Abstract:

Extracting text information from images captured in natural scenes is helpful for the content analysis of images. In this paper,according to the fact that the texts in images is often salient in local regions,a novel visual saliency model with multi-scale bounding box is proposed,based on which a new method combining the edge and texture information is designed for the candidate text detection. In this method,first,Lab color space is used to construct the edge and textural information-based image homogeneity,and by using this characteristic,the image is mapped into the homogeneity domain. Then,the proposed model is employed to generate average homogeneity images. Finally,the weighted Euclidean distance between the homogeneity image and the average homogeneity image is determined,and is taken as the saliency measure to extract text regions. Experimental results of natural scene images show that,as compared with the text detection methods based on the edge or the homogeneity,the proposed method can better restrain the background noise,which helps to further segment the text regions from the background and achieve more accurate text location.

Key words: text detection, visual saliency, homogeneity, image segmentation