In this paper, we propose a novel method to determine upside whether a scanned text document is right side up or down. The text documents discussed here are limited to English, Chinese and Japanese where we find that the punctuation much marks located on the bottom of the text line have a more frequent occurrence than those on the top. Thus, by calculating the the number of punctuation marks on the bottom and top, the orientation of documents image can be detected. The experimental results demonstrate the effectiveness of the proposed method on 683 Chinese, English and Japanese document images. In the text only documents, 98% accuracy of orientation detection is achieved on the documents in three languages with higher performance in Chinese office document image. And even in office documents including tables and pictures and without text segmentation, 87.11% accuracy could be achieved in English documents, 88.52% in Chinese documents and 83.89% in Japanese documents.