Accurate stereo image super-resolution using spatial-attention-enhance residual network

Stereo images can improve the performance of super-resolution (SR) by providing additional information from another viewpoint. However, the existing CNN-based stereo SR methods guide the reconstruction of high-frequency features in an indirect way, which hinders the network representation. In order to solve the issue, we firstly introduce spatial attention mechanism into stereo SR and propose the corresponding spatial-attention-enhance module (SAEM). The SAEM can capture spatial-wise feature correlations and directly guides the high-frequency feature reconstruction in the spatial dimension. This paper presents a novel spatial-attention-enhance super-resolution network (SAESRnet) for stereo images. The network representation is enhanced by SAEM, as extensive experiments show that our SAESRnet can achieve better accuracy and visual improvements against other existing stereo SR methods. Our method can outperform PASSRnet by 0.30 dB, 0.26 dB, and 0.26 dB respectively in the term of PSNR on Middlebury, KITTI2012, and KITTI2015 test datasets. In addition, the results of experiments also prove that our SAEM can also be possible to have a positive effect on improving the performance of single image super-resolution (SISR).