Comprehensive Video Text Extraction Approach Based On Stroke Feature

The artificial text in the images and videos provide very important information about the content. Generally, they are the instruction, explanation or complement of the video content. Due to the existence of the Semantic gap, it is hard to extract the video information from the image directly. Therefore, the extraction and recognition of the video text are so necessary when the video and image fulfill the network in now days. Not only in the video retrieval and analysis field, but also in the IP interactive, it needs to extract the text from video accurately.The text extraction processing includes three procedures text detection, text localization and text segmentation. In this work, it applied the newly stroke feature based filter method in the text detection and localization steps. The original stroke filter response will be degrading or even failure when the character size mismatches the filter size. This work proposed two approaches to improve. The experiment results show the improving effect in filter responses.Based on the stroke feature, this work designed a series of comprehensive detection and localization methods. In the text detection step, applying the histogram non-linear transform and region labeling methods, they intensified the potential text regions and detected these regions by threshold. In the text localization step, it combined the stroke feature into the traditional localizing methods. Different from the based on edge feature projection analysis, this work utilized scalar quantization, DCT coefficients and half-image projection approaches to localize the regions with sufficient stroke feature and eliminate the background regions with complex texture. Compared stroke feature vector proposed in literature [26] with the pixel intensities feature vector, the SVM training achieved better performance under the stroke feature. And using this trained SVM classifier to refine the candidate text block by shrinking, combining and extending, it accurately localized the truth text regions in the image.On the localized text blocks, this work classified the pixels one by one utilizing the Gaussian Mixture Modeling. The training samples are generated by seed-fill and stroke-mode initial segmentation. It designed to combine the color features and text features to construct the sample vectors, where the color features are hue and intensity of pixels and the texture features are presented by wavelet coefficients. The experiments show the recognizing results of the text segmentation.