Residual CRNN and Its Application to Handwritten Digit String Recognition

Abstract

Recently the Convolutional Recurrent Neural Network (CRNN) architecture has shown success in many string recognition tasks and residual connections are applied to most network architectures. In this paper, we embrace these observations and present a new string recognition model named Residual Convolutional Recurrent Neural Network (Residual CRNN, or Res-CRNN) based on CRNN and residual connections. We add residual connections to convolutional layers as well as recurrent layers in CRNN. With residual connections, the proposed method extracts more efficient image features and make better predictions than ordinary CRNN. We apply this new model to handwritten digit string recognition task (HDSR) and obtain significant improvements on HDSR benchmarks ORAND-CAR-A and ORAND-CAR-B.