SSS-Net for Scene Text Recognition

Abstract

CRNN has made extraordinary achievements in the field of natural scene text recognition, but there are still some cases of insufficient recognition accuracy. Aiming at the problem that similar characters in horizontal text are difficult to distinguish, Strip-Squeeze-Spatial-Network (SSS-Net) based on the attention mechanism is proposed. The key point of our proposed SSS-Net is SSS Block. It mainly includes strip convolution block, squeeze attention block and spatial attention block. Add the module directly to DenseNet, and connect with CTC to form a text recognition network SSS-Net. We conducted verification in four datasets: ICDAR2003, ICDAR2013, SVT and SCUT_FORU. The experimental results show that the accuracy of this method is better than the CRNN structure based on DenseNet as the backbone feature extraction network, and the parameter amount is only about 1/4 of CRNN.