Uyghur Text Detection In Nature Scene Images

Text is an important carrier for natural scene image information. the character size, arrangement rules, different orientation of the text in natural scene, the font color changing, uneven illumination and occlusion caused by interference, which will give the text detection enormous difficulties. Unlike the English and Chinese characters,Uyghur has 32 letters and more than 120 characters which have varied form and complex structure,so it makes the detection of natural scenes Uyghur text more challenging.For the unique shape of Uyghur, the paper have depth study Uyghur texts detection on natural scene and proposed two text detection algorithms.then,constructing the Uyghur text detecting system based on two algorithms. The main research and contributions are as follows:Proposed a natural scene Uyghur text area detection algorithm combined with homogenization and baseline. First, through the homogenization to enhance the features of the text area; Then, the gray and texture features were extracted to take the text area; Finally, a special baseline characteristics of Uyghur was used to reduce interference complex background textures. Experimental results show that compared to using only edge detection or homogenization, the method not only detect text Uighur region accurately, but also suppress background interference effectively.Proposed a modified LBP features( ULBP texture-The Local Binary Patterns of Uyghur) extraction algorithm method for the Uyghur text feature extraction.The feature vectors obtained and used to classify the candidate Uyghur text area. The algorithm uses the homogenization of space to build a map of the image; Then, using the corner detection quick access to the candidate text area in the homogenization of space; And finally, ULBP feature extraction in the candidate text area where these features to import training after SVM can get accurate text area. Experimental resultsshow that, ULBP characteristics of the Uyghur text area has a high degree of differentiation, and can greatly reduce the false detection rate, also has strong robustness.