COIN-O-MATIC A fast and reliable system for automatic coin classification Laurens van der MaatenPaul Boon
Introduction Existing systems for coin classification not suitable for heterogeneous coin collections Heterogeneous coin collections: –MUSCLE CIS benchmark dataset –Datasets with historical coins Classification of heterogeneous coin collections requires incorporation of visual features
Introduction MUSCLE CIS benchmark: –692 coin classes with 2,070 coin faces –5,000 coins should be processed within 8 hours –misclassifications have a high penalty –unknown coins in test set –1 GB memory and ~20,000 training coins COIN-O-MATIC was developed with these properties in mind
Sample coins
The system Roughly, consisting of 4 subsystems: –Segmentation –Feature extraction –Classification –Verification
The system
Segmentation Two-stage approach –Fast approach for ‘easy’ cases –Computationally more intensive approach for ‘difficult’ cases
Segmentation Easy cases –Thresholding with t i =60 to remove background –Sobel edge detection with dynamic threshold –Morphological operations –Assume upperleft pixel to be background; perform bucket fill operation –Check whether segmentation was successful
Segmentation Difficult cases are the cases in which a failure of the previous method was detected –Severe blurring of the images removes background structure –Sobel edge detection with dynamic threshold –Idem Two-stage approach successful for 95% of the coin images
Feature extraction Edge-based statistical features –Measure statistical distributions in edge map of the coin Three features –Edge distance distributions –Edge angle distributions –Edge angle-distance distributions Latter feature used in final system
Feature extraction Median filtering and contrast stretching Edge maps are obtained by applying a Sobel edge detection with non-maxima suppression and a dynamical threshold (using Otsu’s method) The borders of the coin are ignored, since they are not discriminative
Edge distance distributions Estimate the distribution of the distances of edge pixels to the center of the coin Rotation invariant feature Can be measured on coarse-to-fine-scales
Edge angle distributions Measure distribution of angles of edge pixels w.r.t. the baseline Not rotation invariant by definition (however, the magnitude of the Fourier transform is) Can be measured on number of fine scales
Edge angle-distance distr. Incorporate both angular and distance information in the coin stamp We measure EADD using 2, 4, 8, and 16 distance bins and 180 angular bins Resulting features are 5200-dimensional
Classification Area preselection (7% margin, measured from image) Thickness preselection (25% margin) Both coin sides are classified seperately using a 5-nearest neighbour classifier
Classification If classifications are equal –Accept this classification (no verification) If not –Perform ranking procedure considering 15 nearest neighbours –Perform verification of classification
Verification Employs averaged prototypes in MUSCLE CIS dataset Coin images are converted to polar space Blurred intensity gradients are computed For the prototypes, this is already done off- line
Verification
Mutual information of coin images with all corresponding prototypes computed (for all circular shifts of the prototypes) Maximum MI assumed to be correct Sum of MI-values serves as rejection value
Implementation In Visual C Employs IPP-library for image processing and Fourier procedures
Results
Misclassifications usually caused by coins that lack contrast for successful edge- detection Application of PCA speeds up the system, however, slight reduction in performance (possibly due to use of Simple PCA)
Recommendations Improved segmentation procedure Speed can be improved by applying LAESA in the 5NN-classifier Reliability can be be improved by always applying verification procedure, and by incorporation of rotation information Classification performance can be improved by applying edge-enhancing filters
Questions