Two image quality assessment methods based on evidential modeling and uncertainty: application to automatic iris identification systems

Abstract

The performance of an Automatic iris Identification System is impacted by both the poor quality of iris images and the uncertainty of information. Assessing image quality and rejecting poor-quality images can substantially improve the performances of the current biometric systems. The main idea behind our proposed Image Quality Assessment approaches is to take advantage, firstly, of the texture of iris images and, secondly, of the uncertainty of these information. This is achieved by defining a set of Contextual Quality Indicators extracted from the image texture and transforming them into Quality Assessment Criteria in the evidential framework, taking into account the information uncertainty degree. The Contextual Quality Indicators are defined based on a priori analysis of the context of the application. We use ‘iris’ as the context of application. Generally, only the normalized iris image is saved, i.e. the acquired iris image is not always available. So, the main advantage of our approaches over other related methods is that it can act in the normalization level of the processing chain to reject poor-quality images. So that, the subsequent Automatic iris Identification System can process only good-quality images, which result in better recognition rate performance. The functioning of our evidential approaches is illustrated using image samples from CASIA 1.0 database. The performance of over the proposed image quality assessment approaches is compared with the standard iris identification system without an image quality assessment step. A statistical test, based on 95% confidence interval, is used to assess if there is a statistically significant difference between the performances of the proposed approaches. The CASIA 1.0 has been used to make the comparison. The comparison results highlight the effectiveness of the proposed approaches for iris domain of applications. The source code of our paper is available at https://github.com/Sonda09/IIQA