On the Semantic Binarization of Fingerprint Images using Support Vector Machines

Abstract

The performance of feature (e.g., minutiae) extraction algorithms and fingerprint recognition techniques rely on the quality of input fingerprint images. In high quality fingerprint images, the valleys and ridges alternate and flow in directions that can be computed and assigned to local regions, making the task of feature extraction easy and accurate. However, in practical fingerprint image recognition and verifications systems, input fingerprint images are far from the ideal. Due to conditions during image capture such as greasy, wet, dry, bruised, damaged, or low quality fingerprints (from manual workers and the elderly), incorrect finger pressure on the capturing device, and sensor noise a sizeable percentage (10 15%) of fingerprint images are of poor quality. In this work, a learning based method using Support Vector Machines (SVMs) is proposed for the binarization of fingerprint images. SVMs are trained on groundtruth images to classify pixels as belonging to ridges or valleys. The identification and classification of individual pixels as background or foreground of the fingerprint image we term semantic binarization. The groundtruth fingerprint images from which SVM training input data is extracted is generated by humans using Grafix, an image processing toolbox developed specifically for pixel level manual processing of images. The performance of SVMs trained on different input vectors is evaluated and compared using images from the FVC2000 Fingerprint Database. Results obtained show error rates less than 0.05% for good training features. The best results are obtained when gray values of pixels are used directly to train SVMs.