Deep learning enables parallel camera with enhanced- resolution and computational zoom imaging

Abstract High performance imaging in parallel cameras is a worldwide challenge computational optics studies. However, the existing solutions are suffering from fundamental contradiction between field of view (FOV), resolution and bandwidth, which system speed FOV decrease as scale increases. Inspired by compound eyes mantis shrimp zoom cameras, here we break these bottlenecks proposing deep learning-based (DLBP) camera, with an 8-μrad instantaneous 4 × at 30 frames per second. Using DLBP snapshot 30-MPs images captured fps, leading to orders-of-magnitude reductions complexity costs. Instead directly capturing photography large scale, our interactive-zoom platform operates enhance using learning. The proposed end-to-end model mainly consists multiple convolution layers, attention layers deconvolution layer, preserves more detailed information that image reconstructs real time compared famous super-resolution methods, it can be applied any similar without modification. Benefiting additional drive optical component, camera provides unprecedented-competitive advantages improving response (~ 100 ×) over comparison systems. Herein, experimental described this work, novel strategy solve inherent among FOV, bandwidth.