A new algorithm of feature matching is proposed after balancing analysis of adjacency matrix of the matching model in a probabilistic framework. A probabilistic model is established and solved using Random Walks with Restart(RWR). Then a balancing analysis to the adjacency matrix of RWR is taken, and an efficient method for bidirectional balance is presented. The approach considers not only all the interaction of the two candidate feature point sets and the weight of each relevance, but also the balancing of all relevance weight. It improves the discriminative and accuracy performance of matching. The experimental results confirm that the method is robust to outliers and geometric deformation, accurate in terms of matching rate in various matching applications, and robust and practicable in the object location and the object recognition.