Coin Detection & Classification

The Coin Detection Challenge from EPFL's Image Analysis and Pattern Recognition course, tackled in a team of three: detect and classify every coin in photos taken over neutral, hand-held and noisy backgrounds, each of which breaks a naive segmenter in a different way.

Rather than one generic segmenter, the pipeline branches per background, selected by a Shannon-entropy measure of local texture. Neutral backgrounds use Canny edge detection followed by morphological closing, hole filling and erosion. Hand backgrounds add rolling-ball background subtraction before Canny and localize coins with a circular Hough transform, which votes in (x, y, r) accumulator space for circle centers and radii. Noisy backgrounds rely on thresholding, median filtering, erosion and region expansion.

Segmented coins are labelled, cropped into patches and augmented by rotation and translation to build rotation-invariant training data. Classification was tried two ways: a cosine-similarity classifier on patch features, and fine-tuned CNNs (ResNet-18/50/152), with the experiments highlighting the accuracy-versus-training-cost trade-off that motivated moving training onto GPU.