Binarization is an important step in o ptical character recognition (OCR),directly affects the accuracy of OCR.At prese nt,the local binarization algorithms based on luminance segmentation have good e ffect,complicated process and long elapsed time.The fast binarization algorithms are simple and noise sensitive.Generally,low-luminance images have nonnegligi ble noise and low contrast of text.In order to obtain low contrast text,fast bin arization algorithms need to be sensitive to luminance gradient.However,in the b inarization result,luminance gradient sensitivity also leads to nonnegligible ba ckground noise,textual breakage and loss.In this paper,for high-quality and fas t binarization,non-local mean filtering is adopted to suppress noise and avoid over-smooth.Improved Bradley algorithm is used to extract the low contrast text in order to solve the problem of textual breakage.In the end,dilation algorithm and erosion algorithm are used to suppress the noise of binarization.Our method is suitable for uneven low luminance pictures and uneven high luminance picture s.Experimental results show that our method performs the same as other fast bina rization algorithms under uneven high luminance,however,extracts more text with less noise under uneven low luminance,solves the problem of textual breakage.The OCR recall rate of the binarization results of this method reached 93.5%.