BVIFormer: a binocular-vision-inspired transformer with binocular competitive fusion for single-image restoration

Human vision remains robust in adverse weather by combining foveated processing with binocular integration. Inspired by this mechanism, we propose BVIFormer, a binocular-vision-inspired Transformer for single-image restoration. BVIFormer adopts a multi-scale encoder-decoder architecture and a dual-branch building block that emulates two-eye perception. Specifically, Parallel Deformable Embedding (PDE) produces two complementary embeddings, Human-Vision-Inspired Attention (HVIA) refines them with a fovea-periphery weighting strategy, and Binocular Competitive Fusion (BCF) performs rivalry-like fusion via softmax-normalized, per-channel competitive gates. We apply BVIFormer to single-image dehazing and deraining. On Dense-Haze and SOTS-Outdoor, BVIFormer achieves 17.30 dB and 37.56 dB PSNR, outperforming the previous best results by 0.68 dB and 0.14 dB, respectively. On Rain1400 and Test2800, it attains 33.82 dB and 34.24 dB PSNR. In addition, we provide visual and statistical analyses of branch behaviors and competitive gating, confirming that the two branches capture complementary information and that BCF adaptively selects reliable cues under different degradations. Code is available at https://github.com/LindaLi113/BVIFormer .