Research And Implementation On Scene Text Localization

The use of text, as the means of information transmission, and image, as the carrier of text, spreads over many the scenarios, such as network media and street signs, with the explosive information available today. So, how to effectively retrieve and understand such huge amount of information has become an important research direction in computer vision. And, scene text localization, as the foundation of content-based information retrieval, of course has attracted considerable attention from researchers.However, due to the complexity of the scene and the diversity of the text itself, scene text localization remains quite challenging. In this paper, we concentrate on how to locate text from natural scenes. The main work focuses on the following aspects.Firstly, we propose an algorithm for scene text preprocessing based on constrained stroke width transform. According to the text of this particular object, we eliminate the opened curve and use the Hough transform to eliminate the straight line segments on the basis of traditional edge detection. Then, we improve the operation of stroke width transform(SWT) by considering the color consistency in local neighborhoods and the gradient direction consistency, and propose an algorithm called stroke width transform with local neighborhood consistency constraints. And then, we limit the stroke width by a threshold and process with the unilateral line. In the end, we get an output image of stroke width transform.Secondly, we propose an algorithm to construct character connected components based on the character features, which filters out lots of non-text regions. Firstly, we connect the pixels into candidate character connected components based on consistent stroke width. On this basis, we generate the candidate connected components. In order to filter out the correct connected components, we extract the low-level features of character, construct a high-level feature of characters based on the analysis of the structure and distribution characteristics of the characters. The high-level feature include: 1、feature of character connected components based on the gradient and the characteristics of binary pattern;2、histogram of contour variation based on chain code. Then, we train the character-level SVM classifiers and get the final character connected components.Thirdly, in order to further filter out the non-text regions, we propose an algorithm for scene text localization based on the text-line features. With the output of character connected components, we connect the connected components of characters into candidate text-lines according to the distribution characteristics of scene text and the relationship between characters. Then we construct text-line features based on Intra-class Dispersion with the description of the commonality between the texts. Also, we construct text-line features based on statistical information, which includes:1、text-line features based on texture and gradient information;2、text-line features based on histogram of step times. In the end, we cascade the two categories of features, and train the text-lines level SVM classifier to filter the candidate text-lines, and get the final text output.To better evaluate our algorithm, our proposed method is tested and evaluated on the ICDAR 2005, ICDAR 2011 databases and our own Chinese database, and is compared with other methods. The experimental results demonstrate that our algorithm is effective and can do text localization in the scenes.