Abstract: Pattern recognition methods of a large number of pattern classes are studied A dynamic multi branches algorithm is presented, which can implement knowledge-increasable to solve the problems of complex or large quantities of classes learning and dynamic extension of system It can reduce object scope by using the divide-and-conquer principle; it can improve recognition rate by using combination classifiers; it can learn as human brain does The algorithm can not only accelerate calculation speed and improve recognition rate but also be extended easily and freely It can be used in pattern recognition such as faces, fingerprints, and characters The constructive method of the system shows the generality The capability analysis validates the practicability of the algorithm and the experimentation result proves its reasonableness and feasibility