Video Text Recognition Using Category-Dependent Feature Extraction Based on Feature Compensation

摘要

When recognizing multiple fonts, geometric features, such as the directional information of strokes, are generally robust against deformation but are weak against degradation. This paper describes a category-dependent feature extraction method that uses feature compensation to overcome this weakness. Our proposed method estimates the degree of degradation of an input pattern by comparing the input pattern to a template of each category. This estimation enables us to offset the degradation in feature values. We apply the proposed method to the recognition of video text suffering from both degradation and deformation. Recognition experiments using characters extracted from videos show that the proposed method is superior to the conventional alternatives in resisting degradation.