Three-dimensional face recognition enhancement using extreme expression rejection
Yu Hen Hu, Kerry R. Widder · 2011
Face recognition has received much attention in the research community in recent years. Although much progress has been made in solving this difficult problem, many issues remain, leaving room for improvement. One of these issues is the impact of expression. In previous face recognition methods, expression has often been handled by increasing the complexity and sophistication of the recognition methods. In this dissertation, an alternative expression handling approach is developed utilizing the idea of classification with rejection. An expanded and flexible framework is presented for the application of rejection. In the face recognition context, this approach seeks to detect images that are difficult to recognize. Once these images are detected, they can be handled by more sophisticated recognition algorithms, or rejected. In the scope of this work, the rejection option will be used. A tool, called the Optimal Rejection Ranking , is developed to aid in the analysis of the face recognition performance of rejection classifiers. It represents the best performance achievable via rejection. A method is developed that detects facial images that have extreme expression. The method utilizes the mouth region of the facial images, which is the region having the greatest degree of variation due to expression. A set of neutral expression eigen-mouths are extracted from the training images using Principal Component Analysis. The coefficients derived from projecting each test image onto the eigen-mouths are compared with the coefficients produced from the neutral expression training images. The decision between normal and outlier is made using a Constant False Alarm Rate detector. This detector performs well at detecting extreme expression, and yields a recognition performance improvement of 1.8% at a rejection rate of 10%. An extension to the eigen-mouth method is developed utilizing additional regions of the face and fusion of the results. Another extension is presented that adds some of the expressive test images to the training set and seeks to learn the difficult images. A comparison of all methods shows different strengths for each. The results provide resources for choosing the best method based on an application's constraints.