Channel attention convolutional recurrent neural network on street view symbol recognition

Abstract

Character recognition has always played an important role in the area of image processing.  In our paper, we focus on the recognition problem of street view symbol, and propose a novel network architecture called Channel-Attention-based Convolutional Recurrent Neural Network (CACRNN) for character detection and character recognition. Different from CRNN, our method is based on convolutional neural network (CNN) structure with the help of attention guidance, which is a new way of combining CNN, recurrent neural network, and SE block. CRNN can be learned directly from sequence labels without detailed annotations, and the recognized objects are not limited in length. In the network, CNN plays the role of text detection, and recurrent neural network have a significant effect on text recognition.  Among them, we use the long short-term memory (LSTM) network as the recurrent neural network. In addition, we introduce SE block and propose an effective attention layer, so that the model is able to dynamically pay attention to certain parts that contribute to the existing task. Then our model can determine the most relevant aspects, and achieve better effect of filtering useless information and removing noise. Finally, the entire network is optimized using the CTC loss function. Through experiments, we found that the recognition accuracy of CACRNN processing characters is higher than other methods.