DO-CGI: deep-optimized illumination patterns for computational ghost imaging at low sampling ratios

Computational ghost imaging (CGI) reconstructs objects from known illumination patterns and bucket-detector measurements, but quality deteriorates at low sampling ratios (SRs).

We present a deep-learning framework that optimizes grayscale diffuser patterns before reconstruction.

In simulations using CIFAR-10 and MNIST images with Split Bregman reconstruction, the learned patterns outperform random patterns in peak signal-to-noise ratio and structural similarity, including at SRs below 5\%.

Patterns trained on CIFAR-10 also transfer to MNIST and remain effective under moderate perturbations of the sensing matrix.

These results support learned pattern design as a route to fewer CGI measurements.