Abstract: This paper proposes a novel method to segment document image based on adaptive feature and multiple phase feedback (AFMPF) model, which is adaptive to the blurring of document image and various patterns of mixed Chinese/English document. First, the whole process of segmentation is divided into several phases and each phase is allocated a primary feature to segment document image. Second, the segmentation result of each phase is fed back to the current or previous phase, and directs the next phase segmentation. This method causes good and effective combination among character segmentation, pre-classification of Chinese/English and character recognition, which improves greatly the accuracy of character segmentation and recognition.