Abstract—Example-based algorithms are most suitable for face image super-resolution since they typically use specific kinds of images as their dictionary. Recently, a significant method using a set of normalized face images as database was proposed, where a high-resolution face image was estimated by taking into consideration their facial parts. This paper reports parametric analysis of the method, regarding the setting of image-patch size during super-resolution process and their relations with image’s scaling factor. Our objective is to find the best patch size for the algorithms, which may produces better output high-resolution images. We generated training images’ patch-databases with different patch size, i.e. 5x5, 7x7, 9x9, 11x11 and 13x13 in pixels. We ran the method onto several sets of low-resolution face images with different scaling factor of magnification, i.e. 2, 3, 4, 5 and 6 times, to generate high-resolution images using different patch size orderly. Then, we observed the average of Peak Signal-to-Noise Ratio (PSNR) values for each set of constructed high-resolution images to analyze which size of patch yielded better results. According to the resulting PSNRs, interestingly we found that the best patch size is adaptable to scaling factor, where if the scaling factor is n, the best setting of patch size in the algorithms can be determined by (2n+1)².

Index Terms—example-based, super-resolution, face image, facial parts, patch size, scaling factor

Cite: Suhail Hamdan, Yohei Fukumizu, Tomonori Izumi, and Hironori Yamauchi, "Face Image Super-Resolution with Adaptive Patch Size to Scaling Factor," Journal of Image and Graphics, Vol. 6, No. 2, pp. 167-173, December 2018. doi: 10.18178/joig.6.2.167-173