Bin-based classifier fusion of iris and face biometrics

Abstract

Accuracy and usability are the two most important issues for a multibiometric system. Most of multibiometric systems are based on matching scores or features of multiple biometric traits. However, plenty of identity information is lost in the procedure of extracting scores or features from captured multimodal biometric data, and the loss of information stops accuracy and usability of the multibiometric system from reaching a higher level. It is believed that matching scores can recover some identity information, which has not been utilized in previous fusion work. This study proposes a framework of bin-based classifier method for the fusion of multibiometrics, to deal with this problem. The proposed method embeds matching scores into a higher-dimensional space by the bin-based classifier, and rich identity information, which is hidden in matching scores, is recovered in this new space. The recovered information is sufficient to distinguish impostors from genuine users more accurately. Therefore, the multibiometric systems which are based on such rich information, are able to achieve more accurate and reliable results. The ensemble learning method is then used to select the most powerful embedding spaces. Experimental results on the CASIA-Iris-Distance demonstrate the superiority of the proposed fusion framework.