Abstarct: Text transfers a lot of data to the viewer in the natural scene. The first step to understanding Contextual Information is auto search text in the natural scene. There are many important applications in this field, for examples, assistance to the tourists, who are not familiar with the local language, helping blind people to get informations in their surroundin and interpret the contents of a video. text localization and recognition in the natural scene is a problem that not be resolved perfectly. In recent years text localization and recognition in the natural scene, has attracted the attention of many researchers. In this thesis, an approach for persian text localization in natural scene images is presented . In this approach, text regions and too non_Text Regions detect using Extremal Regions (ERs) detector. The ER detector is robust to blur, illumination variation and low contrast images. After ER detection, in two classification stage, some of ERs are selected as text regions and some of them are removed. In the first classification stage, the probability of each ER being a character is estimated. if this probability is higher than a predefined value, ERs selecte for the second classification stage but non_text regions exist yet. Adaptive gradient edge detector will be used For eliminating non_text regions. Finally, regions that are almost in a line, are considered as text regions. The method was evaluated on some of collected data. These images are challenging ,blur, with complex background and low contrast that evaluation on these images are satisfactory.