The traditional image recognition algorithm has only a single recognition model and is susceptible to the external illumination interference. In contrast, as for the deep convolutional network model, there exist a large amount of calculation and high cost although its recognition rate is high. An improved based compression algorithm is proposed based on the deep XNOR-network. The compositions of the weld recognition system and the classical convolution neural network model are first introduced. The improved convolution network compression algorithm is described, including the weight update algorithm and the weight compensation algorithm. The data experiments are performed on the self-made datasets and the simulation platform. The research results show that the proposed algorithm has the advantages of high recognition rate, small model, strong adaptability and diversity of recognition models, which can be applied to the weld identification in the welding site.