Abstract:  To extract diversity image features，an improved convolution neural network structure is proposed．The network introduced a multi-layer recursive neural network． Firstly，the network learned the shallow features of input images with convolution neural network，and then learned high-level features through the convolution neural network and multi-layer recursive neural network at the same time． Finally input two kinds of high-level features fusion into the classifier． This paper used the theory of transfer learning to solve the problem of the small sample training set data，and applied the improved convolution neural network structure to the oil pipeline steel grade identification． The experiments explore the influence of recursive neural network number and convolution kernels number separately，and make sure the final structure． The results show that the improved convolution neural network gets a low error rate of 3% ．

Key words: convolution neural network, transfer learning, recursive neural network, small sample set