Projects
Some of my the suff I have worked on.
Anomaly Image Generation with Diffusion and Adapter
To address the data scarcity and out-of-distribution issues in anomaly detection, we proposed a novel image generation method combining DreamBooth and Text-to-Image Adapter. By conditioning on anomaly masks, our model enables diffusion-based generation of realistic anomaly images. A reconstruction loss was added to ensure quality and consistency. This method outperformed SOTA models on AUROC, AP, IS, and IC-LPIPS, and was published in IEEE ACCESS.
View Papers →
Industrial Anomaly Image Generation with User-Specified Masks
In collaboration with ANI, we developed a generative model to synthesize large-scale anomaly images from limited industrial datasets. By conditioning on user-specified masks, our model combines StyleGAN with SPADE to generate anomalies that precisely match mask regions. We also implemented an auto-mask generation module using object position priors for consistency and automation. The proposed model achieved a 1.8% improvement in IS and 10% gain in IC-LPIPS over standard GANs, while maintaining performance under limited memory. Applied to real-world datasets such as display defect and dental amalgam (teeth), the model demonstrated strong performance in industrial anomaly detection tasks. A diffusion-based variant is currently under training.
View Sites →
AI Driving Assistant for Beginner Drivers
Developed a vision-based driving assistant system to help beginner drivers better perceive road conditions. Using dashcam footage from Honda Research Institute, we built a YOLO-based object detection model and a custom distance estimation module based on vehicle centroid height. The system distinguishes road/sidewalk regions, detects nearby hazards within 15 meters, and provides voice alerts. A navigation-integrated web app was also developed. The project won the Capstone Design Grand Prize and resulted in a published paper.
Learn more →
AI Navigation Assistant for the Visually Impaired
Designed a real-time object detection and voice guidance system to assist visually impaired individuals in navigating pedestrian environments. Collected custom datasets for non-COCO hazards (e.g., e-scooters, bollards) to improve detection coverage. Implemented a distance estimation algorithm using average object heights and quantified risk levels based on proximity. The system also recognized traffic lights to infer crosswalk presence and issued real-time alerts. The project won a prize at the Ajou University Software Competition and was filed for patent.
Learn more →