| 摘要 |
A temporary eigenprojection matrix (#12-a) is temporarily generated from a learning image group (#10), in which high-quality images and a low-quality images are paired up, and a temporary projection nuclear tensor (#12-b) that defines the correspondence relationship between the low-quality image and an intermediate eigenspace and the correspondence relationship between the high-quality image and the intermediate engenspace is created. A first sub nuclear tensor is created (#12-c) by first setting from the temporary projection nuclear tensor, and intermediate eigenspace coefficient vectors (#15-b) are calculated by projecting the learning image group using the temporary eigenprojection matrix and the first sub nuclear tensor (#15-a). Representative intermediate eigenspace coefficient vectors (#15-c) are found from the coefficient vector group in accordance with a learning representative number, and an eigenprojection matrix (#17) and a projection nuclear tensor (#18) are recreated on the basis of a determined representative learning image group (#15-d), and used in a restoration step. Consequently, the redundancy of a learning image set can be reduced, and the reduction of the processing load of projection conversion, the increase of the processing speed, and the reduction of the amount of memory can be achieved. |
