Reconstructing Synthetic Lensless Images in the Low-data Regime

CSIR-Central Electronics Engineering Research Institute (CSIR-CEERI), India
Academy of Scientific and Innovative Research (AcSIR), India
BMVC 2023

Our paper shows that it is possible to perform domain-restricted low-shot reconstruction of lensless computational images. We demonstrate an improvement in the reconstruction performance of untrained neural networks by providing very few domain-restricted images for pre-training. We were able to achieve faster convergence with higher reconstruction fidelity with as low as 10 images for pre-training. We compare the model performance against its untrained counterpart at various stages of convergence. We also show comparisons against a fully trained state-of-the-art.


Lensless image reconstruction is a class of inverse problems in computational imaging that is gaining immense popularity in the imaging research community because of its potential to revolutionize traditional imaging systems radically. It is possible to build cameras with extremely flat form factors that enable imaging in challenging scenarios. In this paper, we reconstruct simulated lensless images using a very low number of examples where we pre-train a decoder architecture with the available examples to model the data semantics, and then solve the inverse problem of lensless image reconstruction guided by a physics-informed forward loss function. We contrastively analyze and evaluate the reconstruction performance of our model against its untrained counterpart and show a substantial improvement in the reconstruction quality and convergence time even with a few example images.

Video Presentation


            author = {Banerjee, Abeer; Kumar, Himanshu; Saurav, Sumeet; Singh, Sanjay}, 
            title = {Reconstructing Synthetic Lensless Images in the Low-data Regime}, 
            publisher = {British Machine Vision Conference}, 
            year = {2023},