Abstarct: Nowadays, extraction of text in natural scenes is addressed by many researchers. Scene content can be classified into two important categories: perceptual content and semantic content. Perceptual content includes color characteristics, shape and scene texture. semantic content includes text, human’s face, behaviors and actions. Among various information of scene, textual information has more importance, because they are intelligible by human and computer, and provides possibility of describing the scene content. In this thesis, we propose a method to extract text regions from images of natural scene with complex background. The algorithm contains four main steps. At first step we extract candidate regions for text using gradient cue. At next step we cluster extracted regions according to this fact that the consisting features of a text row in an image have approximately same height and direction. Finally we exploit features of histogram of gradient magnitude and gradient orientation in extracted regions to filter out non-text regions. For this purpose, we train a classifier baxsed on support vector machine (SVM). This classifier is trained by histogram features of gradient magnitude and gradient orientation of text regions and non-text regions. Then, the results will be improved by determining distance criterion baxsed on width of detected textual regions and horizontal projection. The evaluation results obtained from applying the proposed method on scenes which have English and Persian text with different fonts and simple and complex background. According to evaluation of experimental results of text detection on three datasets: ICDAR 2003/2005, Microsoft Street View Text Detection and a collection of Farsi text images. The proposed text detection method can cope with variation in type of font, size, color and slightly direction. In comparison to other methods this result is very promising.