Car plate character recognition using a convolutional neural network with shared hidden layers

In character recognition, convolutional neural network (CNN) outperforms most of the other models. However, to guarantee a satisfactory performance, CNNs usually need a great number of samples. Due to the differences between the Chinese and the alphanumeric characters, the most common way to recognize the two classes is to use two independent CNNs respectively. In this paper, to solve the problem of the Chinese character shortage, we implement a CNN model which has shared hidden layers and two distinct softmax layers for the Chinese and the alphanumeric character respectively. To avoid over-fitting problem in the training process, the early stopping rule is employed. By training and testing on our two small-scale car plate character databases, the model gets a 9.289% and 9.632% relative reduction in test errors for the Chinese and the alphanumeric characters respectively over conventional CNN models.