Abstract: A method for unsupervised text location in color image is presented.The method makes a brightness grading in R、G、B color layers of a color image separately to avoid choosing the number of clustering in common methods witch based on color clustering，and decreases the complexity of the background.Considering obvious directionality and color stability of text strokes，a rough text location is proceeded according algorithm of orientation gradient.And then，precisely discriminating is implemented with a multi-class SVM classifier.The new method constrains the types of candidate areas and depressed the difficulty of training SVM classifier.Those make the new method higher accuracy and robustness.