Fingerprint quality assessment is a crucial task which needs to be conducted accurately in various phases in the biometric enrolment and recognition processes. Neglecting quality measurement will adversely impact accuracy and efficiency of biometric recognition systems (e.g. verification and identification of individuals). Measuring and reporting quality allows processing enhancements to increase probability of detection and track accuracy while decreasing probability of false alarms. Aside from predictive capabilities with respect to the recognition performance, another important design criteria for a quality assessment algorithm is to meet low computational complexity requirement of mobile platforms used for enrolment in national biometric systems, by military and police forces. We propose a computationally efficient means of predicting biometric performance based on a combination of unsupervised and supervised machine learning techniques. We train a self-organizing map (SOM) to cluster blocks of fingerprint images based on their spatial information content. The output of the SOM is a high-level representation of the finger image, which forms the input to a random forest trained to learn the relationship between the SOM output and biometric performance. The quantitative evaluation demonstrates that our proposed quality assessment algorithm is a reasonable predictor of performance.
Citation: NIST Interagency/Internal Report (NISTIR) - 7973
NIST Pub Series: NIST Interagency/Internal Report (NISTIR)
Pub Type: NIST Pubs
Biometric Quality, Fingerprint, Fingerprint Quality, Machine Learning, Self-Organizing Maps, Image processing