TY - JOUR
T1 - Single image super resolution by ℓ2 approximation with random sampled dictionary
AU - Fujisawa, Takanori
AU - Yoshida, Taichi
AU - Mishiba, Kazu
AU - Ikehara, Masaaki
PY - 2016/2/1
Y1 - 2016/2/1
N2 - In this paper, we propose an example-based single image super resolution (SR) method by ℓ2 approximation with self-sampled image patches. Example-based super resolution methods can reconstruct high resolution image patches by a linear combination of atoms in an overcomplete dictionary. This reconstruction requires a pair of two dictionaries created by tremendous low and high resolution image pairs from the prepared image databases. In our method, we introduce the dictionary by random sampling patches from just an input image and eliminate its training process. This dictionary exploits the self-similarity of images and it will no more depend on external image sets, which consern the storage space or the accuracy of referred image sets. In addition, we modified the approximation of input image to an ℓ2-norm minimization problem, instead of commonly used sparse approximation such as.1-norm regularization. The ℓ2 approximation has an advantage of computational cost by only solving an inverse problem. Through some experiments, the proposed method drastically reduces the computational time for the SR, and it provides a comparable performance to the conventional example-based SR methods with an.1 approximation and dictionary training.
AB - In this paper, we propose an example-based single image super resolution (SR) method by ℓ2 approximation with self-sampled image patches. Example-based super resolution methods can reconstruct high resolution image patches by a linear combination of atoms in an overcomplete dictionary. This reconstruction requires a pair of two dictionaries created by tremendous low and high resolution image pairs from the prepared image databases. In our method, we introduce the dictionary by random sampling patches from just an input image and eliminate its training process. This dictionary exploits the self-similarity of images and it will no more depend on external image sets, which consern the storage space or the accuracy of referred image sets. In addition, we modified the approximation of input image to an ℓ2-norm minimization problem, instead of commonly used sparse approximation such as.1-norm regularization. The ℓ2 approximation has an advantage of computational cost by only solving an inverse problem. Through some experiments, the proposed method drastically reduces the computational time for the SR, and it provides a comparable performance to the conventional example-based SR methods with an.1 approximation and dictionary training.
KW - Examplebased
KW - Single image super resolution
KW - ℓ-norm minimization
UR - http://www.scopus.com/inward/record.url?scp=84957657916&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84957657916&partnerID=8YFLogxK
U2 - 10.1587/transfun.E99.A.612
DO - 10.1587/transfun.E99.A.612
M3 - Article
AN - SCOPUS:84957657916
VL - E99A
SP - 612
EP - 620
JO - IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences
JF - IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences
SN - 0916-8508
IS - 2
ER -