Selected Student Projects in ML&DL
Explaining characteristics of Korean Genre Paintings (PungSok-Hwa: 풍속화)
How to help art experts to infer artists of unsigned or unidentified paintings by extracting characteristics of specific artists or eras? This project established a vision to apply deep learning techniques to classify and style transfer the styles of images. Focusing on paintings from two significant artists, Kim Hong-do and Sin Yun-bok, in the late Joseon Dynasty, the project team utilized CNN to learn the feature of paintings and classified paintings by artists. Also, the team used LIME to arrive at explanations that are locally reliable components of the image.
Presentation File
EO Gyeong Seon_Presentation_Team4_See-Saw_Explaining characteristics of Korean Genre Paintings.pdf의 사본Practical Tree Pruning using Genetic Algorithm
Kwangyeon Gill, Jooeun Kim, Heuisu Kim, Joongkyu Lee, Gyeongje Cho, Yourim Choi
Pruning is a necessary process for tree maintenance. Good tree architecture will provide landscaping aesthetics to the community, accelerate tree healing with solid branches, prevent disease/pests, and nourish fruit harvests. However, tree pruning is difficult and expensive. As a complex process, it demands domain expertise, and the job itself is dangerous. This project aimed to use machine learning as a solution (Visual Simulation using GAN) to generate an optimized practical tree pruning guide. Also, the project team took a multifaceted approach for both visual and quantitative evaluations.
Presentation File
GILL Kwangyeon_Presentation_Team6_SixthSensor_Practical Tree Pruning using Genetic Algorithm.pdf의 사본GANbok: Create New Fashion Designs with Hanbok
Seongsu Ha, Yoon-sung Lee, Seoyoung Lee, Juyeon Park, Sooyoun Park, Sungwook Son
With the global popularity of the K-aesthetic values, how can we diversify and globalize the traditional Korean clothes (Hanbok) and change its limited and similar designs via technical approaches such as style transfer, GAN, and feature engineering?
Bearing the question, “can machines understand human aesthetic values? In other words, can machines find patterns within them? “ The project team created multiple new fashion designs with traditional hanbok images based on explainable features. Many findings included but were not limited to: LIME captured the unique shapes on the short jacket (Jeogori) such as collar and tie in hanbok. In addition, the project team generated new mixed styles by adding the effects of one image’s style to an encoded vector space as a background.
Presentation File
Team7_GANbok_final_pres.pdf의 사본Handwriting Analysis of the King of the Joseon Dynasty Using Deep Learning
—Based on the 列聖御筆帖 (Handwriting of Kings)
ByeongChan Kim, Chaeyoung Chung, Dong KyuCho, Eunji Park, Minsoo Cho, Jewon Kang
The 열성어필 (Handwriting of Kings) is an image dataset containing the Medieval calligraphy of twenty-seven Kings over 518 years of the Joseon Dynasty’s reign. It shows a glimpse of the dynasty’s forgotten glamour. Working on a style recognition of medieval Korean Calligraphy using a few-shot font generation model, the project team built a font generator. The generator would reproduce the font style of a specific Korean King using a style transfer model named MXfont. For instance, taejo was known for his bold strokes, while sukjong’s strokes were uniform in size and shape.
Also, through this project, we can imagine broader applications of the font generator such as translation and conversion of the original writing to the target language, physically impaired assistance, or creating samples for the field of forensic pathology.
Presentation File
Presentation_Team8_TCK_Handwriting Analysis of the King of the Joseon Dynasty Using Deep Learning.pdf의 사본