In order to improve the target identification, Markov random field (MRF) is introduced in the target fusion process to build prior probability model of a class. Then aiming at selecting model parameter β, an EM-MAP-HMRF feature-level fusion algorithm is proposed based on non-homogeneous class and direction. HMRF is divided into centric and distributed-based fusion schemes. The simulations show that the two new fusion algorithms can improve the classification accuracy, and enhance the ability to anti-interference. However, they have different advantages. The two new schemes can be used in various fusion systems for different applications and improve the effectiveness of detection and identification for specific targets.