Importance • Latent fingerprints are an important source of forensic evidence • Improving latent fingerprint matching is one of the major goals of FBI’s NGI Program Houston Cold case: Latent fingerprint found on the victim’s car was used to identify the criminal using FBI’s IAFIS.* Motivation • NIST FpVTE: Best matcher achieved rank-1 identification rate of ~99.4% on plain prints [1] • NIST ELFT EFS II: Best rank-1 identification accuracy for latents is ~63.4% in “lights-out” mode [2] NIST FpVTE NIST ELFT-EFS Challenges Poor Ridge Clarity Small Friction Ridge area Complex Background Noise *http://www.fbi.gov/news/stories/2011/october/print_101411 Proposed Feedback Paradigm Latent Probe Image Pre-processing Feature Extraction Exemplar Background Database Bottom-up data flow Matching Match Score • Latent matchers [3] [4]: “bottom-up matching” • Feature extraction for latents is not reliable due to poor ridge clarity and background noise • We propose to incorporate “top-down information” or feedback from exemplars to improve latent feature extraction Latent Fingerprint Matching: Accuracy Gain via Feedback from Exemplar Prints Sunpreet S. Arora, Eryun Liu, Kai Cao and Anil. K. Jain Michigan State University http://www.biometrics.cse.msu.edu Top-down data flow Feedback loop for feature refinement Overlap with other prints Resorting Candidate List Latent Matcher Alignment Matched Minutiae Initial Match Score Initial Matching and Alignment Exemplar Feature Extraction Latent Feature Extraction and Refinement Match Score Computation • Exemplar features extracted using COTS matcher Exemplar Exemplar orientation • Latent matcher matches the latent image to the exemplar (Bottom- up mode) • Candidate list: Top-K candidate exemplars returned by the latent matcher • Matched minutiae used to align the latent-exemplar pair Exemplar Latent Latent Exemplar orientation Refined latent feature • Latent features extracted in Fourier Domain • Refined latent orientation: Latent ridge orientation closest to the exemplar orientation • Refined latent frequency: Latent ridge frequency corresponding to the selected orientation Number of overlapping blocks Refined latent orientation Exemplar orientation Updated Match Score Initial Match Score Feedback Orientation similarity Feedback Frequency similarity Feedback Orientation Similarity Number of overlapping blocks Refined latent frequency Exemplar frequency Feedback Frequency Similarity Match Score Update • Product fusion to obtain the updated match score: The Need for Feedback • Bottom-up matching suffices for good quality latents • Feedback should be applied only when needed Global Criterion Local Criterion • For each latent query, decide if feedback is needed • Model the distribution of match scores obtained by matching the latent to the top- K candidate exemplars • Test for the presence of an upper outlier in the match score probability distribution • Within each latent, decide regions which need feedback • Decide exemplar regions which can provide feedback Latent image Ridge Clarity Map [5] Gray region needs feedback Exemplar image Ridge Clarity Map [5] White region can give feedback 0 0.05 0.1 0.15 0.2 0 10 20 30 40 50 60 Match Scores Count 0 0.05 0.1 0.15 0.2 0 10 20 30 40 50 60 70 80 90 Match Scores Count Upper outlier present No upper outlier Feedback needed No need for feedback Experimental Evaluation Match Score Distribution for Latent Query 1 Match Score Distribution for Latent Query 2 • Wrapped the feedback paradigm around the latent matcher in [4] • Databases: - NIST SD27: 258 operational latents - WVU: 449 latents collected in West Virginia University • Background: ~32k exemplars • Rank-1 identification accuracy improves by ~9.5% and ~3.5% for NIST SD27 and WVU databases, respectively Latent Mated exemplar Refined orientation field Extracted orientation field NIST SD27 WVU The retrieval rank of the mated print of the latent improved from 92 to 5 after feedback References [1] C. Wilson et al., Fingerprint vendor technology evaluation 2003: Summary of results and analysis report, NISTIR7123. [2] M. Indovina et al., ELFT‐EFS Results, NIST Evaluation of Latent Fingerprint Technologies: Extended Feature Sets Evaluation#2. [3] A. Jain and J. Feng, “Latent Fingerprint Matching”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(1):88–100, 2011. [4] A. Paulino et al., “Latent Fingerprint Matching using Descriptor-based Hough Transform”, IEEE Transactions on Information Forensics and Security, 2013. [5] S. Yoon et al., “On Latent Fingerprint Image Quality”, in Proceedings of International Workshop on Computational Forensics, 2012.