Grayscale feature combination in recognition based segmentation for degraded text string recognition

Grayscale feature is very effective for degraded character recognition. While many papers focus on different feature extraction algorithms on single character recognition, few deals with the impact of the selected feature on segmentation. For recognition-based segmentation, a good recognition performance on single character may not always have good performance on segmentation. In this paper, two types of grayscale feature, the R-Feature and the S-Feature, are proposed based on dual-eigenspace decomposition. The RFeature is suitable for single character recognition. The SFeature is suitable for text string segmentation. These two feature are combined to further improve the performance for degraded Japanese text string recognition.