On the use of prior knowledge in deep learning algorithms

Abstract

Machine learning algorithms have seen increasing use in the field of computational imaging. In the past few decades, the rapid computing hardware developments such as in GPU, mathematical optimization and the availability of large public domain databases have made these algorithms, increasingly attractive for several imaging problems. While these algorithms have exceeded in tests of generalizability, there is the underlying question of whether these \black-box" approaches are indeed learning the correct tasks. Is there a way for us to incorporate prior knowledge into the underlying framework? In this work, we examine how prior information on a task can be incorporated, to more eciently make use of deep learning algorithms. First, we investigate the case of phase retrieval. We use our prior knowledge of the light propagation, and embed an approximation of the physical model into our training scheme. We test this on imaging in extremely dark conditions with as low as 1 photon per pixel on average. Secondly, we investigate the case of image-enhancement. We take advantage of the composite nature of the task of transform a low-resolution low-dynamic range image, into a higher resolution, higher dynamic range image. We also investigate the application of mixed losses in this multi-task scheme, learning more eciently from the composite tasks.