This paper presents a face hallucination using training data sets as low and high-resolution patch pairs for an input low-resolution face image. It is complicated to be acquainted with details from a low-resolution image since of severe aliasing and unfortunate face image quality, hence gratitude from the low resolution face image may effect in false gratitude decision. In order to get better gratitude performance, the anticipated expansion method is adopted. Considering the coupled PCA compensation algorithm, this capably exploits the local distribution structure in the training samples. The first and second steps were generate global features the main characteristics of the real image and produces residual image to compensate the outcome of the first step respectively. Experiments give an idea about that the anticipated method generate higher quality face image than recent several methods.