Research On Text Localization Algorithm In The Image

Text detection from natural scene is a hot research topic in the field of image processing, as lot of useful information could be retrieved from news subtitles, license plates and advertisements. Three kinds of typical texts such as subtitle of news, license plates and texts in nature scenes are subject to detection in the thesis.The thesis uses wavelet transform based texture analysis to locate the subtitles in news video which are highly similarc in size, position and colour. Firstly, the algorithm processes video frames using wavelet transform, extracts the LH, HL and HH sub-band statistical feature of transform result, and classifies them by SVM and K-Means. The algorithm is tested with5Chinese and5English videos:SVM Chinese classifier achieves average detection rate of0.867, English classifier achieves0.863; K-Means classifier achieves0.817, and K-Means English classifier achieves0.803. These two methods are compared at the end.Because pixels beside the edges of characters on license plates are all "blue&white" combinations, the thesis approach a colour model according to the distribution characteristics of colour pair. The algorithm filters out inconsistent edge pixels inconsistent with the distribution model, and thus gets candidate text pixels. Those candidate edge pairs are then connected and filled using horizontal scan. After morphological process, the location of plate is retrieved by its geometrical features. The test results demonstrate that the proposed algorithm can accurately locate0.91of license plates correctly for correctly exposed images.For random text in natural scenario, text in the same context should be consistent in colour and stroke width, and aligned vertically. Because of these characteristics, the thesis first computes stroke width with SWT (Stroke Width Transform), calculates the number of pixels between two edges of edged objects (stroke width), and then locates contextual text according to the feature of texts in the same context. The experimental results show that the proposed algorithm is independent of the language detected and can locate text correctly at a rate of0.66.