Implementation and Evaluation of Methods for Contactless Palmprint Recognition

Abstract

Contactless palmprint recognition is a biometric technology which due to its ease of acquisition, non invasive nature and potentially cheap cost can come to stand competitive to more widely used biometric traits such as finger- or facial recognition. This work uses two different feature extraction methods for palmprint images, the Scale invariant feature transform (SIFT) with an iterative random sample consensus (I-RANSAC) algorithm for refining detected keypoints and the Local Line Directional Pattern (LLDP) which finds the directional line responses in the image by exploiting the Gabor filter in twelve directions. A support vector machine (SVM) classifier is trained to combine the extracted features for the final recognition. A method to extract the region of interest (ROI) is developed for use in a real world setting. Preprocessing techniques are investigated for enhancing the SIFT I-RANSAC and the LLDP matching is refined by a local search algorithm. The proposed method as well as state of the art fingerprint algorithm are evaluated on several different databases: the \IITD Palmprint database, the CASIA Palmprint Image database and the Google 11k Hands. In addition to this a database was also collected to reflect the setting and potential problems for a real world contactless recognition system, the Precise Biometrcis Palmprint database. The method shows promising results and the problems and future challenges becomes more clear upon evaluation. To improve the recognition, work on the ROI extraction should be the primary focus.

Popular Abstract

The use of identification systems using our bodies are becoming more common. Almost every smartphone released today either include face or fingerprint recognition as a method to unlock the phone. It is fast while also being secure, but most importantly relieves us from having to remember a password. However, some people are skeptical to share images of their faces and are more inclined towards fingerprints. While fingerprint systems are great, they are less practical since they require contact with a sensor. With contact-less palmprint recognition, the benefits from both such systems can be unlocked.

Using a camera is convenient but difficulties arise since the background and illumination may vary between images. With this in mind we have developed a new method to improve palmprint recognition! This new strategy combines two older methods. One which recognizes hands by comparing all of the lines on the palms, and one which detects "keypoints" which are points of interest found on the hand using advanced computations. Both methods gives a score which tell how similar two palms are. These scores are then given to a powerful machine learning algorithm which can tell us if it is the same hand or not. The method was tested on three large collections of hands. Each collection containing over 5000 images. To test the method on images taken in a more realistic setting, we also designed our own database to see how our algorithm dealt with common everyday problems such as different lighting and backgrounds. These tests showed that the method was both secure and competitive.

When developing this method we could confirm the biggest challenge when using images of hands: finding the palm! If this is not done the algorithms will try to recognize the whole image, even the background, which will cause it to fail. To find the palm, another machine learning algorithm was used. This one had been trained to find hands in images. When the location of the hand is found, computations on the image could be done to remove the background. Subsequently the locations of the "valleys" between the fingers could be found and used to find the location of the palm.

To identify a hand, you need stored images to compare against. These stored images are called templates. When the method was tested on images of hands taken under different lightning conditions, it was discovered that what kind of templates you store is very important. If templates taken only under a specific setting were used, it was hard to recognize palms taken under different illumination conditions. An effective way to deal with this was discovered. By storing templates taken with varying illuminations, the recognition could be improved greatly. There is still work that needs to be done to make palmprint recognition great, but these finds might be a good start on this journey. Perhaps it won't be too long until you can enter a store, realize you forgot your wallet, and say "no problem, I'll pay with my hand".

• author: Breimer, Teodor and Ly, Adam
• supervisor: Magnus Oskarsson
• organization: Mathematics (Faculty of Engineering)
• alternative title: Implementering och utvärdering av metoder för kontaktlös igenkänning av handflatan
• course: FMAM05 20192
• year: 2020
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics
• keywords: Palmprint recognition, Biometrics, Image analysis, Computer vision, SIFT, RANSAC, LLDP, Feature extraction, Convolutional neural networks, Machine learning, Transfer learning, SVM, Gabor filter
• publication/series: Master's Theses in Mathematical Sciences
• report number: LUFTMA-3397-2020
• ISSN: 1404-6342
• other publication id: 2020:E1
• language: English
• id: 9005550
• date added to LUP: 2024-10-07 13:38:29
• date last changed: 2024-10-07 13:38:29