Single-pixel imaging reconstruction based on a complementary frequency-domain filter mask with classifier-free guidance

In single-pixel imaging, reconstructing high-quality images at a low measurement rate is a key goal. Currently, deep learning methods achieve this goal by optimizing the loss between the target image and the original image, which limits the potential of low measurements. Therefore, this study proposes a single-pixel reconstruction algorithm based on a complementary frequency-domain filter mask classifier model. We designed a regulation mask of complementary filters and combined it with the classifier-free guidance method to assist in high-quality image reconstruction. By leveraging the multi-dimensional information advantages of the frequency domain, the algorithm better restores high- and low-frequency details of the image. Experimental results show that at a measurement rate of 10%, the average peak signal-to-noise ratio of the complementary frequency-domain filter mask with classifier-free guidance on the MNIST dataset reaches 28.82 dB, demonstrating excellent performance across multiple dataset scenarios. Further exploration of the <i>ω</i> parameter adjustment scheme for single-pixel reconstruction tasks provides new ideas and references for applications in this field.