Swin Transformer-based Cross-view Attention Network for Stereo Image Super-Resolution
Xue Li, Hongying Zhang, Laiping Zhang, Zixun Ye, Juntao Pu, Mingdong Yuan · 2023
Stereo image pairs offer complementary information that can improve the performance of image super-resolution reconstruction significantly. However, integrating longrange dependencies between stereo image pairs into super-resolution is challenging. This paper proposes a Swin Transformer-based cross-view attention network for stereo image super-resolution (TCASSRnet) to address this issue. Firstly, TCASSRnet extracts stereo image pair features using a Transformer with a shallow residual block (SRB). Next, a cross-view attention model (CAM) is introduced to integrate complementary features of stereo image pairs, resulting in high-frequency and low-frequency attention maps. Finally, deep residual blocks (DRBs) are employed to restore image details layer by layer and reconstruct super-resolution images. Extensive experimental results demonstrate the effectiveness of TCASSRnet. Specifically, TCASSRnet outperforms the baseline network (PASSRnet) by 0.55 dB in terms of PSNR on the Middlebury dataset for 4× SR, while the parameter size of TCASSRnet is reduced by 50% compared to the recent SSRDE-Fnet.