Attention Enhanced ConvNet-RNN for Chinese Vehicle License Plate Recognition

摘要

As an important part of intelligent transportation system, vehicle license plate recognition requires high accuracy in an open environment. While a lot of approaches have been proposed, and achieved good performance to some extent, these approaches still have problems, for example, in the condition of characters' distortion or partial occlusion. Segmentation-free VLPR systems compute the label in one pass using Long Short-Term Memory Network (LSTM), without individual segmentation step, their results tend to be not influenced by the segmentation accuracy. Based on the idea of Segmentation-free VLPR, this paper proposed an attention enhanced ConvNet-RNN (AC-RNN) for accurate Chinese Vehicle License Plate Recognition. The attention mechanism helps to locate the important instances in the step of recognition. While the ConvNet is used to extract features, the recurrent neural networks (RNN) with connectionist temporal classification (CTC) are applied for sequence labeling. The proposed AC-RNN was trained on a large generated dataset which contains various types of license plates in China. The AC-RNN could figure out the vehicle license even in cases of light changing, spatial distortion and partial blurry. Experiments showed that the AC-RNN performs better on the testing real images, increasing about 5% on accuracy, compared with classic ConvNet-RNN [8].