ICACT20250037 Slide.21
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
If you have any questions or are interested in collaboration, please feel free to contact me at moonstarry@etri.re.kr.
|
|
ICACT20250037 Slide.20
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Thank you for your attention.
|
|
ICACT20250037 Slide.19
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
|
|
ICACT20250037 Slide.18
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
This research was supported by the Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency (KOCCA), funded by the Ministry of Culture, Sports, and Tourism (MCST).
|
|
ICACT20250037 Slide.17
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
|
|
ICACT20250037 Slide.16
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
In conclusion, we developed a system that generates and annotates synthetic Korean paintings for AI training. Our results demonstrate the potential of this system to create usable datasets, as validated by expert evaluations and technical metrics.
For future work, we plan to extend this system to other artistic styles and cultural domains, contributing to AI-based cultural research and heritage preservation.
|
|
ICACT20250037 Slide.15
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
|
|
ICACT20250037 Slide.14
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Despite the promising results, some failure cases were observed. For instance, incorrect text-to-image attributions occurred when objects were either blended into the background or not generated distinctly. These issues highlight areas for further improvement.
|
|
ICACT20250037 Slide.13
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
|
|
ICACT20250037 Slide.12
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Additionally, a qualitative evaluation was conducted through a survey involving nine cultural experts. The results were encouraging, with the experts giving an overall satisfaction score of 4.1 out of 5 for the quality of the paintings and object labeling.
|
|
ICACT20250037 Slide.11
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Our experimental phase involved creating a dataset of 256 synthetic Korean paintings and using it to train an object detection AI model. The results showed that the model achieved an AP50 score of 0.322 in detecting objects like butterflies, even though it was trained exclusively on synthetic data.
|
|
ICACT20250037 Slide.10
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
|
|
ICACT20250037 Slide.09
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
By refining the DAAM heat maps, we developed an automated labeling mechanism. This innovation streamlines the annotation process, enabling efficient generation of labeled datasets for AI training in niche artistic domains.
|
|
ICACT20250037 Slide.08
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Hypernetworks: In contrast, these maintained stronger attributions, allowing for more accurate object labeling in our generated Korean paintings.
|
|
ICACT20250037 Slide.07
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Textual Inversion: While it supports style-specific generation, it sometimes weakens text-image correlations.
|
|
ICACT20250037 Slide.06
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
One of the most critical tasks in our pipeline is accurate object labeling. For this, we employed the DAAM technique, which utilizes cross-attention to visualize text-image attribution as heat maps.
These heat maps identify which parts of an image correspond to specific text prompts, assisting in object annotation.
|
|
ICACT20250037 Slide.05
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
The core challenge in our study was generating high-quality Korean paintings using text-to-image models like Stable Diffusion. These models typically require large datasets, which are scarce in niche domains like traditional Korean art.
To address this, we experimented with techniques such as textual inversion and hypernetworks, which allow us to work with small datasets while maintaining the distinct style and quality of the generated images.
|
|
ICACT20250037 Slide.04
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
|
|
ICACT20250037 Slide.03
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Let me begin by introducing the context and motivation for this study.
While generative AI has achieved remarkable success in creating high-quality images, generating content in niche domains, such as traditional Korean painting, remains a challenge due to limited data availability.
Our research leverages advanced techniques like textual inversion and hypernetworks to generate synthetic Korean paintings while automatically labeling the objects within these images. The effectiveness of our approach was assessed through expert evaluations and performance testing on object detection tasks.
|
|
ICACT20250037 Slide.02
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
|
|
ICACT20250037 Slide.01
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Thank you for joining this session. My name is SungWon Moon, a Senior Researcher at ETRI, and I am joined by my colleagues Jiwon Lee and Dowon Nam. Today, we are pleased to present our research titled: "Developing an Automated Synthetic Korean Painting Dataset Generation System for AI Training."
This presentation will explore how generative AI techniques can address challenges in creating and annotating niche-domain datasets, particularly for traditional Korean paintings.
|
