By jointly designing optics and algorithms, we can now create imaging systems that outperform traditional methods. Lens-based systems must form a direct image on the sensor, which constrains the form factors and degrees of freedom which they can record. In contrast, computational imaging systems can form indirect, coded observations on the sensor then decode the desired information algorithmically. This generalization of the image formation problem enables two key improvements. First, imaging hardware can often be simplified, relying on algorithms to overcome certain types of hardware shortcomings. Second, the optics and sensors can be viewed as encoders whose job is to optically preprocess information to be decoded algorithmically.

 

In this talk I will discuss my work creating lensless imagers that use random phase optics in lieu of a lens. I will then show that these cameras naturally fit into the framework of compressed sensing, capturing information (via random projections) that lenses destroy. This enables single-shot observation of additional optical degrees of freedom (depth, time, spectrum) without losing spatial resolution. Finally I will discuss engineering better pseudorandom optics and demonstrate their integration into a miniature microscope capable of video-rate 3D fluorescence imaging in a package smaller than a quarter.