Role

Core team member responsible for algorithm implementation, parameter tuning, and evaluation design

Course

EECE 5626 – Image Processing & Pattern Recognition

Background

Image segmentation is a fundamental problem in computer vision. Compared to deep learning approaches, classical or weakly supervised segmentation methods remain relevant in scenarios with limited or no labeled data. This project focuses on systematically evaluating classical segmentation techniques on natural images.

Objective

• Implement and compare multiple classical image segmentation algorithms
• Analyze strengths and weaknesses of each algorithm under unified dataset and evaluation standards
• Study the impact of parameter settings on segmentation quality and stability
• Build a reusable segmentation evaluation and visualization pipeline

Technologies

Python image processing (NumPy, OpenCV, scikit-image), Classical image segmentation algorithm implementation, Graph theory & clustering methods, Parameter tuning & grid search, Segmentation evaluation metrics & visualization

My Contributions

• Implemented multiple classical segmentation algorithms, including Otsu thresholding, K-Means clustering, Active Contours, Watershed, and graph-based segmentation
• Designed parameter tuning strategies for each method
• Built a unified evaluation pipeline for segmentation results
• Evaluated performance using precision, recall, F-score, and boundary-based metrics
• Conducted experiments on the BSDS500 dataset with qualitative and quantitative analysis

Results

• Successfully implemented and compared multiple classical segmentation methods
• Analyzed trade-offs in boundary accuracy, region consistency, and computational cost
• Demonstrated the sensitivity of segmentation quality to parameter choices
• Delivered a reusable evaluation and visualization framework

Project Value

• Demonstrates systematic understanding of classical computer vision methods
• Emphasizes algorithm principles, parameter tuning, and evaluation methods rather than black-box models
• Reflects complete research workflow from algorithm implementation to experimental analysis

GitHub Repository

Available upon request