An efficient method for text detection from indoor panorama images using Extremal Regions

Text detection in complex real images, such as panorama images, remains great challenging in Computer Vision. A general method often focuses on the small test images with single background which makes it easier to do the detection and recognition. In this paper, we find a novel approach, as it can automatically deal with the indoor panorama images which contains distortion and illumination problems to extract the multi-scale trademark. Our method fuses edge information, color probability detection and geometric characteristics to segment the text and non-text part, and exploits Extremal Regions (ERs) which is robust to blur, illumination, color and texture variation to deal with low contrast text and find the accurate localization. Effectiveness of algorithm has been discussed in the experimental result section where the performance has been compared for different number of feature used.