Research On The Text Localization Algorithm In Complex Video

Text in images and videos, which carries high level semantic information, is atype of important source that is useful for understanding the content of images andvideo. Text location has wide applications, such as image retrieval, video retrieval, etc.However, it is not an easy problem to reliably detect and localize the text embedded inimages and videos. The main cause of failure to achieve good diction rate is thecomplex background. The background can have colors similar to the text color, andthe background may include streaks that appear very similar to character strokes.To overcome these problems, two different methods for text detection andlocalization in images and video frames have been proposed.The main work can besummarized as follows:In the first method, we still uses the gradient-based method for text detection.But taking into account the temporal nature of video and the the direction informationof the character strokes, we proposed an an effective coarse-to-fine algorithms todetect text in video. Firstly, in coarse-detection section, the features ofweighted-average gradient energy and motion energy are employed to detect allcandidate text-block, and then a connected compont analysis is introduced toeliminate the false text regions. Secondly, in the fine-detection section, correct textregions are selected from candidate ones by the distribution feature of gradientdirection. Finally, we propose a simple video tracking algorithms.In the other method, we propose a new overlay text detection and localizationmethod using the transition region between overlay text and background. Thetransition map is generated by extracting transition pixels in both horizontal andvertical directions based on our observation that there exist transient colors betweenoverlay text and its adjacent background. The transition pixels in the backgroundare suppressed by block filtering, and candidate text regions are obtained by intensitybased region growing. Finally, the candidate text regions are verified by improvedlocal binary pattern (LBP)and the detected overlay text regionsare localizedaccurately using the projection of transition map.The experimental results show that our algorithms is robust for the variation oflanguage,font,size and color, and compared with the current text detection algorithmsthe two improved algorithms in this thesis performs better than existing methods in terms of spead, precision rate and false alarm rate. So they have a bright future in theapplication.