一种改进的基于PCA的人脸超分辨率算法

In this paper, based on Circularly Symmetrical Gabor Transform (CSGT) and Principal Component Analysis (PCA), we propose a face hallucination approach. In this approach, all of the face images (both input face image and original training database) are transformed through CSGT at first and then local extremes criteria is utilized to extract the intrinsic features of the faces. Based on these features, we calculate Euclidean distances between the input face image and every image in the original training database, and then Euclidean distances are used as criteria to choose the reasonable training database. Once the training database is chosen, we use PCA to hallucinate the input face image as the linear combination of the chosen training images. Experimental results show that our approach can choose training database automatically according to the input face image and get high quality super-resolution image.