On Latent Palmprint Matching

The evidential value of palmprints in law enforcement is clear from the following statistic: 25% of all crime scenes contain only latent palmprints [1] and 30% of the latents recovered from crime scenes are of palms [2]. In forensic applications, the palmprint image resolution is 500 pixels per inch (ppi) [5] and the matching is based on friction ridges, flexion creases and minutiae points. Palmprint recognition has recently become popular for person authentication in commercial applications (e.g., access control). But, these applications utilize low resolution (∼ 100ppi) images to keep the sensor cost and processing requirements low. The feature extraction and matching techniques used in these low resolution systems cannot be adopted for forensic applications since they do not (i) address latent palmprint matching, and (ii) do not use friction ridge and minutiae information. We propose a palmprint matching system utilizing a variety of palmar features (friction ridges, minutiae, flexion creases and palmar texture) present in 500 ppi images. Our system is able to match either a full or partial/latent input image to a database of full palmprint images. Since there is no public domain 500 ppi palmprint database, our results are based on an in-house database of 100 unique palms with 10 impressions per palm. We are able to achieve a genuine matching accuracy of 98.9% at an FAR of 0.01% for full-to-full palmprint matching. In case Department of Computer Science & Engineering, 3115 Engineering Building, Michigan State University, East Lansing, MI 48824. Emails: {jain,demirkus}@msu.edu