Abstract:  In order to solve the problem of low precision in traditional text localization algorithm, a robust text
localization method based on visual saliency and edge density is proposed from the perspective of natural scene
text properties. Firstly, spectral residual theory is used to extract the saliency region of the image, then search
for the regions of large edge density in the saliency regions and construct candidate connected regions, using a
small amount of prior information to filter out non text area. The experimental results on the standard data set
show that this method can obtain 70% comprehensive detection rate compared to the method of text region
detection using edge feature.

Key words: visual saliency, spectral residual,  text detection, edge density