Spatially variable resolution single-pixel imaging reconstruction based on diffusion transformers

Single-pixel imaging (SPI) achieves efficient image reconstruction through compressive sensing theory, but it struggles with significant degradation in reconstruction quality under low measurement rates. This paper proposes a spatial variant resolution module (SVR)-guided diffusion transformer (DiT) network for SPI reconstruction (SVR-DiT-Net). The framework adopts a two-stage cascaded architecture. Initially, a spatial variant resolution module (SVR) is designed, employing a progressive pixel-sharing mechanism to allocate measurement resources dynamically. Combined with a center-high-resolution-prioritized differential loss function, this enables high-fidelity reconstruction of critical regions under a limited number of measurements. Subsequently, for the first time, to our knowledge, diffusion transformers (DiTs) are introduced to SPI, proposing a dynamically parameterized modulation-based image-conditioned guidance mechanism. By employing the preliminary reconstruction image generated by SVR as conditioning information, the mechanism leverages the global self-attention of transformers to model long-range dependencies. This guides the iterative denoising process of the DiT to generate a high-quality final image. Experimental results demonstrate that our method achieves significant improvements in both global and regional image quality assessments, particularly within visually critical areas. Our framework provides adjustable reconstruction quality for different regions, offering a new, to our knowledge, perspective for research in generative model-based single-pixel imaging.