Lensless extended depth of field imaging using PSF correction and pre-denoising

Lensless cameras have emerged as a common method to extend depth of field (DoF) in computational imaging due to their simple and compact structure. Current lensless extended depth-of-field (EDoF) cameras are primarily designed to generate a depth-invariant point spread function (PSF). This strategy often sacrifices diffraction efficiency to ensure PSF consistency across varying depths. As a result, the cameras usually exhibit PSFs with long tails and narrow modulation transfer functions, causing inaccuracies in the quantized PSFs during measurement and inevitable noise amplification during deconvolution. To address these issues, we first design a PSF correction method that improves the accuracy of the measured PSF through two measurements. Next, we propose a pre-denoising strategy, consisting of a denoising network (DN) and a joint training framework. The DN preprocesses sensor noise, while the joint training framework enables the reconstruction algorithm to guide the DN's convergence, achieving high-quality imaging across a large DoF range in a single exposure. This work effectively overcomes the imaging quality degradation caused by the camera's low diffraction efficiency, with considerably low computational and time consumption.