Domain-specific progressive sampling of face images

Domain specific knowledge is useful in image processing applications where the target image to process is known to be of a particular image class. It is commonly used as prior knowledge, to model structured image classes such as human faces in order to break limitations posed by various problems. This paper proposes to use domain specific codebook and corresponding sampling patterns learned from example faces, to build a progressive image sampling algorithm specifically for face processing applications. Instead of accessing the whole target face image, the proposed system is able to progressively sample from it and make approximation of it during the process, allowing the process to stop when image quality is considered to have met the requirement. The proposed system is able to identify significant information from the target image and retrieve it at early stage of the sampling, without requiring the target image to be pre-processed as conventional PIT methods do. Therefore it is applicable to situations where such pre-processing is not possible. The experiment shows that the proposed method is able to efficiently sample and reconstruct face images to achieve significant improvement of PSNR over state-of-art method.