Latent Fingerprint Image Enhancement using Gabor Functions via Multi-Scale Patch based Sparse Representation and Matching based on Neural Networks

Abstract

Fingerprint identification is one of the most trusted identification techniques acceptable by court of law. Latent fingerprint images are usually smudged, distorted, overlapped by other prints with less clarity and less content of poor quality. Hence it is challenging to achieve well definitive latent fingerprint feature extraction and recognition techniques. The proposed system is the combination of total variation model and sparse representation with multi-scale patching. TV model divides the image into two components: texture and cartoon components. The texture components are characterized as the informative structure of small patterns and the cartoon components are eliminated as non-fingerprint patterns. Initially we apply Gabor functions on a high quality fingerprint images to obtain the characteristics of ridge structures like ridge orientation and frequency. Then the dictionaries are created by repeated learning from a set of well defined fingerprint patterns. Using the knowledge of the dictionary, multi-scale patch based sparse representation is used to enhance and restore the ridge structures in latent fingerprint images. Finally Levenberg-Marquardt algorithm is used to train Neural Networks for fingerprint matching and identification. The proposed algorithm reduces the distortion and enhances the finger print pattern thereby increases the recognition rate.