Abstract: This paper studied the traditional image text detection approaches and proposed a new image text detection method based on double-threshold gradient pattern with a pretty fast speed in both classifier training and implementing.Firstly,in the rough detection phase,the maximally stable extremal regions(MSER) was extracted as a candidate text area,to avoid scanning the whole image,greatly improving the detection speed and real-time.Secondly,in the feature extraction part of refine detection phase,in order to overcome the text area color contrast inversion problem and the problem of noise in natural image,this paper creatively presented a dual threshold gradient mode feature to describe the texture of the text area feature.Finally,to design the detector for text fine detection,this paper designed a new Cascade ELM(Extreme Learning Machine) detector by limit learning machine,which greatly shortens the classifier training time.The experimental results show that this method not only has excellent detection performance,but also greatly shortens the classifier training time and testing time.

Key words: Image text detection,Double-threshold gradient pattern,Extreme learning machine,Multiple classifier combination