In order to solve the insufficient of large-scale data processing ability for intelligent robot, and improve the ability of image classification, one-shot learning method based on convolutional neural network is proposed in this paper. Firstly, a sample is augmented with data augmentation techniques to construct a small target dataset, so that the neural network can capture the key features of the target task as much as possible. Then a pre-training model that performs well on large-scale datasets will be transferred to target task more or less. Due to its existing pre-training model weights and underlying features of the image, the layer freezing technique can be used to freeze the simple geometric shape features. The training and fine-tune will only be performed on the fully-connection layer and the last few conventional networks, which preserve the combined features, finally achieving the identification and classification on the target sample set. The experimental results show that the method has high accuracy and flexibility in the single-sample classification problem, meanwhile, can save a lot of time and computing resources.