Noise reduction based on U-Net in single-pixel imaging using a banner image

A method for reducing noise in single-pixel imaging is proposed, using a banner image as a light source combined with a U-Net-based deep learning approach. In single-pixel imaging using banner images, uneven illumination introduces noise from the displayed characters, making gesture recognition difficult. We reduced this noise by applying U-Net, which is robust in segmenting noisy images. Noise can be effectively reduced by the use of U-Net. Even with 500 measurements, the gesture shape is clearly restored and the SSIM value is 0.8. This indicates the potential of U-Net for denoising single-pixel images illuminated by encoded banner images.