Abstract
Most modern imaging systems process the data they capture algorithmicallybefore-or instead of-human viewing. As a result, performance depends not on howinterpretable the measurements appear, but how effectively they encode detailsfor algorithmic processing. Information theory provides mathematical tools toanalyze this, but developing methods that can handle the complexity ofreal-world measurements yet remain practical enough for widespread use hasproven challenging. We introduce a data-driven approach for estimating theinformation content of imaging system measurements. Our framework requires onlyexperimental measurements and noise characterization, with no need for groundtruth data. We demonstrate that these information estimates reliably predictsystem performance across diverse imaging modalities, including colorphotography, radio astronomy, lensless imaging, and label-free microscopy. Toautomate the process of designing imaging systems that maximize informationcapture we introduce an optimization technique called Information-DrivenEncoder Analysis Learning (IDEAL). The tools we develop in this work unlockinformation theory as a powerful, practical tool for analyzing and designingimaging systems across a broad range of applications. A video summarizing this work can be found athttps://waller-lab.github.io/EncodingInformationWebsite/