Research On Overlay Text Extraction From Images With Complex Background

Text in images and videos carries plenty of semantic information useful for understanding thecontent of images and videos. Thus, it makes text recognition very significant for understanding andretrieval of images and video. However, text is usually embedded in complex background of images,which makes direct optical character recognition almost impossible. Therefore, it becomes necessary toextract text from complex background before recognition. In the last decades, many efforts have beendevoted to developing effective algorithms to extract text from complex background in images andvideo. However, the state-of-the-art of text extraction is far from perfect due to the great difficulty indiscriminating text from complex background completely.In this thesis, my research work aims to improve the performance of text extraction from twoaspects: more robust text segmentation algorithm utilizing hybrid information and effectivepost-processing techniques to eliminate background residues. The features extracted from strokes ofcharacters are experimentally checked to evaluate their feasibility in discriminating text regions fromcomplex background regions. The main contributions of the thesis include:1. A robust text segmentation algorithm is proposed based on hybrid information of text such ascolor and scale. It can generate a more precise estimation of text color via the smart sampling near theedges around text. Also, by utilizing the hybrid information, it becomes more efficient in removingcomplex background regions, compared with most existing algorithms.2. A group of heuristic constraints are designed to eliminate background residues after textsegmentation based on color, edges, scale and the spatial relation of connected components, which caneffectively eliminate a wide range of background residues and can be used in various text segmentationalgorithms.3. A new text detection method for Chinese characters is proposed. It utilizes the featuresextracted from strokes to differentiate text regions from complex background regions, and the relatedexperiments show that the new features are effective.