Research Of Fingerprint Images Quality Evaluation Methods And Applications

Fingerprint recognition is the earliest and the most mature identity authentication that based on biometric traits. Due to its uniqueness, invariance and convenience, the fingerprint recognition technology is widely applied in various fields. Automatic fingerprint recognition technology is also a research hotspot of biometrics, and has lots of achievements in recent years. However, the performance of Automated Fingerprint Identification System (AFIS) is affected severely by fingerprint image quality. The evaluation of fingerprint image quality is important to improve the performance of system, especially in fingerprint image collection, preprocess and match.At present, several aspects of works have been studied to improve the performance of a fingerprint system. Firstly, some researchers pay effort to enhance the performance of the processing steps, such as segmentation, enhancement, feature extraction and match. Secondly, multiple impressions or multiple algorithms are used to increase more information. Thirdly, high-resolution fingerprint techniques and algorithms have been proposed. But little literatures focus on the research of fingerprint image quality. How to estimate the quality of fingerprint image and how to use image quality factor directly to improve the performance of a fingerprint verification system, are meaningful and valuable subjects.In order to solve the problem of the low accuracy of the quality classification of fingerprint image, a new fingerprint image quality classification method based on metric learning is proposed. Seven features are extracted to estimate the quality from fingerprint, which include:1) effective area of the fingerprint image;2) the mean of gray;3) the variance of gray;4) Gabor feature;5) the direction contrast;6) the direction consistency;7) the spectral energy concentration. With LMNN algorithm, we learn a new distance metric. In the new metric space, we use KNN to classify the quality of fingerprint image. The experiment result shows its effectiveness.The verification accuracy of existing algorithms for low-quality fingerprint is not high. We proposed a new method using fingerprint image quality to improve the accuracy of fingerprint verification. Together with matching score, fingerprint quality is seen as another factor to determine whether the pair of fingerprints is genuine. In the decision stage, the information is from one-dimension to two-dimension. Experimental results show that our proposed quality-based method leads better accuracy than conventional method. Moreover, False Reject Rate of our method is lower than conventional method when False Accept Rate is fixed.In the method of quality-based fingerprint verification, the relationship between the match score and quality score may be nonlinear. Machine learning methods are introduced to solve the nonlinear problem. We treat each pair of matching fingerprint as a sample. The matching score and quality score are the two features of a sample. If the matching pair is genuine, the sample label is assigned as1. Otherwise, the sample label is assigned as0. Thus, fingerprint verification is transformed into the classification of samples. Machine learning algorithms are adopted to classify the matching results. The experimental results show that using machine learning algorithm can solve the non-linear problem and improve the accuracy of fingerprint verification.Next works focus on the following aspects:the construction of fingerprint quality databases and the standard of fingerprint quality evaluation; the selection and combination of different fingerprint image evaluation indexes; how to use image quality factor more reasonably to improve the performance of a fingerprint verification system and extend to other biometrics.