The authors have devised a novel method for the recognition of low-resolution character images, based on the extraction of features from the dual eigenspace and synthetic degraded patterns. The first stage of the recognition system is an image normalization step, where a character is normalized to a 24-by-24-pixel enhanced grayscale image. A database of synthetic degraded training patterns is generated, based on a simplified video degradation model proposed by Katsuyama et al. [1]. The second stage of the system is based on extracting the dual eigenspace features, which differs from the traditional principal component analysis (PCA) method in the way the individual eigenspace is built. In the dual eigenspace scheme, the individual eigenspace is built for every category, using the first feature of all the training samples, while, in the PCA scheme, it is based on the mean of every category.
The proposed system has been applied to the recognition of low-resolution Japanese Kanji characters captured by a digital camera. The performance of the proposed system was compared with the traditional PCA method, and the contour directional method proposed by Shridhar and Kimura [2]. Results show that the dual eigenspace feature method is much better than the PCA method, however the contour directional method provides better recognition for low-resolution, simply structured characters.
The authors present their work in a very clear and systematic way. The different stages of the proposed systems are not original contributions of the authors; however, integrating them might be a new contribution. Therefore, the proposed method has made a significant contribution to the field.