Scene Text Image Super-Resolution Reconstruction Based on Dual-Branched Sequence Residual Attention

Abstract

This paper proposes Dual-branch Sequence Residual Attention for Super-resolution Reconstruction Network (DSRASRN) to address the drawbacks of loss of detail information and edge blurring in text images reconstructed by existing scene text image super-resolution reconstruction methods. In DSRASRN, first, a Dual-branch Sequence Residual Attention Block (DSRAB) is adopted to obtain a more comprehensive and accurate representation of contextual information. DSRAB uses a dual-branch structure to extract horizontal and vertical context information and adopts an Efficient Channel Attention (ECA) mechanism to assign higher weights to more important information, thereby enhancing the expression ability of the captured features. Second, DSRASRN adds a Text Edge Awareness Block (TEAB) to enhance the processing of edge details and textures of text images. In TEAB, convolution kernels are applied to capture information in specific spatial directions, and a dilated convolution with different dilation rates to increase the ability to reconstruct high-frequency information is adopted. DSRASRN and four state-of-the-art reconstruction methods: TSRN, TBSRN, TG, and TPGSR are evaluated on the TextZoom dataset. Experimental results show that DSRASRN reconstructs text images with more detail and exhibits superior performance by achieving higher text recognition accuracy. Compared with the other four evaluated methods, DSRASRN improves Peak Signal-to-Noise Ratio (PSNR) by up to 0.27, 0.78, 0.59, and 0.51 dB, respectively. The average text recognition accuracies of ASTER, MORAN, and CRNN are 65.0%, 62.1%, and 52.0%, respectively. In addition, the generalization ability of DSRASRN was evaluated on the ICDAR2015 and SVT datasets. The experimental results show that DSRASRN achieves good generalization.