Project Overview
Participaed in a project with Daewoong Foundation to develop a deep learning system for recognizing skin diseases in pets, aiming to assist veterinarians in diagnosis.
Key Contributions
- Implemented dataset parser for complex data structures
- Used U-Net based image segmentation model using PyTorch
- Collaborated on pipeline combining object detection and classification
Technical Highlights
Dataset Handling:
- Parsed and cleaned ~55,000 images with 6 lesion types
- Converted annotations to YOLO format
U-Net Segmentation Model:
- Addressed challenges of small object segmentation (93% of lesions <5% of image area)
Integrated Pipeline:
- Combined segmentation, object detection, and classification
- Implemented post-processing for improved accuracy
Challenges Overcome
- Inconsistent image quality and small object sizes
- Complex dataset structure with missing data
- Model overfitting in semantic segmentation
Tools Used
Python, PyTorch, Jupyter Notebooks, Git/GitHub