Machine Learning in Coded Optical Imaging

In this chapter, we will go through the theory and application of machine learning in coded optical imaging, where the fundamental task is to solve the inverse problem that infers the desired visual data from coded measurement(s). With the rapid development of machine learning, different decoding approaches have been proposed to advance the progress of coded optical imaging. Taking the snapshot compressive imaging as a representative, we briefly introduce the three widely used schemes: conventional optimization framework imposing various nature visual priors to retrieve the target data in an iterative manner; plug-and-play scheme incorporating deep image priors into an optimization framework; end-to-end deep neural networks applying emerging network backbones for high-quality and fast decoding. All these approaches have their own pros and cons, and all keep developing for high performance, as well as addressing limitations in other aspects, such as efficiency, flexibility, scalability, robustness, etc. We hope this review can sketch the algorithm development on snapshot compressive imaging and inspire new research in other coded optical imaging approaches.