DeepLIR: Attention-Based Approach for Mask-Based Lensless Image Reconstruction

Lensless imaging has emerged as a promising solution to overcome the need for expensive and bulky lenses used in traditional cameras. This technique leverages a mask to op-tically encode the scene, thus generating a sensor pattern. The image is subsequently reconstructed using a computational algorithm. Traditional model-based reconstruction methods often suffer from prolonged convergence time and subpar perceptual image quality. To mitigate these issues, data-driven deep neural networks can potentially offer enhanced reconstruction quality alongside reduced inference time. However, deep learning methods fall short in providing improved results and tend to produce artifacts, primarily because they do not incorporate any prior knowledge about the imaging model. In this work, we propose a DeepLIR, a hybrid approach that combines the physical system model with a deep learning model. This is achieved by unrolling a conventional model-based optimization al-gorithm and incorporating an attention-based deep learning model to denoise the image, thereby enhancing the reconstruction quality. Our empirical analysis confirms that DeepLIR surpasses existing lensless image reconstruction techniques in terms of image quality and computational ef-ficiency. Specifically, DeepLIR achieves a remarkable 1.35 × improvement in perceptual quality over the nearest com-petitor, reflecting its robustness and superiority. Further-more, it demonstrates superior generalization capabilities when applied to real-world imaging. Code available at: https://github.com/arpanpoudel/lenslessimaging.