Recursive reconstruction of non-facial components using support vector data description

To distinguish non-facial components, such as glasses, scarves, and hair, this study proposes a recursive reconstruction method using support vector data description that estimates the approximate face by determining the minimum distance between the transformed input and the trained hyperball in a feature space. The performance of this reconstruction is further improved by replacing the non-facial components with an average face for the initial reconstruction and recursively updating the face using support vector data description. To validate the effectiveness of the proposed method, we perform reconstruction experiments by varying the position and size of non-facial components using images from the AR, the Korean, and the XM2VTS facial database, and measure the reconstruction errors. The experimental results with three databases and some live face images confirm the effectiveness of the proposed method for reconstructing non-facial components.