A robust expression negation algorithm for accurate face recognition for limited training data

Although important and effective contributions on face recognition under varying facial expressions have been reported up to date, most of the methods need multiple images of an individual stored in the database. However, this problem becomes more challenging when a limited number of training samples are available as is the case for expression invariant face identification for surveillance and security applications. This paper proposes a simple and effective method that can be integrated into any face and expression recognition system to improve the overall recognition accuracy even under limitation of training samples. In this approach, neutral component of the expressive image is estimated utilizing prior information obtained from different subjects under the same expression. Basically by analyzing the impact of a particular expression on a neutral face a nullification process is developed to convert the expressive image to a neutral face. In order to make it justifiable to utilize common expression information for different subjects, an alignment strategy is employed where for each expression a specific expression template is used, and the images are warped to their corresponding expression face template. After negating the facial expression from the expressive images, principal component analysis (PCA) is applied to reduce the dimension and cosine similarity matching is used for classification. The experimental results on Cohn-Kanade database exhibit the effectiveness of the proposed method even when there is a single training sample per class is available in the database.