NUMISVISION
Computer vision coin detection, marker-controlled watershed segmentation, and monetary valuation.
Suppresses camera sensor noise while maintaining sharp coin edge boundaries.
Evaluates global variance against local neighborhood thresholds under glare.
Calculates Euclidean distance to edge to isolate local maxima seed points.
Marker-controlled watershed successfully splits touching coins with high accuracy.
Operational & Technical Friction
Standard circle detection algorithms (like naive Hough Circle Transforms) fail catastrophically when coins overlap, touch, or exist under uneven ambient lighting with harsh metallic specular glares.
Systemized Implementation
Formulated a robust multi-stage segmentation pipeline using Gaussian blurring, comparative Otsu vs adaptive thresholding, Euclidean distance transforms to identify local maxima centers, and marker-controlled watershed segmentation to partition merged contours with high geometric precision.
How It Operates in Production
Built in Python using OpenCV, NumPy, and SciPy. The pipeline ingests high-resolution images, computes morphological gradients to preserve boundaries, calculates distance transform relief maps, and seeds watershed markers at local Euclidean peaks.
Engineered automatic contour area calibration converting pixel diameters into exact physical coin denominations (10, 25, 50 piastres, and 1 pound), computing total currency value with rigorous error metrics.
Data Flow & Infrastructure
Raw images undergo bilateral edge-preserving smoothing followed by adaptive thresholding. The binary mask passes through distance transformation to isolate core centers. Watershed segmentation floods outward from peaks, separating boundaries cleanly before passing contours to the denomination classifier.