Aiming at the small sample water meter reading dataset polluted by strong noise

a character recognition method based on convolutional neural network（ CNN） for transfer learning is proposed. The CNN is built in the Tensor Flow framework

and the batch standardization is introduced to construct the network model. The objective function is optimized to improve the network performance. The recognition rate of small sample datasets is improved by transfer learning. Firstly

the structural parameters of CNN are pre-trained by large sample dataset.Then

the parameters of the shallow network layers are shared. The parameters of the deep network layers are trained by the small sample dataset to complete the classification of characters. The experimental results show that compared with the traditional VGG—16 network

the recognition rate of the designed network is increased by2. 19 %. The velocity is increased by 98. 3 %. The computational complexity and memory consumption are also greatly reduced. It is possible for the remote automatic water meter reading.