Abstract: In order to ensure that the recognition system of tire code has preferable recognition correct rate and confidence level, the tire code recognition method was studied. In this method, various kinds of characteristics were extracted among a great deal of farm vehicle tire code character images. Regarding every kind of characteristics as the BP neural network input value, an individual BP neural network was set up and trained. After combining these sub-BP neural networks into a combination neural network, the recognition results were obtained conforming to the averaging or voting rule. The connection numbers of assembled neural network are far less than that of traditional huge neural network, and its training and recognition speed is also faster than that of traditional neural network. A lot of experiments of the tire code sample recognition show that the recognition correct rate and confidence level of assembled neural network are higher than those of the traditional neural network.