Abstract: A method of image orientation recognition is put forward, based on HMM (Hidden Markov Model). Segmenting the target image to recognize into sub-images, a singular value decomposition is conducted for them to extract singular value vectors so as to form an observation sequence, i.e., the singular value vectors of the image are taken as HMM observation vectors. Then, the HMM parameters are determined with their maximum likelihood calculated, and the images are recognized according to the clustering which the maximum likelihood of an image to recognize corresponds to. Test results showed that the recognition rate of 150 target images in 3 clusters is up to 85%.