WaveDM-FSPI: a wavelet-based conditional diffusion model for Fourier single-pixel imaging

Fourier single-pixel imaging (FSPI) enables fast image reconstruction at low sampling rates. However, the absence of high-frequency components usually leads to blurred images. To address this issue, a wavelet-based conditional diffusion model for Fourier single-pixel imaging (WaveDM-FSPI) is proposed. This method fully integrates the frequency characteristics of FSPI and the generative ability of diffusion models, significantly improving image reconstruction quality at low sampling rates. Specifically, we first use the four-step phase-shifting method to obtain the low-frequency Fourier spectrum of the target image and reconstruct the initial image. Subsequently, a lightweight spectrum recovery module (SRM) is introduced to perform preliminary high-frequency enhancement on the image. In order to further optimize the reconstruction effect, the image is decomposed using wavelet decomposition. The conditional diffusion model is used to reconstruct the approximation coefficients; simultaneously, the high-frequency recovery module (HFRM) is used to restore the detail sub-bands. Finally, the high-quality image reconstruction is achieved through end-to-end joint optimization. Experimental results demonstrate that the proposed method significantly improves image quality at a sampling rate of 5%, achieving superior performance across multiple datasets.