KR-102963796-B1 - The Method For Generating Domain-Specific Data Containing Objects Desired By Users And Computing System For Performing The Same

KR102963796B1KR 102963796 B1KR102963796 B1KR 102963796B1KR-102963796-B1

Abstract

The present invention relates to a method for generating domain-specific data containing an object desired by a user and a computing system for performing the same. More specifically, the invention relates to a technology for generating images corresponding to situations or objects that are difficult to naturally obtain during the process of generating training data for training an artificial intelligence model. This technology involves inputting a domain image and a domain prompt, which serve as the basis for the specialized data, into a generative model to generate the domain-specific data, and performing automatic verification thereof, thereby enabling the efficient generation of domain-specific data desired by a user.

Inventors

이건희
윤동식
이예진
정찬성
김종은
이다연
오중균
박종민

Assignees

에이치디씨랩스 주식회사

Dates

Publication Date: 20260511
Application Date: 20250721

Claims (7)

A method for generating domain-specific data containing a user-desired object, performed in a computing system comprising one or more memories and one or more processors, wherein A prompt input step in which a domain prompt is received from a user terminal, wherein the domain prompt includes a command related to the creation or modification of a target object desired by the user; A generative model input step for inputting a domain image that forms the basis of specialized data including the above target object and the above domain prompt into a generative model; A result image output step for outputting a result image corresponding to the domain prompt and the domain image using the generative model above; A data evaluation step for evaluating whether the above result image corresponds to specialized data including the above target object; and A specialized data determination step for determining the result image as specialized data when the result image has an evaluation result greater than or equal to a preset evaluation criterion; The above generative model is, After performing masking at multiple random locations on the input domain image, the location among the multiple masked areas with the highest probability of the target object being located is selected and reconstructed to correspond to the domain prompt, and The above specialized data determination step is, In the above data evaluation step, a step of determining whether the first evaluation score derived as a result of performing the first evaluation step using a detector exceeds a preset first evaluation criterion; In the above data evaluation step, a step of determining whether the second evaluation score derived as a result of performing the second evaluation step using an artificial intelligence-based language model exceeds a pre-set second evaluation criterion; and If it is determined that the first evaluation score exceeds the first evaluation criterion and the second evaluation score exceeds the second evaluation criterion, the method includes the step of determining the result image as the specialized data. The above language model is an artificial intelligence-based model capable of simultaneously processing image data and natural language data in text form, and The detector above is a model that receives the result image and derives location information of a target object on the result image, and A method for generating domain-specific data, wherein data output from each of the language model and detector, which are different models, is scored to inspect the result image, and the result image is evaluated to determine whether it corresponds to specialized data including the target object.
In claim 1, The above data evaluation step is, A first evaluation step of evaluating the result image using a detector; comprising, The above first evaluation stage is, An object location determination step of determining whether the target object is detected on the result image using the detector, and if the target object is detected, deriving first location information of the target object on the result image; A first similarity calculation step for calculating the similarity between the first position information and the second position information of a masked area masked by the generative model by comparing the first position information and the second position information; and A method for generating domain-specific data, comprising: a first evaluation score derivation step for deriving a first evaluation score based on the first similarity above.
In claim 2, The above detector is, It corresponds to a global detector that detects one or more objects included in an input image, and The above object location determination step is, A method for generating domain-specific data, comprising a first position information determination step of comparing position information of each of one or more objects detected by the detector; and second position information of the masking area; and specifying the position information having the highest similarity to the second position information as the first position information of the target object.
In claim 1, The above data evaluation step is, A second evaluation step for evaluating the result image using an artificial intelligence-based language model capable of simultaneously processing image data and text-based natural language data; is included. The above second evaluation stage is, A step of inputting the above result image into the above language model to obtain caption information in the form of text describing the above result image; and A method for generating domain-specific data, comprising the step of comparing the above caption information and the above domain prompt to derive a second evaluation score based on the similarity between the above caption information and the above domain prompt.
delete
In claim 1, The above method for generating domain-specific data is, It further includes a user verification step for verifying whether the specialized data determined in the specialized data determination step corresponds to the specialized data desired by the user; and The above user verification step is, A step of transmitting a popup UI, which allows the user to select whether to use the specialized data determined in the specialized data determination step, to the user terminal along with the specialized data; and A method for generating domain specialized data, comprising the step of determining the specialized data as the final specialized data when receiving input from the above user terminal indicating that the specialized data will be used.
A computing system comprising one or more memories and one or more processors, and performing a method for generating domain-specific data including objects desired by a user, A prompt input unit that receives a domain prompt from a user terminal, wherein the domain prompt includes a command related to the creation or modification of a target object desired by the user; A generative model input unit that inputs a domain image serving as the basis for specialized data including the above target object and the above domain prompt into a generative model; A result image output unit that outputs a result image corresponding to the domain prompt and the domain image using the generative model above; A data evaluation unit that evaluates whether the above result image corresponds to specialized data including the above target object; and It includes a specialized data determination unit that determines the result image as specialized data when the result image has an evaluation result greater than or equal to a preset evaluation standard. The above generative model is, After performing masking at multiple random locations on the input domain image, the location among the multiple masked areas with the highest probability of the target object being located is selected and reconstructed to correspond to the domain prompt, and The above-mentioned specialized data determination unit is, A component that performs a step of determining whether a first evaluation score derived as a result of performing a first evaluation step using a detector in the above data evaluation unit exceeds a preset first evaluation criterion; A component that performs a step in the above data evaluation unit of determining whether a second evaluation score derived as a result of performing a second evaluation stage using an artificial intelligence-based language model exceeds a pre-set second evaluation criterion; and A component that performs the step of determining the result image as the specialized data when it is determined that the first evaluation score exceeds the first evaluation criterion and the second evaluation score exceeds the second evaluation criterion; The above language model is an artificial intelligence-based model capable of simultaneously processing image data and natural language data in text form, and The detector above is a model that receives the result image and derives location information of a target object on the result image, and A computing system that scores data output from each of the language model and detector, which are different models, inspects the result image, and evaluates whether the result image corresponds to specialized data including the target object.

Description

The Method For Generating Domain-Specific Data Containing Objects Desired by Users and Computing System For Performing The Same The present invention relates to a method for generating domain-specific data containing an object desired by a user and a computing system for performing the same. More specifically, the invention relates to a technology for generating images corresponding to situations or objects that are difficult to naturally obtain during the process of generating training data for training an artificial intelligence model. This technology involves inputting a domain image and a domain prompt, which serve as the basis for the specialized data, into a generative model to generate the domain-specific data, and performing automatic verification thereof, thereby enabling the efficient generation of domain-specific data desired by a user. Artificial intelligence (AI) technology, particularly deep learning-based models, plays a key role in various industrial fields such as image recognition, autonomous driving, medical image analysis, and smart factories. The performance of these AI models is critically determined by the quantity and quality of training data, and large-scale, high-quality training data that accurately reflects the characteristics of a specific domain is essential for AI models to effectively solve problems within that domain. However, securing sufficient high-quality training data in real-world environments presents several challenges. First, data collection itself requires a massive amount of time and cost. In particular, various accident scenarios for autonomous driving simulations, data on rare diseases in medical imaging, and defect data for detecting specific product flaws are often very difficult to acquire naturally or are nearly impossible to collect due to ethical or safety concerns. Second, as regulations on personal information protection and data security are strengthened, there are significant restrictions on collecting and utilizing actual data. For example, since images containing individuals' faces or medical information must be managed under strict regulations, it is difficult to freely use them for artificial intelligence training. Synthetic data generation technology is gaining attention as an alternative to address these problems. Synthetic data is virtual data artificially created using generative models that have learned statistical characteristics based on real data. Representative generative models include Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), and these models have been utilized to augment the amount of training data by generating new data with distributions similar to real data. However, conventional synthetic data generation technologies have several limitations. General generative models tend to generate random results by learning the overall distribution of data, making it difficult to generate specific objects or situations with precise control intended by the user. Furthermore, the process of verifying whether the generated data is valid for actual AI model training—that is, whether it accurately reflects the characteristics of the domain—is inefficient. Since most generated results must be manually selected by humans or verified through separate evaluation models, this hinders the complete automation of the data generation process and becomes a factor that incurs additional time and costs. Therefore, to improve the performance of artificial intelligence models, there is a growing need for new technologies that can go beyond simple data augmentation to efficiently generate high-quality, domain-specific data that explicitly includes specific objects or situations desired by users, and automatically validate its validity to automate the entire data generation process. FIG. 1 schematically illustrates the internal configuration of a computing system in which a domain-specific data generation method according to one embodiment of the present invention is performed. FIG. 2 schematically illustrates the process of performing the prompt input step and the generative model input step according to one embodiment of the present invention. FIG. 3 schematically illustrates a domain image and a result image according to an embodiment of the present invention. FIG. 4 schematically illustrates the process of a generative model according to one embodiment of the present invention generating a result image. FIGS. 5 and 6 schematically illustrate the process of performing a first evaluation step according to an embodiment of the present invention. FIG. 7 schematically illustrates the process of performing a second evaluation step according to one embodiment of the present invention. FIG. 8 schematically illustrates the process of performing a specialized data determination step according to one embodiment of the present invention. FIG. 9 schematically illustrates the steps for performing a user verification step according to one embo