Research On Text Localization Based On MGA

With the rapid development of computer and multimedia technology, images become one of the most popular digital storage media, and the identification and classification of the image content has been paid more and more attention. Because many images contain rich text date that reflects the main content of the image and provides an important clue to understand the image, automatic extraction and recognition of text is extremely useful for image semantic understanding and indexing. However, due to the complex background of real world images such as license plate numbers, store name, commodity labels, etc, OCR (Optical Character Recognition) system cannot achieve an ideal effect. It is necessary to locate the text regions and generate a binary image before the OCR system. Therefore, text localization in images with complex background has become a challenging and important study, which is the key step of text extraction and recognition.Multiscale Geometric Analysis (MGA) plays more and more important role in image processing which is a significant study achievement following wavelet analysis in the field of signal processing. In this paper, we aim to locate the text from images with complex background with the advantage of MGA, such as Nonsubsampled Contourlet Transform (NSCT), Discrete Shearlet Transform and so on. The main works are as following:Firstly, we introduce the basic theory and scheme, the state of art and the development of text localization in images. We classify and discuss some relative methods, and the method of text localization combining MGA with other technologies is mentioned emphatically.Secondly, the principles and evaluation criterion of the optimal sparse representation and MGA are narrated. With the study of the theoretical framework of some important MGA methods such as Ridgelets, Curvelets, Contourlets and Shearlets, the main properties and performance of each method are discussed, which is useful for the choice of the text localization tools.Thirdly, a method based on the nonsubsampled contourlet transform (NSCT) is presented for text localization. First, NSCT is applied on the image to decompose it into set of directional subbands in different orientations and scales. With the help of energy-conversion among subbands, text information will be more highlighted. Then, edge detection is used to filter out the background and recognize the edges in multi-directions. And morphological operations can connect the edges together to obtain the connected regions. In order to make better use of multi-direction feature of text, at each scale, the logical AND operator could obtain the coarse text regions; voting-decision is applied among different scales to improve the accuracy and robustness of the method. Finally, the final text regions are achieved with the help of heuristic knowledge. In this paper, precision rate, recall rate and run time are used as the standards to evaluate the algorithm, which are base on the principle of rectangle matching.Finally, because the Discrete Shearlet Transform has the characteristics of simple structure, flexibility on directional selectivity and lower computing complexity, a method based on the Discrete Shearlet Transform is proposed for fast text localization. Using the shearlet transform with translation invariance, an original image is quickly decomposed into subbands in different orientations and scales, which have rich edges. Under our algorithm, a better result of non-text information removal could be obtained with the help of dynamic thresholding rather than simple edge detection. So the precision rate and recall rate of localization is significantly improved. The experimental results show that the proposed method has good performances.