Research On The Technology Of Video Caption Detection, Localization And Segmentation

Video captions contain large amounts of key objective information, and they’re themost direct description and explanation of the content of a video clip. So extractingcaptions from videos is the foundation of the video retrieval and comprehension.Generally, there are four stages in the process of video caption extracting: captiondetection, caption localization, image segmentation and OCR. The technology of OCRis highly developed, so this paper mainly studied the key algorithms about video captiondetection, localization and segmentation. The main works are as follow:In the video caption detection and localization stages, the main contributions are:(1)In order to detect the low-contrastive video captions from complicated background,we studied a caption detection algorithm of self-adapted threshold’s filtering based onlocal background’s complexity and of text area recovery based on the edges density. Itcan enforce the caption pixels and restrain the background pixels at the same time, sowe can locate the caption area more easily;(2)In order to overcome the shortcomings ofthe projecting and cutting algorithm in the stage of video caption localiztion, weproposed a new localizing algorithm based on template-matching in multi-scales. Ittakes full advantage of the texture density, texture difference and texture uniformity tosearch the candidate captions. By using integral image the calculation speed of thisalgorithm can be highly accelerated;(3)After video caption detecting and localizing, weadopted the muti-frame verification algorithm to obtain the precise caption based on theimage correlation peak judgment. The experiments show the good performance that wecan detect and localize the low contrastive video captions from complicated backgroundin multi-scales.In caption image segmentation stage, the main contributions are:(1)A OTSU andGaussian local adapted thresholding method was proposed to segment the video captionimge with low contrast. The experimental results confirm the good performance of theproposed method;(2)In order to overcome the shortcomings of the over-segmentatingand under-segmentating problems of the traditional segmentation methods, this paperproposed a method based on “white pixels” increment ratio. This method finds theoptimal threshold using a feed-back model by gradually changing the segmentingthreshold. In order to verify the performance of the proposed segmentation method,several experiments under different complex conditions were designed. The experiment results show that the proposed method can resolve the over-segmenting andunder-segmenting problems of the traditional segmentation methods such as the OTSUalgorithm and the K-means clustering algorithm.