| As an important carrier of information, text has played an important role in human life, so far from ancient times. In our daily life, text is all around us. For example, billboards, posters, license plates, road signs and leaflets, all contain masses of text. The semantic information embedded in text could help us understand the scene better. As a consequence, scene text detection has a great research value to lots of applications of computer vision, such as content-based image retrieval,assistive navigation, intelligent transportation and automatic geocoding. However,different from traditional document image, image in natural scenes always has complex background, and text in natural scenes may be diverse. Except that, some other interference factors are exist in the wild, for instance, non-uniform illumination,partial occlusion, blur and distortion, which bring great challenge to scene text detection. Though a rich body of approaches have been proposed in recent years,localizing the text region from scene image rapidly and effectively still have been the difficult issues in this field. This paper carries out study on scene text detection, and the main research contents in this thesis are as follows.Firstly, we investigate the opponent color theory and perform to extract Maximally Stable Extremal Regions from opponent color channels, which make up for the deficiency of the results in intensity images. We take the multi-channels combined by opponent color channel and intensity image for text detection, then we experiment the performance in different color channels.After the extraction of candidate text regions, we apply two classifiers, a text component classifier and a text-line classifier, sequentially to filter non-text regions.The kernel descriptors are then introduced for the representation of image feature, the gradient, color and local binary pattern are introduced to construct different kernel descriptors. With the kernel descriptors, we can train two classifiers with linear SVM.Considering that the importance of different descriptors to classification is diverse, we propose to learn the relative importance of descriptors with multiple kernel learning technique, the relative weights of different features can fill the gaps of multiple feature classification, and improve the performance of classification.We have experimented our proposed strategy on ICDAR 2003 and ICDAR 2011,the experimental results demonstrate its effectiveness and efficiency. We can achievegeneral equivalently good performance with several compared state-of-the-art methods, and the proposed method is insensitivity to some interference factors, like illumination, blur and so on. |
