An Innovative Network based on Double Receptive Field and Recursive Bi-directional Long Short-Term Memory

Abstract

\n In the scenes of character recognition, this paper studies the influence of network structure and parameters of “An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition (CRNN) ” on the recognition results in detail, and proposes an improved CRNN network based on double receptive field and recursive Bi-directional Long Short-Term Memory(BILSTM), which is named “An innovative network based on double receptive field and Recursive Bi-directional Long Short-Term Memory༈CRNN_RES༉”. In the CRNN_RES network, the innovations of this paper are adjusting the structure of CNN to enhance the feature extraction ability of the CNN network and using the shared parameter BILSTM network with recursive residuals to reduce the number of network parameters and improve the accuracy of the model prediction. In fact, the number of parameters of CRNN_RES network proposed in this paper is 7148325, which is 1182976 fewer than that of CRNN. On the same open datasets:ICDAR 2003༈IC03༉,ICDAR 2013༈IC13༉༌IIIT 5k-word༈IIIT5k༉, and Street View Text (SVT), the proposed method achieves 96.90%, 89.85%, 83.63%, and 82.96% recognition accuracy, which are higher than that of CRNN 1.40%, 3.15%, 5.43% ༌and 2.16%.