply the reinforcement learning-based search method to the option search and obtain the model parameters that yield the highest accuracy An object of the present invention is to provide an apparatus and method for image recognition through machine learning.

An object of the present invention is to provide an apparatus and method for image recognition through machine learning, which enables the user to determine the location and class of an object that a user wants to detect using deep learning technology using Google's tensorflow framework. have.

The present invention analyzes the distribution of objects, extracts the parts to be actually processed using the ROI SEL module, and selects a suitable structure through real-time processing request speed analysis, user setting, average size analysis of objects, and ratio analysis of objects in the image. It is an object of the present invention to provide an apparatus and method for image recognition through machine learning, including a preprocessing configuration.

The present invention is a model parameter for calculating the best accuracy by repeating the process of analyzing the compensation value and the current state, the option and sequence generation, model training and accuracy verification, the model option storage and the compensation value for the training is appropriate to the controller It is an object of the present invention to provide an apparatus and method for image recognition through machine learning, including an option decision configuration to obtain a.

The objects of the present invention are not limited to the above-mentioned objects, and other objects that are not mentioned will be clearly understood by those skilled in the art from the following description.

The apparatus for image recognition through machine learning according to the present invention for achieving the above object is to analyze the distribution of the position of the object in the image to be detected, so that the user can select the setting for the net structure, etc. A data preprocessor configured to determine a net structure by performing an average size analysis of an object to be detected; a source data processor configured to learn, verify, and test data; a binary file generator configured to output a file encoded with a data set and information; An option decision unit which applies a reinforcement learning-based search method to option search to determine the model parameter that yields the highest accuracy; model determination by learning using the model option determined by the option decision unit. A learning execution unit for monitoring the evaluation, the learning process and the evaluation process, and storing the completed model; Characterized in that it comprises a.

Here, the data preprocessing unit may select a part to be actually processed according to an object distribution analysis result of the object distribution analyzer OBJ LOC DISTRIB and the object distribution analyzer OBJ LOC DISTRIB, which analyzes the location or distribution of the object in the image. A user who sets the ROI SEL to be extracted, the TAC-TIME EVAL for analyzing the real-time processing speed of the inspection apparatus, and the user setting required for the deep-net configuration. USER SETTING, OBJ SIZE EVAL, which analyzes the average size of the object, and OBJ-IMG RATIO EVAL, which analyzes the proportion of the object in the image, and object distribution and processing requirements. And a net structure determination unit (DEEP-NET STRUC DECESION) for determining a suitable net structure based on the speed, the average size of the object, and the object ratio analysis.

The source data processor may include an image data processor for processing a learning, verification, and test data set, an XML file processor for processing label information of a data set, and a model for managing deep-net information. Model Config, learning parameter information processing unit (Train Config) to process the learning parameter information, verification parameter information processing unit (EvalConfig) to process the verification parameter information, data set path management unit (Input set management) Config).

The option determiner may include a controller that analyzes a reward value and a current state of appropriateness of learning, an option sampler that generates data augmentation options and sequences, and a training environment. Train Config Update, Model Construction for importing and configuring the model needed for the Trainer Object Detection API from the source data, Model Construction, and verifying model training and accuracy. Train and Eval, Save Aug Opt, which stores data augmentation options after analyzing accuracy, and Optimal Options Store, which stores the model options that yield the highest accuracy. (Best Option) and a compensation value providing unit (RewardCAL) for providing a compensation value obtained from the accuracy to the controller (Controller).

The option decision unit evaluates the performance of the learner through the combination of the options represented by this sample when one sample is selected in space, and uses the evaluation performance as a reward for reinforcement learning. Detecting the distribution.

In addition, the option decision unit uses a long short term memory network (LSTM) as a detection algorithm, and the output of the LSTM network selects an odd term option and an even term selects the option according to a time step. It is characterized by.

The final tensorflow object detection API includes a train stage for learning by using the model options determined by the option determiner, and a model evaluation state for evaluating model performance in the learning process. And a monitoring unit (Tensorboard) for monitoring the learning process and the evaluation process, and a model storage unit (ModelExporter) for storing the completed model.

The method for image recognition through machine learning according to the present invention for achieving another object is to analyze the distribution of the position of the object in the image to be detected, and to allow the user to select settings for the net structure, etc. A data preprocessing step of determining a net structure by performing an average size analysis of an object to be detected; a source data processing step of processing learning, verifying, and test data; generating a binary file outputting a file encoded with the data set and information; An option decision step of applying a reinforcement learning-based search method to the option search to determine the model parameter that yields the highest accuracy; model determination by learning using the model option determined in the option decision step. Assessment, learning and monitoring of the assessment process and the storage of completed models It characterized in that it comprises a; execution phase.

Here, the data preprocessing step may include an object distribution analysis step of analyzing an existence position or a position distribution of an object in the image, a region of interest selection step of extracting an actual part to be processed according to an object distribution analysis result, and real-time processing of an inspection device Processing to analyze the required speed The required speed analysis step, the user setting step to set the user required for Deep-Net configuration, the size analysis step to analyze the average size of the object, and the ratio analysis of the ratio of the object to the image It characterized in that it comprises an analysis step.

The option determining step may include analyzing a reward value and a current state of appropriateness of learning, generating a data augmentation option and sequence, an updating step of updating a train config, The model construction step of importing and constructing the model required for the Tensorflow Object Detection API from the source data, the verification step of verifying model training and accuracy, and the data augmentation option after analyzing the accuracy in the buffer. And an option storing step for storing an option storing the model option for calculating the highest accuracy, and a step of providing a compensation value for providing the controller with a compensation value obtained from the accuracy.

The apparatus and method for image recognition through machine learning according to the present invention have the following effects.

First, we apply reinforcement learning based retrieval techniques to option exploration and make option decisions to obtain model parameters that yield the highest accuracy.

Second, the Google tensorflow framework allows users to make models and options decisions that are optimized for the location and class of objects they want to detect, without the need for deep learning technology.

Third, model parameters that calculate the best accuracy by repeating the process of analyzing the compensation value and current status of the appropriate training, generating the option and sequence, model training and verifying accuracy, storing the calculated model option and transmitting the compensation value to the controller You can make an option decision.

1 is a block diagram of an apparatus for image recognition through machine learning according to the present invention
2 to 5 are detailed block diagrams of an apparatus for image recognition through machine learning according to the present invention.
6 is a flowchart illustrating a method for image recognition through machine learning according to the present invention.

Hereinafter, a preferred embodiment of an apparatus and method for image recognition through machine learning according to the present invention will be described in detail.

Features and advantages of the apparatus and method for image recognition through machine learning according to the present invention will become apparent from the following detailed description of each embodiment.

1 is a configuration diagram of an apparatus for image recognition through machine learning according to the present invention, and FIGS. 2 to 5 are detailed configuration diagrams of an apparatus for image recognition through machine learning according to the present invention.

An apparatus and method for image recognition through machine learning according to the present invention is to apply a reinforcement learning-based search method to the option search and to determine the option to obtain the model parameters that yield the highest accuracy.

To this end, the present invention analyzes the distribution of objects, extracts the parts to be actually processed using the ROI SEL module, and is suitable for real-time processing request speed analysis, user setting, average size analysis of objects, and ratio analysis of objects in the image. It may include a pretreatment configuration for structure selection.

The present invention is a model parameter for calculating the best accuracy by repeating the process of the compensation value and current state analysis, option and sequence generation, model training and accuracy verification, model model storage and compensation value for the training is appropriate to the controller It may include an optional decision configuration to obtain.

The apparatus for image recognition through machine learning according to the present invention, as shown in Figure 1, analyzes the distribution of the position of the object in the image to be detected, and allows the user to select settings for the net structure, etc. The data preprocessing unit 100 determines a net structure by performing an average size analysis of an object to be detected, a source data processor 200 that processes learning, verification, and test data, and a file that encodes the data set and information. A binary file generator 300 for outputting, an option decision unit 400 for performing an option decision for obtaining a model parameter for calculating the highest accuracy by applying a reinforcement learning-based search technique to an option search, By using the model option determined by the option determiner 400, the model performance is evaluated, the learning process and the monitoring process are monitored, and the completed model is stored. And a learning execution unit 500.

The configuration of the data preprocessor 100 is the same as in FIG. 2.

The data preprocessor 100 includes an object distribution analyzer (OBJ LOC DISTRIB) 10 and an object distribution analyzer (OBJ LOC) for analyzing the position or distribution of positions of objects in the image through binary vision and blob detector package detectors. A region of interest selector (ROI SEL) 13 for extracting a portion to be actually processed according to the object distribution analysis result of the DISTRIB) 10, and a process request rate analyzer (TAC-) for analyzing a real-time process request rate of the inspection apparatus. TIME EVAL) 11, USER SETTING 14 for user setting required for Deep-Net configuration, OBJ SIZE EVAL 12 for average size analysis of the object, and OBJ-IMG RATIO EVAL (15), which analyzes the proportion of the object in the image, and the net structure for determining the appropriate net structure based on the object distribution, processing speed, object average size, and object ratio analysis. A DEEP-NET STRUC DECESION 16 is included.

The data preprocessor 100 assists the option decision in the option determiner 400 and processes the predecision options that can be determined by the user in advance or by analyzing the data to be learned.

The object distribution analyzer 10 analyzes the distribution of the position of the object in the image to be detected.

If the size of the object to be detected is smaller than the size of the entire image to be tested, the detection performance is lowered when the object is detected by searching the entire image.In this case, the location of a bag, box, or package that is likely to exist is found. You need to get it first.

In addition, depending on the type of sensor, there may be an area where an object is not always present in a part of the upper and lower margins, the left and right or the center in the image, and detecting all of these areas is inefficient in terms of calculation amount and detection performance.

In order to solve such a problem, the object distribution analyzer 10 analyzes the existence position or position distribution of the object in the image and detects the ROI (Region) to be detected by the ROI SEL 13. Of interest).

And the processing request rate analysis unit (TAC-TIME EVAL) 11 is to analyze the real-time processing request rate of the inspection equipment.

The real-time requirements of the inspection equipment to be applied vary depending on the application. Some systems require faster detection speeds, while others are slower but more important.

The processing speed and performance are tradeoffs, and you can choose the structure of the neural network to apply to deep learning.

That is, the processing request rate analyzing unit (TAC-TIME EVAL) 11 has a high performance net structure, a high speed structure instead of slow processing, and some loss of performance through the user setting unit (USER SETTING) 14. This is to allow the user to select settings for the net structure.

The OBJ SIZE EVAL 12 performs an average size analysis of the object to be detected, and determines an average ratio of the object to be detected in the image through analysis of the aspect ratio of the object size and the distribution of the area. .

In addition, the net structure determination unit (DEEP-NET STRUC DECESION) 16 includes an object distribution analyzer (OBJ LOC DISTRIB) 10, a processing request rate analyzer (TAC-TIME EVAL) 11, and an size analyzer (OBJ SIZE). EVAL) 12 determines the optimal net structure to apply based on the analysis results.

The detailed configuration of the source data processing unit 200 is the same as in FIG. 3.

As shown in FIG. 3, the source data processor 200 may include an image data processor for processing a learning, verification, and test data set, an XML file processor for processing label information of a data set, and a Deep file. Model Config for managing the Net information, the learning parameter information processing unit (Train Config) for processing the learning parameter information, the verification parameter information processing unit (EvalConfig) for processing the verification parameter information, data set path management And a data set path management unit (Input Config).

The detailed configuration of the optional option selector 400 is as shown in FIG. 4.

The option determiner 400 applies a reinforcement learning-based search technique to the option search. The controller 41 analyzes the reward value and the current state of the appropriate learning and the data augmentation option. And an optional sampler (Data Aug Opt sampler) 42 for generating a sequence, a Train Config Update 43 for updating Train Config, and a model required for the Tensorflow Object Detection API from the source data. Model Construction (44), Train & Eval (45) to validate model learning and accuracy, and Option Save (Save Aug Opt) to store the data augmentation options in the buffer after analyzing the accuracy. 46, an optimal option storage unit 47 for storing a model option having the highest accuracy, and an appropriate compensation value obtained from the accuracy to the controller 41. Including Study (RewardCAL) (48) It will obtain the Model parameters that yield the best accuracy by repeating these processes.

Here, data augmentation is a method of changing pixels without changing the label of the image, so that learning can be performed using the modified data.

The option decision unit 400 according to the present invention applies a retrieval learning-based search technique to the option search.

Among the main algorithms of reinforcement learning, the policy gradient method uses the assumption that the optimal distribution can be obtained by iterating several data samples obtained from this distribution, assuming that a combination of multiple factors represents a probability distribution. will be.

Several selectable options consider each option to constitute one axis in multidimensional space, and the combination of options results in some probability distribution of performance in multidimensional space.

When one sample is selected in space, the performance of the learner can be evaluated by combining the options represented by this sample, and the optimal performance is detected based on the policy gradient using the evaluation performance as a reward for reinforcement learning. .

In addition, the detection algorithm (Controller) may use Long Short Term Memory networks (LSTM), which is one of deep learning learning techniques.

The output of LSTM network selects option for odd term and option for even term according to time step.

The output of each term is passed to the input token of the next term.

For example, when there are five option lists, the first output selects one of five options, the selected option is passed to the input of the second output term, and the second output through the LSTM network outputs whether the selected option is available or not. The second term output is then passed back to the next term.

The data Aug Opt Sampler is used to construct the training model (Model Construction) and the optimization (Train & Eval) to output the best performance distribution (Best Option).

The detailed configuration of the learning execution unit 500 is the same as in FIG. 5.

Final Tensorflow Object Detection API (500) is a training stage (Train Stage) 51 for learning by using the model option determined in the option determiner 400, and evaluates the performance of the model in the learning process A model evaluation unit (Evaluation State) 52, a monitoring unit (Tensorboard) 53 for monitoring the learning process and the evaluation process, and a model storage unit (ModelExporter) 54 for storing the trained model.

The processing procedure of the apparatus for image recognition through machine learning according to the present invention having such a configuration is as follows.

6 is a flowchart illustrating a method for image recognition through machine learning according to the present invention.

First, an analysis on the distribution of the position where an object is placed in the image to be detected is performed (S601).

The average ratio of the object to be detected in the image is determined by analyzing the aspect ratio of the object size and the distribution of the area (S602), and the structure of the neural network to be applied to the deep learning learning is selected (S603).

Subsequently, training, verification, test data set processing, label information, deep-net information, learning parameter information, verification parameter information, and data set path processing of the data set are performed, and a file encoded with the data set and information is generated. )

Then, reinforcement learning based retrieval technique is applied to option retrieval and option decision is made to obtain model parameters that yield the highest accuracy (S605).

Subsequently, learning is performed using the Model Option obtained in the reinforcement learning process, model performance evaluation, monitoring the learning process and the evaluation process, and storing the completed model (S606).

Here, in the pre-processing process (S601) (S602) for selecting the structure of the neural network to be applied to deep learning learning, the object distribution analyzing the position or distribution of the position of the object in the image through binary vision and blob detector package detector The analysis step, the region of interest selection step for extracting the actual part to be processed according to the object distribution analysis result, the processing request rate analysis step for analyzing the real-time processing request rate of the inspection device, and the user setting required for Deep-Net configuration A user setting step, a size analysis step of analyzing the average size of the object, and a ratio analysis step of analyzing the ratio of the object occupies in the image.

In operation S605, an option decision for obtaining a model parameter that calculates the highest accuracy may include analyzing a compensation value and a current state of whether learning is appropriate, generating a data augmentation option and sequence, and calculating Train Config. An update step to update, a model construction step to load and construct a model required for the Tensorflow Object Detection API from source data, a verification step to verify model learning and accuracy, and an option to store data augmentation options in the buffer after analyzing the accuracy It may include a storing step, an optimal option storing step of storing the model option that has calculated the highest accuracy, and providing a compensation value providing the controller with an appropriate compensation value obtained from the accuracy.

Apparatus and method for image recognition through machine learning according to the present invention described above can be applied to reinforcement learning-based search method to option search and to determine the option to obtain the model parameters that yield the highest accuracy It is.

As described above, the present invention uses Google's tensorflow framework to accurately determine the location and class of the object that the user wants to detect using deep learning technology.

It will be understood that the present invention is implemented in a modified form without departing from the essential features of the present invention as described above.

Therefore, the described embodiments should be considered in descriptive sense only and not for purposes of limitation, and the scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the equivalent scope are included in the present invention. It should be interpreted.

100. Data preprocessor 200. Source data processor
300. Binary file generation unit 400. Option determination unit
500. Learning Execution Unit

Claims (10)

A data preprocessor configured to analyze the distribution of the position of the object in the image to be detected, allow a user to select a setting for the net structure, and determine a net structure by performing an average size analysis of the object to be detected;
A source data processor for learning, verifying, and testing data;
A binary file generator for outputting a file encoded with the data set and the information;
An option decision unit configured to apply a reinforcement learning-based search method to option search to determine an option for obtaining a model parameter that yields the highest accuracy;
Learning execution using the model option determined by the option determiner to learn the model performance evaluation, the learning process and the evaluation process and to store the completed model; learning for image recognition through machine learning comprising a Device. The data preprocessing unit of claim 1,
An ROI selector for extracting the part to be actually processed according to the object distribution analysis result of the object distribution analyzer OBJ LOC DISTRIB and the object distribution analyzer OBJ LOC DISTRIB. (ROI SEL),
A user setting unit (USER SETTING) for setting a user setting required for a TAC-TIME EVAL and a deep-net configuration for analyzing a real-time processing request rate of an inspection apparatus;
OBJ SIZE EVAL, which analyzes the average size of the object, and OBJ-IMG RATIO EVAL, which analyzes the proportion of the object in the image.
An apparatus for image recognition through machine learning, comprising a DEEP-NET STRUC DECESION for determining a suitable net structure based on an object distribution, a processing speed, an average size of an object, and an object ratio analysis. . The method of claim 1, wherein the source data processing unit,
An image data processing unit for processing a learning, verification and test data set,
XML file processing unit (XML Files) for processing the label information of the data set,
Model Config that manages Deep-Net information,
A learning parameter information processing unit (Train Config) for processing the learning parameter information,
A verification parameter information processing unit (EvalConfig) for processing the verification parameter information,
Apparatus for image recognition through machine learning, characterized in that it comprises a data set path management unit (Input Config) for data set path management. The method of claim 1, wherein the option determiner,
Controller that analyzes the current state and the reward value for learning is appropriate,
An optional data augmenter to generate data augmentation options and order,
An update unit (Train Config Update) for updating the training environment (Train Config),
A model construction unit for retrieving and constructing a model required for a training execution unit (Tensorflow Object Detection API) from source data,
Train & Eval to verify model learning and accuracy,
An option storage unit (Save Aug Opt) for storing the data augmentation option in a buffer after analyzing the accuracy;
An optimal option store for storing the model options that yield the highest accuracy,
And a reward value provider (RewardCAL) for providing a compensation value obtained from the accuracy to a controller. The method of claim 1 or 4, wherein the option determination unit,
When one sample is selected in space, the combination of options represented by this sample is used to evaluate the performance of the learner and to detect the optimal distribution based on policy gradient using the evaluation performance as a reward for reinforcement learning. Apparatus for image recognition through machine learning, characterized in that. The method of claim 1 or 4, wherein the option determination unit,
The detection algorithm uses Long Short Term Memory networks,
Output of the LSTM network according to the time step (time step), odd terms option selection, even terms select the option or not, the apparatus for image recognition through machine learning. The method of claim 1, wherein the learning execution unit (Final Tensorflow Object Detection API),
A training stage that trains using the model option determined by the option decision unit,
A model evaluation state for evaluating model performance in the learning process,
A tensorboard for monitoring the learning process and the evaluation process,
Apparatus for image recognition through machine learning, characterized in that it comprises a model storage unit (ModelExporter) for storing the completed model. A data preprocessing step of analyzing a distribution of positions of objects in the image to be detected, allowing a user to select a setting for a net structure, etc., and determining an net structure by performing an average size analysis of the object to be detected;
A source data processing step of processing learning, verification, and test data;
A binary file generation step of outputting a file encoded with the data set and the information;
An option determination step of applying a reinforcement learning-based search method to the option search to perform an option decision to obtain a model parameter that yields the highest accuracy;
Learning by using the model option determined in the option decision step, the model performance evaluation, the learning process and the learning execution step of storing the completed model; learning for image recognition through machine learning comprising a Way. The method of claim 8, wherein the data preprocessing step comprises:
An object distribution analyzing step of analyzing an existence position or a distribution of positions of an object in the image;
A region of interest selection step of extracting a portion to be actually processed according to an object distribution analysis result;
A processing request rate analyzing step of analyzing a real-time processing request rate of the inspection apparatus;
User setting step to set user setting required for Deep-Net configuration,
A size analysis step of analyzing the average size of the object, and a ratio analysis step of performing a ratio analysis of the object occupies in the image. The method of claim 8, wherein the option determination step,
Analyzing the rewards and current status of appropriate learning;
Creating a data augmentation option and order,
An update step to update the Train Config,
A model construction step of importing and constructing a model required for a training execution unit (Tensorflow Object Detection API) from source data,
Validation steps to verify model training and accuracy,
An option storage step of storing data augmentation options in a buffer after analyzing the accuracy;
An optimal option storage step that stores the model options that yield the highest accuracy,
And a compensation value providing step of providing a compensation value obtained from the accuracy to a controller.

Images