Autoencoder for Single-pixel imaging

Single-pixel imaging (SPI) as a new type of computing imaging technology, has important application value in the fields of biomedical images, long-distance imaging, remote sensing and so on. Limited by imaging mechanisms, high-spatial resolution single pixel imaging is inefficient. Obtaining high-quality imaging at low sampling rates is a hot and challenging issue in the field of single pixel imaging. In recent years, artificial intelligence methods represented by deep learning have been widely used in single pixel imaging. Generally, their excellent performance is imputed to their ability to learn realistic image priors from a large number of example images. However, how to get sufficient training sets in many practical applications, e.g., long-range targets single-pixel imaging is very difficult. Here, we introduce and motivate a new training principle for unsupervised learning of autoencoder (AE) based on the single pixel imaging physical model. This approach can be used to train autoencoders, and these denoising autoencoders can be stacked to initialize deep architectures. Comparative experiments clearly show the surprising advantage of AE in low sampling rate of time-varying modulation patterns.