Research On Text Extraction From Image With Complex Background

Text in the images and video contains rich semantic information, which is an important clue to understand the contents of the images and video. For content-based image information retrieval system, extracting the text effectively is an important technology. As the text in the images and video is usually superimposed on the complex backgrounds, the interference of the background components makes the current OCR technology difficult to recognize them, thus confining the successful application of OCR technology. Therefore, how to extract the text from the images and video with complex backgrounds is a problem which is needed to solve.For text detection in the images and video, the method combined with the multi-scale features of the wavelet transformation and marking algorithm was used. First, remove the image noise with method of wavelet semi-soft threshold; Second, obtain the high-frequency featured information of the image with the method of Mallat transformation, through which calculate the amount of characteristic energy of all components; Then, according to the characteristics of text in the images, use the newer dual-threshold method to detect the text from the backgrounds, achieving roughly position of the text; Finally, use the marking algorithm to position the text precisely, making the candidates of the text pixels to form the integrity of text region. The experiments showed that this method could detect the text in the images and video rapidly and accurately.After the text detection in the images, a segmentation method used by non-supervised parameter estimate to create GMM was used. First, a new edges-based sampling method was revised. In other words, on the basis of Canny edge detection, sample the pixels between the edges used by multi-scale scanning method; Then, unlike the previous features to create GMM, a HIS color space was created, and use the Hue as the features; Finally, segment the text rapidly with the established color model and the characteristics of the similar Hue, the same Gauss Possibility Distribution. Such a sample-possibility method could make the text segmentation completely automated, improving the efficiency of segmentation.Finally, in terms of character split and recognition, the dissertation proposed a projection method to split the single character. That is to say, scan the image through from top to bottom, from left to right, making the text in the images split into the single character which OCR system can use. During the process of recognition, the dissertation did not use the single structural feature or the statistical feature, but combine the advantages of both features, making every single character corresponded to 4-stroke directions, which were horizontal,vertical,neglect and restrain. Then calculate the grid stroke vectors as features, and use the most close neighbor method to classify the vectors in order to recognize the character. The experiments made by a large number of the images with the text showed that, this method could achieve a satisfying result of character recognition.