Abstract

3D fingerprint capture is less sensitive to skin moisture levels and avoids skin deformation, which is common in contact-based sensors, in addition to capturing depth information. Unfortunately, its adoption is limited due to high cost and system complexity. Photometric stereo provides an opportunity to build low-cost, simple sensors capable of high-quality 3D capture. However, it assumes that the surface being imaged is lambertian (unlike our fingers). We introduce the Split and Knit algorithm (SnK), a 3D reconstruction pipeline based on the photometric stereo for finger surfaces. It introduces an efficient way of estimating the direct illumination component, thus allowing us to do a higher-quality reconstruction of the entire finger surface. The algorithm also introduces a novel method to obtain the overall finger shape under NIR illumination, all using a single camera. Finally, we combine the overall finger shape and the ridge-valley point cloud to obtain a 3D finger phalange. The high-quality 3D reconstruction also results in better matching accuracy of the captured fingerprints.

Overview

Pipeline

The Split-and-Knit Algorithm consists of three primary components: (a) Reconstruction pipeline for finger ridge-valley point cloud. The white light images are preprocessed to obtain cropped phalange. Using CLAHE and global-direct component separation by U-Net, the non-lambertian nature of the finger image is reduced. Further, using photometric stereo and shapelet reconstruction, finger ridge-valley point cloud is obtained (b) Reconstruction pipeline for finger shape point cloud. The NIR light images are preprocessed to obtain cropped phalange. Further, we apply photometric stereo and Frankot-Chellappa reconstruction to obtain the finger shape point cloud. (c) The phalange point cloud is obtained by pixel-wise addition of the ridge-valley and finger shape point cloud.

Reconstruction Steps

Output Glimpse

Qualitative Comparison

Qualitative Comparison: a) Comparison of the front view and side view of the phalange point cloud produced by SnK: (Split-and-Knit), PR: (Partial Result), EA: (Existing Algorithm). The EA considers fingers to be lambertian and reconstructs using grayscale images leading to undetailed fingerprints and distorted global shape. The PR is the intermediate result obtained after reducing the non-lambertian nature before extracting the ridge-valley and adding global shape leading to a detailed fingerprint but distorted global shape. The SnK gives the best result with a detailed fingerprint and proper global shape. b) Comparison of zoomed ridge-valley point cloud for SnK and EA. SnK retrieves superior quality ridge-valley point cloud.

Surface Normal Matching using LBP features

ROC curves for matching fingerprint surface normals for SnK: Split-and-Knit algorithm, PR: Partial Result after reducing the non-lambertian nature of finger image with the distorted overall shape, and EA: Existing Algorithms output based on the photometric stereo. The best ROC curve and 92%TAR @ 0.01FAR are obtained for SnK, displaying its highest reconstruction quality.

Ridge-Valley Point Cloud Matching using LBP features

ROC curves for matching ridge-valley point cloud obtained from SnK: Split-and-Knit algorithm and EA: Existing Algorithms based on the photometric stereo. The 95% TAR @ 0.01 FAR and 90.5% TAR@0.001 FAR of SnK as opposed to 93% TAR @ 0.01 FAR and 78.8% TAR @ 0.001 FAR of EA proves the high quality of the ridge-valley pattern by SnK.

Summary of Matching Experiments

The above Bar-Graph compares the matching accuracies of SnK: Split-and-Knit algorithm output and EA:Existing Algorithms output based on the photometric stereo. The higher % TAR of SnK over EA for ridge-valley point cloud matching and surface normal matching for all cases show the higher quality of 3D phalange produced by the SnK.

3D to 2D fingerprint Conversion

We tried to prove the high quality of the ridge-valley point cloud by visualizing the ridge-valley point cloud as a range image. Hence, we obtain the 2D fingerprint in a novel way directly from the range image of the ridge-valley point clouds without any image processing on the point cloud. The traditional method to obtain 2D fingerprints from contactless finger images is by following the enhancement and scaling methods. The 2D fingerprint obtained from visualizing the ridge-valley range image appears similar to the 2D fingerprint obtained by processing the contactless fingerprint. It again proves the excellent quality of the 3D reconstruction of the ridge-valley point cloud by SnK. Further, we obtained the enhanced fingerprint using Gabor filter-based enhancement. Also, obtaining the 2D fingerprints makes the Split-and-Knit algorithm (SnK) compatible with 2D fingerprint matching algorithms.