Text Detection And Localization In Natural Scene Images

In recent years, the capacity of the digital images all over the world is growing at an alarming rate with the rapid development of multimedia technology and computer networks. Every day produce capacity equivalent to thousands of megabytes of images, the digital images contain a large amount of useful information. The information included in characters of images is very important in people’s daily life, for example, traffic signs, street signs, billboards, posters, book covers and so on. It is very helpful for automatic understanding of high-level semantics, image indexing and retrieval, which can automatic detect text in natural scene images. This paper presents a novel method for detecting text in natural scene images based on the relevant knowledge of applied mathematics, the following work:Firstly, Canny edge detection operator extract the edge map from original image with NiBlack operator, which can restrain the background and noise and separate the adjacent area of the characters. At the same time it prepare for the follow-up feature extraction of characters with their respective advantages. Secondly, we fill the edge image so that form connected components. Then we begin to analysis connected component. We remove part of the background or noise according to character features such as character ratio satisfies a certain proportion, stroke width similarity, characters with more angular points and longer perimeter and other features. Once again, the candidate text regions are colored compared with the original color image. According to the text row character will keep similar color, we use the K-means clustering to extract color feature and locate the candidate text regions. These candidate text regions are clustered into three categories. Because the text regions have more angular points than other regions, Harris corner detection algorithm can distinguish the text regions from non-text regions in these candidate regions by judging numbers of angular point in connected areas. Thereby we go further remove non-text regions and get character candidate regions accurately. Finally, because there are non-text regions in these candidate regions, we use a trained SVM classifier with the feature of HOG and LBP to distinguish the text region from non-text region in these candidate regions. Histogram of Oriented Gradient is got by calculating gradient direction histogram of local image areas and it is not sensitive to illumination changes and small offset. The Local Binary Pattern is able to describe the image texture features. Feature LBP extracting is simple and the effect is good, so it commonly used to image classification and recognition. Principal component analysis is to use a few representative variables to represent the original of the vast number of information, and these representative variables are independent of each other. Ultimately we get text detection and localization in natural scene images.This algorithm proposed in this paper can detect various characters and characters whose values are relatively low in the background. The experimental results show that localization effect is good, and it is robust against illumination, character fonts, complex background and not affected by the influence of other factors.