Researchers have developed a lightweight mathematical module that helps AI reconstruct lensless camera images while staying closer to the measurements the sensor actually recorded.
The study, published in Optics Express, addresses a persistent problem in computational imaging: a neural network can produce an attractive picture that does not fully satisfy the camera’s physical measurement model.
The team’s output-constraint module adjusts a reconstruction through an analytical projection. This calculation avoids the repeated refinement steps used by iterative reconstruction methods, allowing the correction to accompany a single pass through a neural network.
In the reported experiments, the module reduced measurement inconsistency by approximately threefold compared with unconstrained deep learning. Inference was approximately nine times faster than the tested iterative optimization method, ADMM running for 100 iterations.
The researchers also reported the strongest reconstruction quality among the measurement-consistent single-pass methods they compared.
The results point to a practical way of combining neural imaging speed with physical constraints.
They also clarify an important distinction: matching a conventional camera photograph and matching a lensless sensor’s measurements are different objectives. The method improves the latter without requiring a lengthy iterative reconstruction process.