High-quality Reconstruction from Flat Camera Measurements

Develop improved algorithms for reconstructing high-quality images from lensless flat camera (FlatCam) sensor measurements, which consist of multiplexed projections of the scene reflectance and are modeled by the separable linear system Y = Φ_L X Φ_R, in order to resolve the ill-posed inverse problem and produce high-quality reconstructions that surpass existing methods.

Background

Flat cameras, such as FlatCam, replace traditional lenses with a static amplitude mask placed close to the sensor, producing multiplexed measurements that render the reconstruction problem highly ill-posed. Prior approaches—including optimization-based methods and deep learning models—have not consistently delivered sufficiently high-quality reconstructions, highlighting a persistent gap in practical performance.

The paper introduces DifuzCam, which leverages a pre-trained diffusion model with ControlNet and a learned separable transformation to improve reconstruction quality. Despite these advances, the authors explicitly note that achieving high-quality reconstruction from flat camera measurements is not yet solved, motivating the need for better algorithms tailored to the separable imaging model and the complexities of real-world measurements.