Extracting Caption Content on Sports Videos by Caption Identification

Abstract

The paper describes a novel caption extraction scheme to detect unconstrained captions on sports videos and identify the captions for extracting the caption content directly. A caption detection process based on a multi-frame averaging approach is applied to locate a reliable caption region accurately. Furthermore, a caption-content extraction process based on caption identification and model masking approaches is applied to directly extract the caption content without a conventional segmentation process. Experimental results show that the proposed caption detection approach is very efficient to provide a high tolerance to noisy and complex-background video frames. Furthermore, a learning-based approach for the caption identification process is capable of generalizing the learning knowledge to identify various sports captions and get an average identification rate of 89.99% for testing data.