Objective Quality Metrics for Ghosting Artifacts in Video and HDR Images
                        Funding: Google Korea (2023-2024) | Affiliation: Pi Lab
                        Tools: PyTorch, OpenCV, scikit-image, NumPy, Pandas
                        
                       - Built a dataset of 2,500+ real images and annotated over 37,000 patches for spatial/temporal artifact detection.
- Designed data pipelines for collection, annotation, and feature extraction from diverse imaging sources.
- Developed a test framework simulating multiple camera noise conditions for robust artifact evaluation.
- Developed a detection model evaluated using subjective and objective quality metrics.
- Trained SOTA deep learning models for domain specific tasks (e.g ghosting artifacts, object detection).
Multi-Scale Attention Model for Low-Light Image Enhancement
                        Affiliation: Pi Lab (2024)
                        Tools: PyTorch, LoL Dataset v1/v2, GCP A100, SSIM, MS-SSIM, LPIPS
                        
                            - Designed a lightweight model (12M parameters) achieving 0.88 SSIM, 0.93 MS-SSIM, 0.207 LPIPS on LoL datasets using GCP A100.
- Worked on integrating attention Mechanisms and transfer learning.
- Implemented an adaptive enhancement pipeline balancing perceptual quality and computational efficiency.
- Reviewed and implemented AAAI, CVPR, and ICCV research papers to optimize model performance.
Traffic Sign Recognition with Advanced Neural Network Techniques
                         Affiliation: Pi Lab (2022-2024)
                         Tools: TensorFlow, OpenCV, Grad-CAM, LIME, GTSRB, ITSD, PTSD
                         
                           - Developed an interpretable CNN (2.6M parameters) achieving 98.4% accuracy and 74.34 ms inference.
- Streamlined ML model development and deployment with MLflow for tracking and reproducibility.
- Optimize GPU acceleration and parallel processing to optimize the system performance.
- Worked on explainable and adversarial robust traffic sign classification to improve model accuracy.
Advanced Medical Imaging Analysis for Disease Detection
                        Affiliation: Pi Lab (2023-Present)
                        Tools: TensorFlow, Keras, PyTorch, Grad-CAM, LIME, OpenCV
                        
                            - Developed lightweight convolutional architecture achieving 99.51% mAP with 17.2 ms inference for brain tumor detection.
- Worked on multi-modal analysis of brain tumors and chest diseases (COVID-19, pneumonia, tuberculosis).
- Designed an adaptive channel attention mechanism with multi-path CNN architecture for medical image classification.
- Implemented explainable AI techniques (Grad-CAM, LIME) reducing false positives by 78% and enhancing diagnostic interpretability.
FireXplainNet: Interpretable Wildfire Detection System
                         Affiliation: Pi Lab (2022-2023)
                         Tools: PyTorch, Grad-CAM, Matplotlib
                         
                            - Developed a lightweight interpretable CNN (5.3M parameters) for early wildfire detection.
- Applied gradient-based attribution (Grad-CAM) for decision explainability in high-risk outdoor scenarios.
- Achieved high accuracy under variable conditions; received Best Paper Award at KCC 2023.
YOLO-Based Real-Time Object Detection and Tracking System
                         Affiliation: Pi Lab (2021-2023)
                         Tools: YOLOv5, YOLO-NAS, DeepSORT, Kalman Filter, ByteTrack, PyTorch, OpenCV
                         
                            - Built a real-time multi-object detection and tracking system using YOLOv5/YOLO-NAS and DeepSORT.
- Achieved mAP > 92% and 30+ FPS on COCO/MOT datasets, enabling real-time visual intelligence.
- Leveraged GPU parallelization strategies to optimize inference performance.