Approaches To Text Information Extraction In Natural Scene

There are not only many pictures but also rich texts containing important information in natural scene. These texts are of great value to the description and understanding of scene contents and can be used as the vital clue to indexing and retrieving. Therefore, an automatic tool is seriously needed for text cognition in natural scene for retrieving, querying, browsing and understanding scene contents in order to increase the efficiency of scene image management.Text cognition in natural scene is performed to recognize and understand the text with little or no restriction. Considerable success has been achieved in traditional document analysis. However, in general, it can handle text only in the printed document, not in natural scene. Therefore, the study on the algorithm of text information extraction (TIE) in natural scene becomes a hot issue.This dissertation studies the related techniques for text cognition in natural scene, with the emphasis on TIE. TIE in natural scene refers to the determination of the presence of the text, confirms the location of the text region and segments the characters from the background. By means of characteristic analysis of text, a TIE method based on fuzzy homogeneity is proposed in this dissertation. A learning-based text detection method and a text location method based on text pixel density are introduced to determine the location of the text in an image and generate the bounding boxes around the text. A text extraction method based on multi-scale transform and template matching is employed to segment the text from the background. Compared to other approaches, it has following advantages. Homogeneity is largely related to the local information extracted from an image and reflects how uniform a region is. It is believed that the fuzzy set theory is a suitable and critical tool for handling the image with complex background because of the uncertainty associated with the vagueness and impression of the human vision. To reduce the difficulties of extracting the text in a complex color image, multi-scale transform and template matching makes full use of the structural characteristics. This dissertation includes the following aspects:1. This dissertation studies space mappings, introduces homogeneity mapping to text detection and proposes the novel definition of homogeneity mapping. The characteristics of the text region can be enhanced by the space mapping to improve the performance of text detection. Several space mappings are compared to demonstrate that the homogeneity mapping possesses the best divisibility and lay the foundation of text detection.2. A text detection approach based on fuzzy homogeneity is proposed. On the basis of homogeneity, fuzzy homogeneity mapping is defined utilizing fuzzy set theory to reflect the local uniform of a region. The proposed method is applicable to the text detection in a complex image, especially, with the background similar to the text region.3. The range and position of the text should be known to extract the text from the text region. A text location method based on text pixel density is proposed here. Characteristic of the text and geometric information is applied to the selection of the candidate region. Multi Resolution Analysis (MRA) is employed to solve the problem caused by the various sizes of the text. Range and position of the text is confirmed by the fusion of the text locations based on MRA. The proposed method is tested on the ICDAR'2005 datasets. Experimental results show that the proposed text location method has better performance.4. A text extraction algorithm based on multi-scale transform and template matching is proposed in this dissertation. Traditional character segmentation scheme cannot segment characters in an image correctly due to the variances of text. First, the normal and existent template of the text region and multi-scale transform are defined to describe the variances of the text region. Then, template matching algorithm based on multi-scale transform is proposed for text extraction. The proposed method is applied to vehicle license plate (VLP) location and character segmentation. Experiments show that the algorithm has good performance for text location and character segmentation with fixed format, especially, when the characters are broken or characters are merged, and it has high noise immunity.Dealing with both the artificial text and scene text, the proposed method is initially studied on general TIE method in natural scene and several achievements have been obtained. TIE in natural scene illustrates a good prospect of application whether it is applied to the intelligent man-machine interface or to the services of content-based image retrieval.