A Machine Learning Approach to Assessment of Image Quality through Prediction of Iris Recognition Success

It is clear from prior research that quality metrics can predict and improve recognition performance. As we add new quality metrics, however, determining a weighting scheme using simple rules for their combination becomes infeasible. This argues for a machine learning approach that will automatically determine the optimal weighting. We propose to develop a large set of quality metrics based on requirements set down in the IREX II/IQCE evaluation. These metrics will form a feature vector supplied to a novel manifold-learning algorithm developed recently for function approximation, which will be trained to predict the performance of stages in various iris recognition algorithms as well the algorithms' final recognition performance. Applying standard feature selection techniques, our system will also automatically determine a weighting for the features to be used, identifying features that contribute little and can be eliminated. The final quality metrics will be combinations of the outputs of the feature extractor associated with the metrics and their weighting as determined by our system. Such a flexible image quality assessment system would have application beyond iris recognition to other areas in biometrics, such as face recognition, but also to domains such as stereovision, visual odometry, and general object recognition. Image quality metrics are largely at the level of proof-of-concept demonstration on available datasets. Our work will take this technology from TRL 3 to TRL 4 by integrating the metrics with recognition algorithms by acting as a front-end predicting performance and so indicating whether an image is adequate to the task.