Counting Turkish Coins with a Calibrated Camera

Abstract

We present a computer vision application that detects all coins in a test image, classifies each detected coin and computes the total amount. Coins to be counted are assumed to be lying on a flat surface. The application starts by estimating the extrinsic parameters of the input camera relative to this flat surface ($[\mathbf R \,|\,t]$), whose intrinsic parameters ($\mathbf K$) are assumed to be known beforehand. Then, a bilateral filter is applied to the image to remove textural details and noisy artifacts. Circles in the filtered image are detected and smaller concentric circles are eliminated. Finally, the geometric parameters (the center and the diameter) of the remaining circles are computed by back-projecting the reciprocal points from the circle contours using the estimated camera parameters. Having thus computed the diameter of each detected coin, the classification is performed by comparing the computed diameter with the actual coin diameters. The experiments performed with a dataset consisting of 50 images containing different combinations of Turkish coins show that the proposed method achieves 98% accuracy rate and works even when some coins are partially occluded, as the method does not use any texture information.