Learned partial transform ensembles for exceptional optical compressive sensing

In most sensing processes, samples are acquired from a partial ensemble of measurements taken from a general transform defined by physically variable parameters. To follow the principle of parsimony, it is often desired to take the most compact ensemble of measurements possible. Compressive Sensing (CS) addresses this issue by significantly reducing the number of samples required in imaging systems. However, choosing the optimal set of physically realizable samples for CS can be challenging. In this work, we propose a novel Deep Learning (DL) method that jointly optimizes physically realizable CS matrices together with the reconstruction algorithm. Our method achieves unprecedented levels of optical compression, a few orders of magnitude higher compression ratio than what is typically achievable with classical CS techniques. We demonstrate the effectiveness of our approach through results on face imaging using as few as ten samples. Our method has the potential to significantly improve and enhance a wide range of imaging and sensing modalities.