Self-supervised neural reconstructions for lensless imaging

Recent advances in lensless imaging reconstruction have primarily relied on supervised neural models trained using target images captured by lensed cameras via a beam splitter. However, we argue that using reference images from a different optical system introduces bias into the reconstruction process. To mitigate this issue, we propose a self-supervised approach that leverages data-fidelity guidance, similar to deep image prior, to train neural models for single-iteration lensless reconstruction. Through simulations and prototype camera experiments, we demonstrate that combining simple convex optimization methods with a denoising UNet improves perceptual quality (LPIPS), accelerates inference compared to traditional optimization techniques, and reduces potential unwanted biases in the reconstruction network.