KR20220039533A - Self improving object recognition method and system through image capture - Google Patents

Abstract

Provided are a method and a system for recognizing a self-improving object through image capture. The method for recognizing an object includes the steps of: collecting a first photographed image through a user terminal; predicting a photographing condition of the first photographed image; verifying the first photographed image using the predicted photographing condition and adding the image to a verification dataset; training an object recognition model using the verification dataset; and using the learned object recognition model, and obtaining a recognition result of the object indicated by a second photographed image.

Description

Self-improvement object recognition method and system through image capture

The present invention relates to a method and system for self-improving object recognition through image capture.

Artificial intelligence image recognition technology that recognizes objects using artificial intelligence is a technology that classifies a specific object in an image (classification), a technology that detects multiple objects simultaneously (detection), and a technology that identifies and divides objects in units of pixels ( segmentation) and the like. Although the quality level of artificial intelligence image recognition technology is increasing due to deep learning, deep learning requires large-scale data and computing power in the learning process.

In order to overcome this limitation of artificial intelligence learning, self-supervised learning to reduce the annotations required for learning by finding the characteristics necessary for solving problems and expressing them appropriately, can greatly contribute to learning. Research on various technologies such as active learning, which selects and annotates existing data first, and transfer learning, which learns a target domain from a source domain, is active.

An object of the present invention is to provide an object recognition method and system capable of providing a dataset in which various conditions are combined as learning data for learning an artificial intelligence model for object recognition.

Another object of the present invention is to provide an object recognition method and system capable of accurately recognizing objects from photographed images under various photographing conditions by using an object recognition model and a photographing condition prediction model together.

In particular, AI systems can be disrupted by unknown new data, in which case human (data scientist or engineer) intervention is required to sort, classify, and label that data and then re-introduce it into a new learning model. What you need is usually Accordingly, according to embodiments of the present invention, since the entire process is automated without human intervention such as an engineer or a data scientist, a separate human intervention is not required to prevent confusion due to new data and introduce it into the learning model. .

Additionally, AI systems typically do not know what data is missing or what new data is needed to improve accuracy under various conditions, which traditionally requires the intervention of a data scientist or engineer, which is time consuming and costly. it will cost a lot Accordingly, according to embodiments of the present invention, through an automated process, in order to improve the accuracy of the artificial intelligence system, the problem is self-diagnosed without the input of a human (engineer or data scientist), and specific data required to increase the recognition rate is recognized and This can save you money and time.

An object recognition method according to an embodiment of the present invention includes: collecting a first photographed image through a user terminal; predicting a shooting condition of the first captured image; verifying the first photographed image using the predicted photographing condition and adding it to a verification dataset; training an object recognition model using the verification dataset; and using the learned object recognition model, obtaining a recognition result of the object indicated by the second captured image.

The predicting of the shooting condition may include: providing the first captured image to a metadata classification server; and predicting the shooting condition of the first captured image by using a plurality of shooting condition prediction models of the metadata classification server.

The photographing conditions of the first photographed image include a first detailed photographing condition and a second detailed photographing condition, some of the plurality of photographing condition prediction models are used to predict the first detailed photographing condition, and others are It may be used for prediction of the second detailed photographing condition.

The method for recognizing an object may further include training the plurality of photographing condition prediction models using the verification dataset including the predicted photographing condition as a label.

The step of adding the captured image to the verification dataset may include: providing a quiz to a user; verifying the first captured image based on an answer to the quiz; and adding the verified first captured image to the verification dataset.

The providing of the quiz to the user may include: generating a Prioritization Ranking List (PRL) based on an Accuracy Breakdown Report (ABR) indicating validation accuracy; and providing a quiz to the user based on the PRL.

The method for recognizing objects may further include selecting an expert (oracle) from among a plurality of users based on the user's quiz score, and providing the user with the quiz may include: providing the quiz to the expert; may include

The selecting of an expert from among the plurality of users may include: calculating an accuracy score based on a plurality of previous quiz scores and a current quiz score; and selecting a user whose accuracy score falls within a predetermined upper range from among the plurality of users as the expert.

The object recognition method may include: providing competition challenges to a user; and providing a reward to the user based on an answer to the competitive challenge provided from the user.

Providing the user with a competitive challenge may include: generating a PRL based on an ABR indicating verification accuracy; and providing a competitive challenge to the user based on the PRL.

The method for recognizing an object may further include generating an initial object recognition model using a pre-prepared data set.

The object recognition method may further include generating an initial object recognition model without using a pre-prepared dataset.

An object recognition system according to an embodiment of the present invention includes: a metadata classification server for predicting a shooting condition of a first captured image collected through a user terminal; a verification server that verifies the first captured image using the predicted shooting condition and adds it to a verification dataset; an artificial intelligence (AI) learning server for learning an object recognition model using the verification dataset; and a prediction server configured to obtain a recognition result of an object indicated by the second captured image by using the learned object recognition model.

The metadata classification server may predict a shooting condition of the first captured image by using a plurality of shooting condition prediction models.

The photographing conditions of the first photographed image include a first detailed photographing condition and a second detailed photographing condition, some of the plurality of photographing condition prediction models are used to predict the first detailed photographing condition, and others are It may be used for prediction of the second detailed photographing condition.

The AI learning server may train the plurality of shooting condition prediction models by using the verification dataset including the predicted shooting condition as a label.

The verification server may provide a quiz to the user, verify the first photographed image based on an answer to the quiz, and then add the verified first photographed image to the verification dataset.

The object recognition system further includes a verification priority server that generates a Prioritization Ranking List (PRL) based on an Accuracy Breakdown Report (ABR) indicating validation accuracy, wherein the verification server includes: A quiz may be provided to the user.

The verification server may select an expert (oracle) from among a plurality of users based on the user's quiz score, and provide a quiz to the expert.

The verification server calculates an accuracy score based on a plurality of previous quiz scores and a current quiz score, and selects a user whose accuracy score falls within a predetermined upper range among the plurality of users as the expert (oracle) can be selected.

The object recognition system may further include an incentive server that provides competition challenges to the user and provides a reward to the user based on answers to the competition challenges provided by the user.

The object recognition system may further include an incentive priority server that generates a PRL based on an ABR indicating verification accuracy, and the incentive server may provide a competitive challenge to the user based on the PRL.

The AI learning server may generate an initial object recognition model using a pre-prepared dataset.

The AI learning server may generate an initial object recognition model without using a pre-prepared dataset.

A computer-readable medium according to an embodiment of the present invention includes the steps of: collecting, in a computer, a first photographed image through a user terminal; predicting a shooting condition of the first captured image; verifying the first photographed image using the predicted photographing condition and adding it to a verification dataset; training an object recognition model using the verification dataset; and using the learned object recognition model, a program for executing a step of obtaining a recognition result of an object indicated by the second captured image may be recorded.

According to embodiments of the present invention, it is possible to effectively increase the recognition accuracy of a photographed image of a previously unknown or uncertain object without spending a large amount of data and computing power, and thus the recognition quality can be reliably improved. .

1A and 1B are block diagrams illustrating an object recognition system according to an embodiment of the present invention.
2 is a block diagram illustrating an object recognition system according to an embodiment of the present invention.
3A and 3B are flowcharts illustrating an object recognition method according to an embodiment of the present invention.
4 is a block diagram illustrating an object recognition system according to an embodiment of the present invention.
5 is a flowchart illustrating a method for recognizing an object according to an embodiment of the present invention.
6 is a flowchart illustrating a method for recognizing an object according to an embodiment of the present invention.
7A and 7B are block diagrams for explaining an object recognition system according to an embodiment of the present invention.
8 is a flowchart illustrating a method for recognizing an object according to an embodiment of the present invention.
9 is a block diagram illustrating a computing device for implementing an object recognition method and system according to embodiments of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described in detail so that those of ordinary skill in the art can easily implement them with reference to the accompanying drawings. However, the present invention may be implemented in several different forms and is not limited to the embodiments described herein. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

Throughout the specification and claims, when a part "includes" a certain element, it means that other elements may be further included, rather than excluding other elements, unless otherwise stated.

In addition, terms such as "... unit", "... group", and "module" described in the specification may mean a unit capable of processing at least one function or operation described in the present specification, which It may be implemented in hardware or software or a combination of hardware and software.

1A and 1B are block diagrams illustrating an object recognition system according to an embodiment of the present invention.

Referring to FIG. 1A , an object recognition system 1 according to an embodiment of the present invention includes a prediction server 10 , a metadata classification server 12 , an AI learning server 14 , a dataset server 16 , and a mobile It may include a user terminal 18 , 34 in which the application is executed, an incentive priority server 20 , an incentive server 22 , a verification priority server 30 and a verification server 32 .

In this embodiment, the prediction server 10, the metadata classification server 12, the AI learning server 14, the dataset server 16, the incentive priority server 20, and the incentive server 22 shown in FIG. ), the verification priority server 30 and the verification server 32 are only logically separated and do not represent a physical division. That is, the prediction server 10, the metadata classification server 12, the AI learning server 14, the dataset server 16, the incentive priority server 20, the incentive server 22, the verification priority of FIG. 1A Of course, at least some elements of the server 30 and the verification server 32 may be implemented as one entity, and elements referenced by one number in FIG. 1A may be implemented as a plurality of entities.

In addition, in this specification, the element "server" may refer to a hardware device having a processor and memory, such as a computing device as shown in FIG. 9, and software that can be executed on any computing device to provide a service It may refer to itself, and sometimes may refer to a form implemented as a combination of hardware and software.

On the other hand, the user terminals 18 and 34 in which the mobile application is executed may refer to a smartphone, a tablet computer, a notebook computer, etc., but the scope of the present invention is not limited thereto, and the mobile application or the mobile application is not necessarily limited thereto. It may refer to any computing device on which any operating system capable of executing any application is installed.

Now, with reference to FIG. 1A , an object recognition system 1 capable of automatically providing an image of conditions necessary for learning to increase image recognition accuracy and improving the recognition rate by itself to recognize an object from a photographed image under various photographing conditions let me explain

First, the mobile application executed on the user terminal 18 may photograph an object. Specifically, the mobile application may use a camera mounted on the user terminal 18 to photograph an object and collect photographed images. In the present specification, it is assumed that the object is a vehicle for convenience of description, but the scope of the object is not limited to the vehicle.

The captured image captured by the mobile application may be provided to the prediction server 10 , the metadata classification server 12 , or the dataset server 16 .

The prediction server 10 may obtain a recognition result of an object indicated by a photographed image by using the learned object recognition model, and may provide the recognition result to a mobile application. For example, the prediction server 10 may recognize a vehicle indicated by a photographed image, obtain a model and a manufacturer as a recognition result, and display it on the user terminal 18 by providing it to a mobile application. Accordingly, the user of the user terminal 18 can obtain the model and manufacturer information of the vehicle simply by photographing the vehicle.

The AI learning server 14 may train the object recognition model used by the prediction server 10 to recognize an object, and provide the learned object recognition model to the prediction server 10 .

Here, the object recognition model may be a prediction model based on a neural network. Here, the neural network may be implemented as a convolutional neural network (CNN), which is widely used for image recognition, but the scope of the present invention is not limited thereto. The neural network-based AI prediction model has limitations in recognition rate due to the complexity of image information. In other words, the AI prediction model may sometimes make inaccurate judgments, such as not recognizing an object in an image or recognizing it as another object. Accordingly, research on methods to increase the accuracy of artificial intelligence prediction models is active.

In this regard, there are methods to manually add curated data or images to the dataset, methods to automate the collection of large amounts of data, and methods for filtering with human or other artificial intelligence models, but these methods improve the performance of predictive models You will collect a significant amount of data or images that are not very effective for There is a disadvantage that it is inefficient because there are few data or images that can be used.

Also, although the AI prediction model can recognize an object under ideal conditions, when the conditions change, the recognition rate for the same object is often lowered. Therefore, in order to guarantee high accuracy even for various conditions in reality that are not ideal conditions, it is necessary to include a large number of data or images in which various non-ideal conditions (or abnormal conditions) are combined in the dataset.

On the other hand, even if data used in artificial intelligence predictive models can be collected automatically, it is another matter to automatically, accurately and efficiently classify or label new data among the collected data, and usually humans manually classify new data. , label it, or use another AI model. With other AI models, you need to be able to properly classify and label new data by training it to recognize images through a dataset.

In order to solve these problems, the embodiment of the present invention introduces a metadata prediction mechanism, a user verification mechanism, and a user compensation mechanism in addition to the object recognition model used for object recognition in the prediction server 10 . Because of these metadata prediction mechanisms, user validation mechanisms, and user reward mechanisms, engineers or data scientists can There is no need to classify or label the data, or to manually identify problematic images, and ultimately, it is possible to obtain specific data or images that can most effectively improve the object recognition model. Accordingly, the object recognition system 1 may recognize an image that was previously unknown (Zero Accuracy/Unknown) or an image with low accuracy (Low Accuracy).

Accordingly, the object recognition system 1 can accurately recognize data that has been difficult to recognize, and can acquire more training data that can most effectively improve the object recognition model. For example, if a user shoots an image of a blue 2020 Mercedes-Benz CLS on a rainy night at an angle from the rear, a typical prediction system would produce very low-accuracy results because there are too many uncertain factors in object recognition. will provide

These uncertain low-accuracy images are transmitted to the servers 30 and 32 implementing the quiz system for automatic verification by the user (especially, the user who identified the 2020 Mercedes-Benz CLS and was rewarded), and the metadata classification server ( 12) automatically checks the image provided by the user through the shooting condition prediction model, and then identifies important shooting conditions or variables (eg, lighting conditions, shooting angle, color, shooting time, etc.).

That is, the metadata classification server 12 may predict a shooting condition of a captured image by using a plurality of shooting condition prediction models. Here, shooting conditions refer to any conditions (or environment) can be Accordingly, the metadata classification server 12 may predict a shooting condition from the captured image, and the prediction result may be provided to the user through the user terminal 18, but will be used to increase the recognition rate of the object represented by the captured image. may be

The AI learning server 14 trains a plurality of shooting condition prediction models used by the metadata classification server 12 to predict shooting conditions, and sends the learned multiple shooting condition prediction models to the metadata classification server 12 . can provide

The dataset server 16 may include a verified dataset server 160 and an unverified dataset server 162 . The verification dataset server 160 is a server that manages a set of photographed image data in which the object represented by the photographed image is verified, and the unverified dataset server 162 is a set of photographed image data before verification of what the indicated object is. It may be a managing server. In particular, in an embodiment of the present invention, the term verification may mean verification by a user, and the term unverified may mean not verified by the user.

The verification priority server 30 and the verification server 32 implementing the quiz system provide a quiz to the user of the user terminal 34 and perform verification on the photographed image in such a way that the answer is collected, and a competitive challenge system The incentive priority server 20 and the incentive server 22 implementing ? may provide Competition Challenges to the user of the user terminal 18 and provide a reward to the user based on the answer.

The metadata classification server 12 may also be enhanced by a quiz system, and a competitive challenge system may be combined with the quiz data.

The metadata classification server 12 predicts shooting conditions (the rear of the vehicle is photographed at an inclined angle, rainy night, the vehicle color is blue), and to verify the shooting conditions predicted by the metadata classification server 12 It could automatically create a competitive challenge system to show the user that image (i.e. an image of the 2020 Mercedes-Benz CLS taken from the rear of the vehicle at an angle on a rainy night).

In this way, the object recognition system 1 strengthens the dataset by automatically verifying images that were not known at all (Zero Accuracy/Unknown) or images with low accuracy (Low Accuracy), so that in the future, the object recognition system 1 It allows tangible objects to be recognized with high accuracy.

Owing to the excellent mechanism described above, the object recognition system 1 can automatically and quickly recognize new objects for which there was no information before, and use training data for training object recognition models to recognize new or unfamiliar objects. can be obtained automatically. In addition, by performing prediction on metadata, high-complexity images that have lowered the accuracy of artificial intelligence prediction models under abnormal conditions can be recognized with high accuracy, and new data can be automatically entered into the training dataset. can That is, data that has lowered the accuracy of the artificial intelligence prediction model due to high complexity or new objects are ultimately used as training data that can improve the object recognition model most effectively. By self-improving , it is possible to automate the task of increasing the recognition rate for previously unknown (Zero Accuracy/Unknown) or low accuracy (Low Accuracy) images, and even when there is not enough data, the learning model can be can be created, and the recognition rate of the learning model can be increased over time.

On the other hand, the AI learning server 14 provides the learned object recognition model to the prediction server 10, or a method of training the model in order to provide a plurality of learned shooting condition prediction models to the metadata classification server 12 may include a Fine-tuned Bootstrapping (FB) method or a Zero-Lite Bootstrapping (ZLB) method.

The FB method may be a method of using a previously prepared dataset (Traditional Dataset, TD) when the AI learning server 14 first trains the model, as shown in FIG. 1A . That is, a model, that is, an object recognition model and a plurality of shooting condition prediction models, can be trained for the first time with a dataset prepared before the object recognition system is activated. According to the FB method, when the object recognition system is activated, a high initial accuracy can be obtained. That is, when the system is activated, there is a high probability that the first image (or data) provided to the prediction server 10 will be correctly identified.

Alternatively, the ZLB method may be a method that does not use a pre-prepared dataset when the AI learning server 14 trains the model for the first time, as shown in FIG. 1B . That is, in model learning, it refers to a case where there is no initial dataset, and the object recognition model and the plurality of shooting condition prediction models can be learned without using any data prepared before the object recognition system is activated. there is. According to the ZLB method, when the object recognition system is activated, the initial accuracy may be lower than that of the FB method.

The ZLB method may include a Zero Knowledge Bootstrap (ZKB) method and a Partial Knowledge Bootstrap (PKB) method. In the ZKB method, no category of the AI learning server 14 is defined, and an image initially input to the prediction server 10 may be recognized as “unknown”. That is, it can be said that the initial accuracy of the object recognition model of the prediction server 10 is 0. The image recognized as “unknown” is transmitted to the verification server 32 , and a category for the image can be set, ie, classified, by the user through the verification server 32 . The classified and identified data as described above may be provided to the verification dataset server 160 as verified data. And the verified data may be used by the AI learning server 14 to train an object recognition model and a plurality of shooting condition prediction models for the prediction server 10 and the metadata classification server 12 .

In the PKB method, a human (eg, engineer or data scientist) sets the category, and at the time of setting, there may be no data in the corresponding category (ie, only the skeleton of the category can be set). In the PKB method, unlike the ZKB method, the initial accuracy of the object recognition model of the prediction server 10 is not 0, but is still lower than the FB method using a pre-prepared dataset. When the object recognition system is activated, an image (or data) is provided to the prediction server 10 , and there is a very high possibility that the image (or data) will be incorrectly recognized with a very low confidence rate. The image (or data) incorrectly recognized with very low accuracy is transmitted to the verification server 32 , and the image (or data) can be classified by the user through the verification server 32 . The classified and identified data as described above may be provided to the verification dataset server 160 as verified data. And the verified data may be used by the AI learning server 14 to train an object recognition model and a plurality of shooting condition prediction models for the prediction server 10 and the metadata classification server 12 .

Hereinafter, an object recognition system and an object recognition method of the present invention will be described in detail with reference to FIGS. 2 to 8 .

2 is a block diagram illustrating an object recognition system according to an embodiment of the present invention.

Referring to FIG. 2 , the object recognition system 2 according to an embodiment of the present invention includes a prediction server 10 , a metadata classification server 12 , an AI learning server 14 , a dataset server 16 and a mobile It may include a user terminal 18 in which the application is executed.

The mobile application of the user terminal 18 may allow the user to photograph the vehicle. That is, the mobile application may collect the first captured image IMG1 through the user terminal 18 . The mobile application may induce the user to photograph the vehicle, thereby increasing the amount of collection of photographed images.

Since the first captured image IMG1 captured by the user is an unverified captured image, it may be transmitted to the unverified dataset server 162 . The unverified dataset server 162 stores the first captured image IMG1 as an unverified dataset, and provides the incentive priority server 20 and the verification priority server 30 for verification of the dataset. there is.

In addition, the first captured image IMG1 may be transmitted to the prediction server 10 , and the prediction server 10 recognizes the vehicle indicated by the photographed image using the object recognition model M0, and the model and You can get manufacturer information. Here, the object recognition model M0 may have a CNN (Convolutional Neural Network) model structure, but the scope of the present invention is not limited thereto. As a result of the recognition, the model and manufacturer information of the corresponding vehicle may be provided back to the mobile application of the user terminal 18 .

Also, the first captured image IMG1 may be transmitted from the mobile application to the metadata classification server 12 , and the metadata classification server 12 uses a plurality of photographing condition prediction models M1, M2, and M3. After predicting the shooting condition of the first captured image IMG1 , the predicted shooting condition PC may be provided again to the mobile application of the user terminal 18 . That is, the metadata classification server 12 may predict a shooting condition by using a collection of prediction models. Of course, the scope of the present invention is not limited thereto, and the metadata classification server 12 may use only a single predictive model.

Here, the object recognition model M0 may recognize the model and manufacturer of the vehicle indicated by the photographed image IMG1, and the plurality of photographing condition prediction models M1, M2, and M3 may predict the photographing conditions of the photographed image IMG1. can In particular, the photographing condition of the photographed image IMG1 may include several detailed photographing conditions. Detailed shooting conditions are, for example, the shooting time (morning, afternoon, day, night, etc.), weather (sunny, cloudy), shooting angle, light level (bright environment, dark environment), etc. that may affect the shooting image. may include the conditions (or environment) of Some of the plurality of shooting condition prediction models (M1, M2, M3) may be used to predict some of several detailed shooting conditions, and other parts of the plurality of shooting condition prediction models (M1, M2, M3) include: It can be used to predict other parts of several detailed shooting conditions. For example, the shooting condition prediction model M1 may be used to predict the weather at the time of shooting, the shooting condition prediction model M2 may be used to predict the shooting angle, and the shooting condition prediction model M3 may be It can be used to predict the amount of light at the time of shooting.

Meanwhile, the metadata classification server 12 receives the verified photographed image (VD) or the unverified photographed image (UD) from the dataset server 16 , and uses a plurality of photographing condition prediction models M1, M2, and M3. After predicting the shooting conditions for the received captured images (VD, UD), the predicted shooting conditions (PC) may be labeled on the captured images (VD, UD) and transmitted back to the dataset server 16 . That is, the metadata classification server 12 may affect the captured image, such as the shooting time (morning, afternoon, day, night, etc.), weather (sunny, cloudy), shooting angle, and light amount (bright environment, dark environment). After predicting the shooting conditions for the captured images (VD, UD) without information on the shooting conditions including any condition (or environment) in which there is an arbitrary condition (or environment), the prediction result may be labeled. The captured images VD and UD labeled with the shooting conditions may be used as training data for learning the models M0, M1, M2, and M3.

The AI learning server 14 may receive learning datasets for various vehicles. Specifically, the AI learning server 14 may receive the verification dataset (VD) as the training dataset from the dataset server 16 . The AI learning server 14 may train the object recognition model M0 with the input training dataset, and may provide the learned object recognition model M0 to the prediction server 10 .

In particular, the AI learning server 14 may receive a verification dataset (VD) labeled with a shooting condition as a training dataset from the dataset server 16 . The AI learning server 14 may train a plurality of shooting condition prediction models (M1, M2, M3) with a learning dataset in which the input shooting conditions are labeled, and a plurality of shooting condition prediction models (M1, M2, M3) may be provided to the metadata classification server 12 .

Meanwhile, the AI learning server 14 may transmit an Accuracy Breakdown Report (ABR) to the dataset server 16 after completing learning. Here, ABR indicates validation accuracy, and may be provided to the incentive priority server 20 and the verification priority server 30 for verification of the dataset. Here, ABR is defined as the size of the dataset, verification accuracy, user's report indicating that it was misclassified, the vehicle model predicted with low confidence by the AI model, and the difference between the prediction result by the AI model and the expert classification result. may contain information such as vehicle models or conditions with significant discrepancies in .

In various embodiments of the present disclosure, the metadata classification server 12 may make the prediction performance of the object recognition system 2 more accurate and robust in an automated manner.

In order to support a large number of vehicles or objects, and to be able to be used in a huge number of different situations and conditions, the object recognition system 2 can be used in various conditions such as low light, night time, snowy weather, sparse colors, unusual angles, etc. A number of vehicles or objects must be recognized in Such unusual conditions can reduce AI recognition accuracy.

The object recognition system 2 collects a large amount of confusing data that must be classified and labeled before being used to train a new artificial intelligence model, in order to improve AI recognition accuracy. By the way, this task, which can traditionally be done manually, is too time consuming to sort and classify such a large amount of data or images.

The metadata classification server 12 can automatically identify the conditions, label the data appropriately and make it available for future artificial intelligence predictive model training. This classified and labeled data is analyzed and used for the training of new and more accurate AI predictive models, which improve AI recognition accuracy.

Specifically, the accuracy weaknesses of AI systems can only be addressed by identifying conditions in which AI prediction models do not work properly. Therefore, data analysis must be performed. For example, through data analysis, it must be identified that AI recognition works well during the day but not at night, or that AI recognition works well on images of a vehicle from a front angle but not with an image of a vehicle from the rear. , can improve accuracy. Accordingly, the metadata classification server 12 may generate a label for the condition, which is necessary to generate a detailed AI prediction model accuracy report (ie, the ABR generated by the AI learning server 14). This artificial intelligence prediction model accuracy report is used in the verification priority server 30 to prioritize the data to be verified, or to the incentive priority server 20 to prioritize the data to be collected. can be used

When an image of the BMW 3 Series taken from a front angle is given to the AI prediction model during the day, the recognition accuracy will be high, but when an image of the BMW 3 Series shot from a rear angle is given, the recognition accuracy may drop. Accordingly, the metadata classification server 12 predicts what conditions exist in the image or data by using the artificial intelligence prediction model in order to automatically label the data.

Thereafter, the metadata classification server 12 becomes able to identify and classify the photographed images taken from the rear angle at night as a result of the learning, and this allows the incentive priority server 20 to prioritize the data collection (at night). The recognition accuracy of the object recognition system 3 can be increased by specifying the BMW 3 Series filmed from the rear angle) and providing learning data for new artificial intelligence learning.

In particular, the object recognition system works together with the metadata prediction mechanism, user verification mechanism, and user reward mechanism to automatically discover learning data to train the AI prediction model, thereby continuously improving the accuracy and self-functioning. Because it can be extended, it is possible to create a recognition model with only a small amount of data and to improve the recognition accuracy.

3A and 3B are flowcharts illustrating an object recognition method according to an embodiment of the present invention.

Referring to FIG. 3A , the method for recognizing an object according to an embodiment of the present invention includes generating an initial object recognition model using a pre-prepared dataset (S301); Collecting the first captured image (IMG1) through the user terminal 18 (S303); providing the first captured image IMG1 to the metadata classification server 12 (S305); predicting the shooting conditions of the first captured image IMG1 by using the plurality of shooting condition prediction models M1, M2, and M3 of the metadata classification server 12 (S307); verifying the first photographed image IMG1 using the predicted photographing condition PC and adding it to the verification dataset (S309); self-learning the object recognition model (M0) using the verification dataset (S311); and using the learned object recognition model M0 to obtain a recognition result of the object indicated by the second captured image IMG2 ( S313 ).

Next, referring to FIG. 3B , the method for recognizing an object according to an embodiment of the present invention includes generating an initial object recognition model using a ZKB or PKB method without a pre-prepared data set ( S302 ); Collecting the first captured image (IMG1) through the user terminal 18 (S303); providing the first captured image IMG1 to the metadata classification server 12 (S305); predicting the shooting conditions of the first captured image IMG1 by using the plurality of shooting condition prediction models M1, M2, and M3 of the metadata classification server 12 (S307); verifying the first photographed image IMG1 using the predicted photographing condition PC and adding it to the verification dataset (S309); self-learning the object recognition model (M0) using the verification dataset (S311); and using the learned object recognition model M0 to obtain a recognition result of the object indicated by the second captured image IMG2 ( S313 ).

That is, in the object recognition method according to FIG. 3A, model learning is performed according to the FB method described above with respect to FIG. 1A, and in the object recognition method according to FIG. 3B, model learning is performed according to the ZLB method described with reference to FIG. 1B. can

In relation to step S307 , the shooting conditions of the first shooting image IMG1 include the first detailed shooting conditions and the second detailed shooting conditions, and some of the plurality of shooting condition prediction models M1 , M2 , and M3 are It may be used for prediction of the first detailed photographing condition, and the other part may be used for prediction of the second detailed photographing condition.

In addition, the method may further include training the plurality of shooting condition prediction models M1 , M2 , and M3 using a verification dataset including the predicted shooting condition PC as a label.

For more detailed information on the method for recognizing an object according to the present embodiment, reference may be made to the above-described content with reference to FIGS. 1A to 2 , and a description thereof will be omitted for overlapping content.

4 is a block diagram illustrating an object recognition system according to an embodiment of the present invention.

Referring to FIG. 4 , the object recognition system 3 according to an embodiment of the present invention includes a prediction server 10 , a metadata classification server 12 , an AI learning server 14 , a dataset server 16 , It may include a verification priority server 30 , a verification server 32 , and a user terminal 18 , 34 on which a mobile application is executed.

After the user captures the vehicle through the mobile application of the user terminal 18 , the user may transmit the first captured image IMG1 to the dataset server 16 . The dataset server 16, in particular, the unverified dataset server 162 transmits the unverified dataset (UD) to the metadata classification server 12, and the metadata classification server 12 transmits the predicted shooting condition (PC). ) can be labeled on the unverified dataset (UD). In addition, the metadata classification server 12 may transmit the labeled unverified dataset (UD) to the dataset server 16 .

The dataset server 16 , in particular the verification dataset server 160 , may transmit a verification dataset (VD) labeled with shooting conditions to the AI learning server 14 , and the AI learning server 14 may send the labeled verification data The models M0, M1, M2, and M3 may be trained using the set VD. Also, the AI learning server 14 may generate an ABR indicating verification accuracy and then transmit the ABR to the dataset server 16 .

Here, ABR is defined as the size of the dataset, verification accuracy, user's report indicating that it was misclassified, the vehicle model predicted with low confidence by the AI model, and the difference between the prediction result by the AI model and the expert classification result. may contain information such as vehicle models or conditions with significant discrepancies in .

The dataset server 16 may transmit the ABR to the verification priority server 30 . The verification priority server 30 may generate a Prioritization Ranking List (PRL) and transmit the PRL to the verification server 32 . The verification server 32 may provide a quiz QUIZ to the user through the mobile application of the user terminal 34 based on the PRL.

Here, the PRL may include a priority ranking for a vehicle model or a challenge. These rankings are for determining which data to collect in order to maximize the accuracy of the AI model, and for this purpose, various ranking algorithms may be used.

As an example of the ranking algorithm, there is a method of scoring each image taken based on ABR. For example, the priority of the photographed image of the vehicle photographed at night may be higher than that of the photographed image of the vehicle photographed at the front during the day. On the other hand, the priority of the photographed image having little data on the manufacturer and the condition in the dataset server 16 may be higher than that of the photographed image already having a large amount of data on the manufacturer and the condition.

That is, a shot image with a high potential for improvement of the AI model has a higher priority, and a shot image with a low room for improvement of the AI model has a lower priority. To achieve this purpose, the ranking score is based on the amount of training data that already exists, the verification accuracy for vehicle models and manufacturers, user reporting to inform that they are misclassified, the reliability of artificial intelligence predictions, prediction results by artificial intelligence models and experts It may be determined in consideration of whether there is a discrepancy between the classification results of .

Meanwhile, users may verify model and manufacturer information for a given vehicle under various shooting conditions as part of the quiz. For example, users may select a model and make of a vehicle displayed in a photographed image as part of a multiple-choice or multiple-choice question. Of course, the shooting conditions themselves may be presented as a quiz.

Quiz may be generated based on the PRL generated by the verification priority server 30 , and these quizzes are provided to the mobile application of the user terminal 34 . Quizzes can be designed to validate captured images, which in the future can train AI models more effectively.

The user may give an answer (ANS) to the quiz (QUIZ) through the mobile application of the user terminal (34), and data on the quiz (QUIZ) may be provided to the verification server (32) again. The verification server 32 collects data about users' quizzes (QUIZ), verifies the first photographed image therefrom, and then adds the verified first photographed image to the verification dataset, that is, the verification dataset server ( 160) can be stored.

In particular, the quiz may have content for verifying the shooting condition predicted by the metadata classification server 12 , and the verification of the shooting condition may be performed using the data about the quiz, and the verified shooting condition may be By learning the shooting condition prediction model of the metadata classification server 12 using the labeled captured image, the prediction performance of the object recognition system 3 can be improved.

5 is a flowchart illustrating a method for recognizing an object according to an embodiment of the present invention.

Referring to FIG. 5 , a method for recognizing an object according to an embodiment of the present invention includes generating a PRL based on an ABR indicating verification accuracy ( S501 ); providing a quiz to the user based on the PRL (S503); verifying the first captured image IMG1 based on the answer to the quiz (S505); and adding the verified first captured image IMG1 to the verification dataset ( S507 ).

For more detailed information on the method for recognizing an object according to the present embodiment, reference may be made to the above-described content with reference to FIGS. 1A to 4 , and the overlapping description will be omitted.

6 is a flowchart illustrating a method for recognizing an object according to an embodiment of the present invention.

Referring to FIG. 6 , the method for recognizing an object according to an embodiment of the present invention includes calculating an accuracy score based on a plurality of previous quiz scores and a current quiz score ( S601 ); and selecting a user whose accuracy score falls within a predetermined upper range from among a plurality of users as an expert (oracle) (S603); providing a quiz to an expert (S605); verifying the first captured image IMG1 based on the answer to the quiz (S607); and adding the verified first captured image IMG1 to the verification dataset ( S609 ).

For more detailed information on the method for recognizing an object according to the present embodiment, reference may be made to the above-described content with reference to FIGS. 1A to 4 , and the overlapping description will be omitted.

A specific example of an expert system is as follows.

1. The user can scan the real vehicle as described above.

2. A full copy version of the user scan is sent to the quiz repository, a database of newly scanned vehicle images.

3. The metadata classification server creates attributes for each scanned image, such as the vehicle's orientation, color, weather conditions, lighting conditions, landscape, and people in the image.

4. When a new quiz or competitive challenge is created, new vehicle scan images can be randomly selected or the selection can be prioritized according to specific criteria.

You can prioritize new user scans in the following cases:

・When a user reports that the AI classification is incorrect

When there are few or no images for a specific vehicle model in the AI training data source database.

Few or no images, especially for important attributes (night scans, back scans, etc.)

When many experts have classified scans differently compared to AI-aware servers

When the AI car model recognition server flags the scan as having low recognition confidence (for example, when the AI model output signal is below a certain threshold value of 0.5)

When a recognized vehicle model (from an incentive system) or a custom vehicle model is observed to have low accuracy when trained with the AI recognition system, when the AI vehicle model recognition system has difficulty accurately identifying scans for specific conditions (e.g. a scan of a black BMW i8 shot at night)

Prioritized scans are important because they can be used to identify and fix weaknesses in currently deployed AI models.

5. A new vehicle scan may be selected for a quiz or challenge.

6. Selected scans are used to create a quiz or challenge, and you can adjust the ratio of new scans to verified verified scans. That is, Y % of new scans and (100 - Y) % of already validated scans can be mixed and chosen. For example, Y = 20 may mean that the quiz includes scans of 20% new vehicles and 80% are scans of already verified vehicles. The difficulty of the quiz can be adjusted based on the percentage of user accuracy of vehicle model identification obtained in the previous quiz.

7. The user is asked to identify a displayed vehicle scan for a displayed quiz or competitive challenge via the mobile application.

Users are rewarded (in-game rewards, points, prizes, crypto tokens, prizes, etc.) according to their chosen accuracy compared to the validated scan vehicle model classification, and users can receive bonus rewards if previous quizzes are answered with high accuracy. Yes (compensate for consistent accuracy and prevent other users from answering the quiz).

8. For each quiz or competitive challenge completed by the user, an accuracy score is calculated, and the average of the previous and current quiz scores is used to determine the user's score. Users with high scores can accurately identify vehicle model classifications in user scans. An example of score calculation is as follows.

Score = ([Average accuracy of previous M quizzes] - [Standard deviation of previous M quizzes]) * p + ([Accuracy of current quiz] * (1-p))

where M is the number of previous quizzes used in the calculation and p is a percentage. Punishing users with high standard deviations from the accuracy of previous quizzes can either eliminate users who are inconsistent or allow other users to answer the quiz.

9. The user with the highest score is selected as an expert (Oracle). For example, you can select the top 10% of users with the highest scores as experts.

10. For each new scan included in the quiz, the expert's answer is used to determine the vehicle model displayed in the scan. These expert answers are made through several quizzes, and when enough expert users classify a new vehicle scan, an aggregated answer is calculated. For example, if a new vehicle scan has 100 expert answers and 90% of them identify the scan as a BMW X3, the scan will be classified as a BMW X3 and the scan will be marked as validated.

If a consensus cannot be reached among experts, this could mean some kind of manual intervention. Alternatively, the incentive server may focus on prioritizing user scans in expert quiz answers.

11. Successfully classified scans are copied to a validated user scan database (data set server) and incorporated into future quizzes. The validated user scan database (data set server) is periodically shared with the AI system and used to train future car recognition AI models.

7A and 7B are block diagrams for explaining an object recognition system according to an embodiment of the present invention.

Referring to FIG. 7A , the object recognition system 4 according to an embodiment of the present invention includes a prediction server 10, a metadata classification server 12, an AI learning server 14, a dataset server 16, It may include an incentive priority server 20 and an incentive server 22 .

After the user captures the vehicle through the mobile application of the user terminal 18 , the user may transmit the first captured image IMG1 to the dataset server 16 . The dataset server 16, in particular, the unverified dataset server 162 transmits the unverified dataset (UD) to the metadata classification server 12, and the metadata classification server 12 transmits the predicted shooting condition (PC). ) can be labeled on the unverified dataset (UD). In addition, the metadata classification server 12 may transmit the labeled unverified dataset (UD) to the dataset server 16 .

The dataset server 16 , in particular the verification dataset server 160 , may transmit a verification dataset (VD) labeled with shooting conditions to the AI learning server 14 , and the AI learning server 14 may send the labeled verification data The models M0, M1, M2, and M3 may be trained using the set VD. Also, the AI learning server 14 may generate an ABR indicating verification accuracy and then transmit the ABR to the dataset server 16 .

The dataset server 16 may transmit the ABR to the incentive priority server 20 . The incentive priority server 20 may generate a PRL and transmit the PRL to the incentive server 22 . The incentive server 22 may provide a competitive challenge (CC) to the user via the mobile application of the user terminal 18 based on the PRL.

A competitive challenge (CC) may be to allow a user to find a model or make of a vehicle under a specific condition (a specific shooting condition). For example, the user may be prompted to input the model or manufacturer of the vehicle indicated by the photographed image captured under specific conditions. Alternatively, the competitive challenge (CC) may induce the user to photograph a vehicle under a specific condition and input the corresponding condition. In this way, the user can answer the competition challenge (CC) through the mobile application of the user terminal 34, and can provide a reward to the user based on the answer to the competition challenge (CC) provided by the user. there is.

The data collected in a reward giving method to users who have solved or performed such a competitive challenge (CC) may be used to train the models M0, M1, M2, and M3.

In particular, the competitive challenge (CC) may have content for verifying the shooting conditions predicted by the metadata classification server 12, and data collected in a reward giving method to users who have solved or performed the competitive challenge (CC) can be verified, and the prediction performance of the object recognition system 2 is improved by learning the shooting condition prediction model of the metadata classification server 12 using the captured image labeled with the verified shooting condition. can do.

On the other hand, the AI learning server 14 provides the learned object recognition model to the prediction server 10, or a method of training the model in order to provide a plurality of learned shooting condition prediction models to the metadata classification server 12 may include an FB method or a ZLB method, and FIG. 7A shows that a previously prepared dataset is used when the AI learning server 14 first trains a model according to the FB method, and FIG. 7B shows the AI learning server 14 It indicates not to use a pre-prepared dataset when training the model for the first time.

8 is a flowchart illustrating a method for recognizing an object according to an embodiment of the present invention.

Referring to FIG. 8 , a method for recognizing an object according to an embodiment of the present invention includes generating a PRL based on an ABR indicating verification accuracy ( S801 ); providing a competitive challenge to the user based on the PRL (S803); and providing a reward to the user based on the answer to the competitive challenge provided by the user ( S805 ).

For more detailed information on the method for recognizing an object according to the present embodiment, reference may be made to the above-described content with reference to FIGS. 1A to 7B , and a description thereof will be omitted for overlapping content.

To summarize the operation of the object recognition system described so far, the user collects an (unverified) original image through a mobile application, and the original image may be transmitted to the prediction server 10 or transmitted to the dataset server 16 . . The prediction server 10 may provide a recognition result for the original image to the user.

The dataset server 16 may provide the image to the metadata classification server 12 so that classification and labeling are performed by the metadata classification server 12 . The metadata classification server 12 may transmit the newly classified (unverified) and labeled (photographing condition) images back to the dataset server 16 .

The (unverified) classified and labeled images are provided to the verification priority server 30 , and the verification priority server 30 may output a PRL based on the report received from the AI learning server 14 . The PRL of the vehicle model, shooting conditions, training data size, accuracy, etc. may be transmitted from the verification priority server 30 to the verification server 32, and the verification server 32 takes a quiz for the user to verify the data. can create

The quiz data answered by the user is transmitted to the verification server 32, the verification server 32 classifies the data, and provides the verified classification result to the dataset server 16 as "verification data" to be saved.

The verified image is transmitted to the AI learning server 14 to generate a new artificial intelligence prediction model dataset.

On the other hand, using the newly verified dataset, the AI learning server 14 may generate an ABR and transmit it to the dataset server 16 , and the dataset server 16 transmits the ABR to the incentive priority server 20 . may be provided, and the incentive priority server 20 may generate a PRL and provide it to the incentive server 22 . The incentive server 22 may generate a competitive challenge (CC) based on the PRL and provide it to the user through a mobile application.

According to the embodiments of the present invention described so far, it is possible to effectively increase the recognition accuracy of a photographed image of a previously unknown or uncertain object without spending a large amount of data and computing power, so that the recognition quality can be reliably improved. can

In addition, the following effects can be expected.

Automating data collection, data labeling, and data cleaning (traditionally time consuming to do manually)

It takes data that typically confuses AI and produces recognition results with low accuracy and turns it into valuable training data that can help make future AI models much more accurate.

Determining and grouping the accuracy of recognition under various conditions

View user-generated real-world data to automatically identify weaknesses in models themselves (self-diagnosing) (a task traditionally performed by data scientists)

Prioritize data collection that previously confused AI-aware servers, classify that data, and then reuse it to train more accurate AI models (self-updating, self-healing) repairing), self-correcting)

Create competitive challenges or quizzes to encourage users to collect specific data required by the system (based on data identified by the system as being of high need or relevance)

·Automatically check whether the data captured by the user meets our requirements and is useful

Continuously increase the accuracy of AI recognition servers by automatically adding new unknown objects to the system and prioritizing collection of newly discovered objects.

Save time and money by eliminating the need for engineers or data scientists to classify or label images or data, or manually identify problematic images

No need to input the initial data for learning the recognition model

Thus, our object recognition system, coupled with a metadata classification server, works together to automate the most time-consuming aspects of creating AI models (i.e., automating data labeling and cleaning and potentially improving AI models). It can help you specifically target the highest data). In addition, the object recognition system of the present invention can continuously improve the data set to train better AI models and automatically increase recognition accuracy.

When a new vehicle that has never been seen before comes to market, the object recognition system of the present invention has a process to automatically identify the new vehicle model, label the data and integrate it into the AI model of the future. For example, when a new vehicle model is first released in 2022, the new vehicle model will be filmed by a mobile application and added to an unverified (unverified) data set, and then the new vehicle will appear in a quiz.

The quiz has an option to specify that the vehicle model is a new model, and the user can suggest a new vehicle model name. When entering a suggested name, an autocomplete suggestion option is available that lists previously suggested names (by other users) for new vehicles. The user can select one of the autocomplete options. If enough expert (Oracle) users agree on the same model name, that new model name becomes an additional classification label. The incentive server starts including this new vehicle model in the competitive challenge, encouraging users to capture more images of this new vehicle model.

When enough data is collected for a new vehicle model, a new AI model can be trained that includes the new vehicle model classification.

9 is a block diagram illustrating a computing device for implementing an object recognition method and system according to embodiments of the present invention.

Referring to FIG. 9 , an object recognition method and system according to embodiments of the present invention may be implemented using a computing device (or computer) 50 .

The computing device 50 includes at least one of a processor 510 , a memory 530 , a user interface input device 540 , a user interface output device 550 , and a storage device 560 in communication via a bus 520 . can do. Computing device 50 may also include network interface 570 electrically connected to network 40 , such as a wireless network. The network interface 570 may transmit or receive signals with other entities through the network 40 .

The processor 510 may be implemented in various types such as an application processor (AP), a central processing unit (CPU), a graphic processing unit (GPU), and the like, and executes a command stored in the memory 530 or the storage device 560 . It may be any semiconductor device that does The processor 510 may be configured to implement the functions and methods described with reference to FIGS. 1A to 8 .

The memory 530 and the storage device 560 may include various types of volatile or non-volatile storage media. For example, the memory may include a read-only memory (ROM) 531 and a random access memory (RAM) 532 . In an embodiment of the present invention, the memory 530 may be located inside or outside the processor 510 , and the memory 530 may be connected to the processor 510 through various known means.

In addition, the object recognition method and system according to embodiments of the present invention may be implemented as a program or software executed in the computing device 50, and the program or software may be stored in a computer-readable medium.

In addition, the object recognition method and system according to embodiments of the present invention may be implemented as hardware that can be electrically connected to the computing device 50 .

Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto. Various modifications and improvements by those with knowledge also fall within the scope of the present invention.

Claims (25)

collecting a first photographed image through a user terminal;
predicting a shooting condition of the first captured image;
verifying the first photographed image using the predicted photographing condition and adding it to a verification dataset;
training an object recognition model using the verification dataset; and
using the learned object recognition model to obtain a recognition result of an object represented by a second photographed image;
How to recognize objects. According to claim 1,
Predicting the shooting conditions comprises:
providing the first captured image to a metadata classification server; and
and predicting a shooting condition of the first captured image by using a plurality of shooting condition prediction models of the metadata classification server. 3. The method of claim 2,
The shooting condition of the first captured image includes a first detailed shooting condition and a second detailed shooting condition,
and a part of the plurality of shooting condition prediction models is used to predict the first detailed shooting condition, and another part is used to predict the second detailed shooting condition. 3. The method of claim 2,
The method of claim 1, further comprising: learning the plurality of shooting condition prediction models using the verification dataset including the predicted shooting condition as a label. According to claim 1,
The step of adding the photographed image to the verification dataset includes:
providing a quiz to the user;
verifying the first captured image based on an answer to the quiz; and
and adding the verified first captured image to the verification dataset. 6. The method of claim 5,
The step of providing a quiz to the user,
generating a Prioritization Ranking List (PRL) based on an Accuracy Breakdown Report (ABR) indicating validation accuracy; and
and providing a quiz to the user based on the PRL. 6. The method of claim 5,
Based on the user's quiz score, further comprising the step of selecting an expert (oracle) from among a plurality of users,
The step of providing a quiz to the user,
Including the step of providing a quiz to the expert, object recognition method. 8. The method of claim 7,
The step of selecting an expert from among the plurality of users,
calculating an accuracy score based on the plurality of previous quiz scores and the current quiz score; and
and selecting a user whose accuracy score falls within a predetermined upper range from among the plurality of users as the oracle. According to claim 1,
providing users with Competition Challenges; and
The method further comprising the step of providing a reward to the user based on an answer to the competition challenge provided from the user. 10. The method of claim 9,
The step of providing a competitive challenge to the user comprises:
generating a PRL based on the ABR indicating verification accuracy; and
and providing a competitive challenge to the user based on the PRL. According to claim 1,
The object recognition method further comprising the step of generating an initial object recognition model by using a pre-prepared dataset. According to claim 1,
The object recognition method further comprising the step of generating an initial object recognition model without using a pre-prepared dataset. a metadata classification server that predicts a shooting condition of the first captured image collected through the user terminal;
a verification server that verifies the first captured image using the predicted shooting condition and adds it to a verification dataset;
an artificial intelligence (AI) learning server for learning an object recognition model using the verification dataset; and
and a prediction server for obtaining a recognition result of an object indicated by a second captured image by using the learned object recognition model.
object recognition system. 14. The method of claim 13,
The metadata classification server is configured to predict a shooting condition of the first captured image by using a plurality of shooting condition prediction models. 15. The method of claim 14,
The shooting condition of the first captured image includes a first detailed shooting condition and a second detailed shooting condition,
A part of the plurality of shooting condition prediction models is used to predict the first detailed shooting condition, and another part is used to predict the second detailed shooting condition. 15. The method of claim 14,
The AI learning server trains the plurality of shooting condition prediction models by using the verification dataset including the predicted shooting condition as a label. 14. The method of claim 13,
The verification server provides a quiz to the user, verifies the first captured image based on an answer to the quiz, and then adds the verified first captured image to the verification dataset. 18. The method of claim 17,
Further comprising a validation priority server for generating a Prioritization Ranking List (PRL) based on the ABR (Accuracy Breakdown Report) indicating the validation accuracy (validation accuracy),
The verification server provides a quiz to the user based on the PRL. 18. The method of claim 17,
The verification server, based on the user's quiz score, selects an expert (oracle) from among a plurality of users, and provides a quiz to the expert. 20. The method of claim 19,
The verification server calculates an accuracy score based on a plurality of previous quiz scores and a current quiz score, and selects a user whose accuracy score falls within a predetermined upper range among the plurality of users as the expert (oracle), an object recognition system. 14. The method of claim 13,
The system further comprising an incentive server that provides competition challenges to a user, and provides a reward to the user based on answers to the competition challenges provided by the user. 22. The method of claim 21,
Further comprising an incentive priority server for generating a PRL based on the ABR indicating the verification accuracy,
and the incentive server provides a competitive challenge to the user based on the PRL. 14. The method of claim 13,
The AI learning server generates an initial object recognition model using a pre-prepared data set, an object recognition system. 14. The method of claim 13,
The AI learning server generates an initial object recognition model without using a pre-prepared dataset. on the computer,
collecting a first photographed image through a user terminal;
predicting a shooting condition of the first captured image;
verifying the first photographed image by using the predicted photographing condition and adding it to a verification dataset;
training an object recognition model using the verification dataset; and
A computer-readable medium recording a program for executing a step of obtaining a recognition result of an object indicated by a second photographed image by using the learned object recognition model.

Images