Text String Extraction from Scene Image Based on Edge Feature and Morphology

Extraction of text from scene image is much difficult than extraction from simple document image. A lot of researches succeeded in extracting single text string from image, but can not deal with image including many text strings. Meanwhile, the result may be mixed with noises be similar to text. This paper describes an algorithm that uses mathematical morphology to extract text effectively, and edge border ratio is utilized to differentiate text region from noise region, using the edge contrast feature of the text region in real scene. This paper also describes the method which can connect characters into text strings, and distribute text strings to different subimages according to their width of strokes. The algorithm is implied to scene image like signs, indicators as well as magazine covers, and its robustness is proved.