Text Location Based On Corner And Color In Natural Scene Image

In the scene of natural there are lots of graphics and images which contain a wealth of text information. This information is very important for us to describe and understand the natural scene and it plays an important role in our life. As the bottleneck of the promotion of natural scene character recognition, improvement of text localization technical is necessary.This paper focuses on the method of natural scene text detection and localization method. Corner detection based natural scene text localization methods, color segmentation and clustering based natural scene text localization is used in this paper. A method combined the above two algorithm is also used to locate the text. Stroke and periodic of texture are join in the method to improve its performance.Because of massive existence of corner in text area, this paper use the comer feature to the text localization and use SUSAN corner detecting algorithm to find the corner. Firstly, detecting the corner point in natural scene images, and then using the templates to classify the corner point, finally using the Euclidean distance based clustering to obtain the text area. This algorithm in complex background text has a better positioning capability.By analysis the distribution of the colors in the natural scene image, the second location method is based on color space segmentation and color clustering. First, the color of natural scenes is divided into several categories in accordance with certain rules, each category can be conversed to a sub-image, then use the connected component analysis method to extract the text area, and finally combine the text area of the sub-image into the final result of the text area. This method is mainly for the initial segmented contingent image.Each method has its own insignificance, so this paper find a new method of text localization that using the above two methods of text localization algorithm. This method is the combination of rough text location base on corner detection and clustering, better text location based on color segmentation and clustering, selecting based on stroke texture information and periodic of texture. Finally get the precise and smooth text region by the selecting based on prior experiment. Excise results show that this method of natural scene text detection has a better positioning accuracy and robustness.