High resolution fingerprint additional features analysis

Abstract

Fingerprint recognition as one of the most-widely used biometric technologies has been extensively studied in the past half of a century. Although it can achieve very high accuracy given fingerprint images of good quality and covering sufficiently large fingerprint areas, its accuracy is still far from being satisfactory when only low quality or small partial fingerprint images are available. For the purpose of further enhancing the fingerprint recognition accuracy, this thesis explores the fingerprint additional features (including pores, dots, and incipient ridges) that can be reliably detected on high resolution fingerprint images. As an important application of fingerprint additional features, a pore-based fingerprint alignment method has been developed for high resolution partial fingerprints. The method utilizes the pores and their surrounding valley structures which are abundant and discriminative on partial fingerprints. Consequently, it achieves much higher alignment accuracy than both minutia-based and orientation field based methods. To more accurately extract additional features from fingerprint images, models of their intensity appearance have been established and novel extraction methods based on the models have been proposed. The proposed models can more accurately describe the additional features by using orientation and scale parameters. The extraction methods apply the established models to fingerprint images for detecting the additional features on them, and employ the automatic scale selection technique to better cope with the varying sizes of the features. For matching the fingerprint additional features, some coarse-to-fine approaches have been designed. They first establish initial correspondences between the features based on their local descriptors and then further check the correspondences to remove false ones by using certain transformation estimation and refinement techniques. The methods can directly align and match additional features on fingerprints, thereby obviating the dependency on other features like minutiae and enabling more effective fusion between these features.To better understand the fusion between fingerprint features, an extensive study has been conducted on fingerprint additional features and minutiae. In addition to different parallel fusion and score normalization methods, hierarchical fusion has been also analyzed by considering different fusion orders and manners. Based on the experiments, a more effective fusion scheme has been presented, and some enlightening conclusions have been arrived at regarding the fingerprint features fusion. Based on the above algorithms proposed in this thesis, a high resolution fingerprint recognition system has been finally developed by using a custom-built optical high resolution fingerprint sensor. As another approach to handling poor quality fingerprint images, a diffusion-based fingerprint image enhancement method has been also proposed for fingerprint image pre-processing. The method can well regularize fingerprint ridge orientation field which is a basis of most fingerprint image analysis methods. Compared with the traditional Gabor-based and other diffusion-based fingerprint image enhancement methods, it can more consistently enhance fingerprint images and better preserve the ridge structures around singular points