An approach for document orientation detection and classification by using support vector machine (SVM) theorem is proposed in this paper. First, all the characters in a document image will be isolated and some valid ones are selected. Using the valid characters, the document image will be vectorized to a 32-dimensional vector by the feature extracting. By training lots of samples, an SVM classifier can be obtained,and then the orientation of unknown document images can be classified. Experimental results show the accuracy of the proposed method is considerably higher than Bray Curtis distance, even for some bad samples.