Lu T., Palaiahnakote S., Tan C.L., Liu W. Video Text Detection

Springer, 2014. — 272 p.With the increasing availability of low cost portable digital video recorders, we are witnessing a rapid growth of video data archives today. The need for efficient indexing and retrieval has drawn the attention of researchers towards handling the video databases. However, efficiently handling video content is still a difficult task in the pattern recognition and computer vision community, especially when the size of the database increases dramatically. A lot of ideas have been proposed, but like other frontiers in this community, there is no reliable approach that has theoretical grounding in video content analysis.
Fortunately, studies have shown that humans often pay their first attention to text over other objects in video. It is probably due to the ability of humans to simultaneously process multiple channels of scene context and then focus the attention on texts in video scenes. This fact makes video text detection a feasible and probably the most efficient way for indexing, classifying, retrieving and understanding the visual contents in videos. This can be further used in transportation surveillance, electronic payment, traffic safety detection, sport videos retrieval, and even commercial online advertisements. One example is video-based license plate recognition systems, which are accordingly necessary to help improve the convenience of checking vehicle status at roadside and designated inspection points efficiently. Another example is online video advertising. Driven by the advent of broadband Internet access, today’s online video users face a daunting volume of video content from video sharing websites, personal blogs, or from IPTV and mobile TV. Accordingly, how to develop advertising systems especially considering contextual video contents through efficient video text detection techniques has become an urgent need.
Actually, video text detection has not been systematically explored even though people have developed a lot of optical character recognition (OCR) techniques, which are considered as one of the most successful applications in the past decades. For example, to explain a typical Google street video scene view, popular visual understanding methods detect and identify objects such as car, person, tree, road and sky from the scene successfully. However, regions containing text tends to be ignored. It is probably due to the fact that text from video is sometimes difficult to detect and recognize. The performance of OCR thereby drastically drops when applied to video texts which are either artificially added (graphic text) or naturally existing on video scene objects (scene text). There are several reasons for this fact. First, the variety of color, font, size and orientation of video text bring difficulties to OCR techniques. Second, video scenes exhibit a wide range of unknown imaging conditions which in general add sensitivity to noises, shadows, occlusion, lights, motion blur and resolution. Finally, the inputs of most of the OCR engines are well segmented texts which have been distinguished from background pixels. Unfortunately, the segmentation of video text is much harder.
This book tries to systematically introduce readers to the recent developments of video text detection for the first time. It covers what we feel a reader who is interested in video text detection ought to know.Introduction to Video Text Detection
Video Preprocessing
Video Caption Detection
Text Detection from Video Scenes
Post-processing of Video Text Detection
Character Segmentation and Recognition
Video Text Detection Systems
Script Identification
Text Detection in Multimodal Video Analysis
Performance Evaluation