|
|
|
“Advanced Techniques for Latent Fingerprint Enhancement and Recognition”
by Jian Li
May 2017
Fingerprints provide one of the most popular biometric data, and have been widely used in individual person identification and verification. The Automated Fingerprint Identifi- cation System (AFIS) offers important evidences for criminal investigation, and serves as an important tool for law enforcement. As compared with conventional exemplar fingerprints, latent fingerprints are typically collected in a crime scene. They are often degraded and corrupted, leading to very low identification rates. In a practical system, a latent fingerprint has to be enhanced prior to feature extraction to ensure a reliable finger- print matching performance. In this research, we study techniques for latent fingerprint enhancement and orientation field estimation to achieve a higher matching rate. Our studies include traditional image processing techniques as well as a new method based on the emerging convolutional neural network (CNN). Several major contributions are detailed below. In Chapter 3, starting from traditional image processing basis, we propose a new method using the Markov random field (MRF) model and the sparse representation (SR) of ridges to enhance latent fingerprint. The proposed MRF-SR method is inspired by the recent success of dictionary-based methodologies (including both the orientation field dictionary and the ridge dictionary). The idea is detailed below. First, given a set of training local fingerprint patches, we obtain the over-complete sparse dictionary to form a pool of ridge patch candidates. Second, the texture component of latent fingerprints is extracted using two total variation (TV) models: namely, the adaptive-directional TV (ADTV) model and the TV with the L1 fidelity regularization (TV-L1) model. Third, we assess local image quality using the structure similarity (SSIM) index to determine the proper total variation model to use for a local fingerprint patch in the next optimization step. Fourth, we define an MRF model with an unary potential and a neighbor interaction potential and use it to select the optimal patch candidates for the enhancement of a given patch. As compared with existing fingerprint enhancement techniques based on the ridge dictionary for denoting, the proposed MRF-SR method offers a better scheme for latent fingerprint enhancement with sparse optimization. In Chapter 4, we show that the MRF-SR method can also be used to extract the orientation field. To begin with, we generalize several well-known traditional orientation estimation algorithms to the context of latent fingerprints. Then, we include the orientation field estimation technique by adding the orientation unary potential based on fingerprint pose to the cost function of the MRF formulation, which is MRF-SR- Modified. It provides a valuable supplementary tool to other orientation estimation algorithms in the literature. Finally, a fusion technique is adopted to boost the overall performance of latent fingerprint enhancement. As an essential feature of fingerprints, the orientation field can enhance a fingerprint image with directional and contextual filtering. Experimental results on orientation field estimation as well as latent fingerprint enhancement are given to demonstrate the effectiveness and robustness of the proposed fusion methodology. In Chapter 5, we explore the feasibility and study the performance of applying the CNN to latent fingerprint enhancement. Being motivated by recent developments of CNN in image enhancement and restoration applications, we propose a novel encoding- decoding neural network, called the FingerNet. The FingerNet is trained in a pixelwise end-to-end manner for direct fingerprint enhancement. In particular, we develop a novel