Predict and Improve Iris Recognition Performance Based on Pairwise Image Quality Assessment

The iris recognition performance is partially dependent on the relative quality variations of pairwise iris images. So bridging the gap between the quality and the matching score of pairwise iris images is helpful to predict and improve iris recognition performance. This paper formulates the relationship between matching score and quality of pairwise iris images as a statistical regression problem. Firstly, a number of quality measures of iris images such as focus, motion blur, illumination, off-angle, occlusions and dilation are computed as the performance related feature vector of iris images. And then partial least squares regression is used to establish two models to predict the intra score and inter score from pairwise iris image quality respectively. Finally, we define the uncertainty interval of matching scores. The uncertain match pairs are discarded to improve the recognition performance. Extensive experiments on ICE 1.0, CASIA-Iris-Lamp and CASIA-Iris-Thousand demonstrate that the proposed method can accurately estimate the distributions of matching scores. It can simultaneously improve the performance, even using simple features in recognition.