A new method for multi-oriented graphics-scene-3D text classification in video

Abstract

Text detection and recognition in video is challenging due to the presence of different types of texts, namely, graphics (video caption), scene (natural text), 2D, 3D, static and dynamic texts. Developing a universal method that works well for all the types is hard. In this paper, we propose a novel method for classifying graphics-scene and 2D–3D texts in video to enhance text detection and recognition accuracies. We first propose an iterative method to classify static and dynamic clusters based on the fact that static texts have zero velocity while dynamic texts have non-zero velocity. This results in text candidates for both static and dynamic texts regardless of 2D and 3D types. We then propose symmetry for text candidates using stroke width distances and medial axis values, which results in potential text candidates. We group potential text candidates using their geometrical properties to form text regions. Next, for each text region, we study the distribution of the dominant medial axis values given by ring radius transform in a new way to classify graphics and scene texts. Similarly, we study the proximity among the pixels that satisfy the gradient directions symmetry to classify 2D and 3D texts. We evaluate each step of the proposed method in terms of classification and recognition rates through classification with the existing methods to show that video text classification is effective and necessary for enhancing the capability of current text detection and recognition systems.