Latent Fingerprint Segmentation And Enhancement Based On Sparse Representation

Latent fingerprints are the finger skin impressions left at the crime scene by accident.Compared with other fingerprints,latent fingerprints are usually of poor quality with unclear ridge structure and various overlapping patterns.Therefore,segmentation is necessary to be performed on latent fingerprints before matching,which segments the valid fingerprint regions as foreground from non-fingerprint background.After that,enhancement is performed to reduce the noises,restore the corrupted ridge and valley and improve the fingerprint quality.Latent fingerprint segmentation,which segments the valid fingerprint region from complex backgrounds,is an important preprocessing step for latent fingerprint recognition system.This thesis proposes a latent fingerprint segmentation algorithm based on sparse representation.First,we use total variation model to decompose a latent fingerprint into cartoon and texture components.The cartoon component is mainly composed of non-fingerprint patterns such as digit number and handwriting while the texture component is composed of valid fingerprint texture such as edge and valley.The texture component is used for further processing but the cartoon component will be discarded as noise.Second,we construct a dictionary with a set of Gabor functions and compute the sparse representation of the texture image against the dictionary.A quality map is computed based on a sparsity measure of representation coefficients.Finally,thresholding and morphology processing are applied on the quality map to produce the final segmentation result.After segmentation,the valid fingerprint region is still corrupted by various noises resulting in broken ridges and artifacts.To address this problem,the thesis proposes a latent fingerprint enhancement algorithm based on orientation guided sparse representation.First,a local ridge orientation field is computed for each fingerprint and a reliability is calculated based on the orientation consistency.Second,for the fingerprint region of low reliability,a sparse representation is computed on the global Gabor dictionary and the local fingerprint is reconstructed with linear combination of the sparse coefficients and dictionary elements.Finally,we compute the local orientation of the reconstructed image.If the orientation has large deviation from the initial one,a local dictionary is constructed using a set of Gabor functions with orientation specified by the initial one and the local fingerprint is reconstructed with the local dictionary and spare coefficients.In the experiments,we test the proposed segmentation and enhancement algorithms on the NIST SD27 latent fingerprint database.Experimental results indicate that the proposed algorithms can not only segment the fingerprint region from complex backgrounds but also restore the corrupted ridge structure and improve the fingerprint quality.