Robust telop character recognition in video for content-based retrieval

The recognition of telop characters in video has two problems: edge degradation and background noise. To overcome these problems, the paper proposes: (1) a feature that describes the shape of background region in addition to character region; (2) a classifier that dynamically decreases the influence of noise using the difference in the number of pixels within a local area. Experiments show that the two proposals achieve a recognition rate of 76% and a cumulative classification rate within the best ten of 92%. Both values are significantly better than those of conventional methods