Image Understanding

Coursework projects (CENG 391 / Image Understanding) implementing classical and learning-based computer vision techniques. All implementations are in Python (OpenCV, scikit-learn, scikit-fuzzy) without deep learning frameworks — the focus is on understanding the underlying signal processing and machine learning algorithms.

---

Projects

Directory	Topic	Technique
retinal_blood_vessel_extraction/	Medical image segmentation	FCM + Decision Tree + Gabor (paper reproduction)
people_counting_on_a_mass_site/	Crowd density estimation	Morphological features + MLP (paper reproduction)
ball_tracking_program/	Real-time object tracking	Background subtraction + contour detection
fire_detection/	Video fire detection	Multi-colorspace (YCbCr + HSV) rule fusion

---

Key Results

Retinal Blood Vessel Extraction (DRIVE test set, 2nd manual annotations)

Metric	This Implementation	Paper (Hashemzadeh 2019)	2nd Human
Accuracy	0.9255	0.9531	0.9464
AUC	0.8607	0.9752	—
Sensitivity	0.6760	0.7830	0.7796
Specificity	0.9603	0.9800	0.9717

The AUC gap (0.86 vs 0.975) is attributable to the paper's proprietary Root Guided Decision Tree, which is co-designed with its Gabor feature representation and has no direct sklearn equivalent.

People Counting (Method 3 — best overall)

Metric	Result
MAPE	10.64%
Accuracy within 10%	66.67%
Accuracy within 15%	86.27%

---

Shared Technical Stack

• Python 3.x, OpenCV, NumPy, scikit-learn, scikit-fuzzy, scikit-image, Matplotlib
• Conda environment: image

---

Skills Demonstrated

• Medical image processing: CLAHE, Gabor filterbanks, morphological operators, Kittler-Illingworth optimal thresholding
• Unsupervised learning: Fuzzy C-Means (FCM) clustering
• Supervised classification: Decision Tree, MLP regression
• Multi-colorspace video analysis (RGB, YCbCr, HSV, CIELab)
• Quantitative evaluation: AUC, sensitivity/specificity, MAPE, pixel-level accuracy
• Paper reproduction methodology: understanding the gap between MATLAB and Python implementations