An Improved Deep Neural Network Based on a Novel Visual Attention Mechanism for Text Recognition

Abstract

Text recognition from images captured by handheld mobile devices has attracted considerable research interest because of its commercial applications. The state-of-the-art printed text recognition methods are often based on attention mechanisms. However, these methods perform poorly on images captured due to poor illumination conditions, blur, noise, and low resolution. To address these unfavorable conditions, we propose a new text recognition method based on an encoder-decoder model. Particularly, we present a novel attention mechanism using a multi-scale cascade fashion combined with a channel attention gate module. Our model is also strengthened by an EfficientNet-like backbone. Extensive experiments on three popular datasets, including SROIE 2019, B-MOD, and CORD, show that our proposed method outperforms the baseline attention mechanism and achieves competitive accuracy compared to other state-ofthe-art approaches.