ISSN 2630-0583 (Print)

ISSN 2630-0656 (Online)


Journal of Current Science and Technology

Rangsit Journal of Arts and Sciences. Vol.1 No.2 , July - December 2011.

A novel tensor framework for face hallucination

Kanjana Boonim and Parinya Sanguansat


                Normal regression analysis methods can be used to find a relationship between high-resolution (HR) and low-resolution (LR) features. Previous work neglected to use regression analysis methods in finding a relationship between the error of face reconstruction and the LR feature in tensor objects. Because of this limitation, the image as a featured matrix is transformed into one-dimensional vectors, causing a loss of spatial information. By using our method this problem is eliminated. In our proposal we have developed a new face hallucination framework, using a Tensor Regression Analysis (TRA) to further enhance the quality of an image. Error estimation of in the validation system providing the correct final result is further applied into our framework through regression analysis. In doing this, we present our framework based on a two-tier approach. First a global step and second a local step. In the global step, we apply the TRA in order to find a relationship between the features of LR and HR based on multilinear Principal Component Analysis (PCA). The regression coefficients are acquired through the relationship between featured HR and LR. In the local step, the TRA is used to find the relationship between the LR features and the error of validation process, obtained from the global step. Experimental results obtained from a well-known face database show that the resolution and quality of the hallucinated face images obtained with our proposed method are greatly enhanced and improved in comparison with the traditional methods used.

Keywords: Tensor Regression Analysis, face hallucination, super-resolution

Download Full Paper.