Individual-adapted facial shape transformation based on subspace learning

As the use of high-resolution digital images continues to increase, there is a growing interest in facial image warping or facial image transformation, which is the process of making visual changes to facial images. Several facial image transformation techniques have been proposed and used in many applications. However, most of these techniques warp or transform a facial image to a pre-defined target image. As a result, the transformed image may not maintain the individual features, and sometimes an unnatural facial image is produced. In this paper, we propose a novel subspace learning based method for automatic individual-adapted facial shape transformations. In our proposed method, we first prepared a database containing several tens of pairs of facial images and their manually generated target images, and then constructed a facial shape subspace. We then projected both original facial shapes and their target image shapes onto the subspace and used the B-spline method to estimate the mapping functions from the original shapes to their target shapes. By using the mapping function, the target image shape points of a test facial image could be automatically generated, thus maintaining the original individual features.