Estimation of Missing Values in Multimodal Biometric Fusion

The computation of any similarity score will be precluded by missing values. Missing values can be attributed to poor quality biometric data, poor data capture, or classifier error from computing the similarity scores. The presence of missing values in biometric systems can be inconvenient to the user, as the system will reject the submitted biometric data and request for another. It is therefore important for a biometric system to be prepared for, and able to deal with missing values. Currently methods for dealing with missing values can be categorised into: 1) (deletion) - deleting missing values; 2) (maximum likelihood) - computing maximum likelihood of observed data, while integrating out the missing values; 3) (imputation) - replacing missing values with estimated values. This paper adapts the popular k-nearest neighbour (k-NN) imputation method to produce three imputation methods for dealing with missing values in classification. We also introduces a forth category for dealing with missing values, called the exhaustive fusion framework. This method eliminates the need to predict or delete missing values. We show experimentally that our proposed methods provide an improved performance over the original k-NN and the widely used mean method for predicting missing data. These experiments were carried out using the newly developed BioSecure database and the popular XM2VTS database.