Lensless light-field imaging through diffuser encoding

Conventional photography captures only spatial information, losing angular data. Light-field imaging, in contrast, captures both, offering capabilities like viewpoint shifting, refocusing, depth sensing, and depth-of-field extension. Existing light-field imaging methods face challenges. Multi-sensor approaches are bulky and expensive, while multi-exposure methods are time-consuming. Microlens array-based plenoptic cameras, while common, suffer from a trade-off between spatial and angular resolution. Computational light-field imaging techniques, employing modulation masks, avoid this trade-off by computationally reconstructing the light field. This paper proposes a novel computational light-field imaging method using a diffuser as an encoder, eliminating the need for lenses. Each point source creates a unique pseudorandom pattern on the sensor, encoding its spatial and angular information. The diffuser's properties allow the encoding of all light rays within the field of view, facilitating lensless light-field imaging.

The introduction section extensively reviews existing light-field imaging techniques. It categorizes these into three main approaches: multi-sensor single-exposure, single-sensor multi-exposure, and single-sensor single-exposure (using microlens arrays or similar). The limitations of each approach are discussed, highlighting the trade-offs inherent in microlens array-based systems and the time constraints of multi-exposure methods. The review also mentions alternative computational light-field imaging techniques that utilize modulation masks to overcome resolution limitations. This establishes the context and the need for a novel approach like the one proposed in the paper.

The core methodology involves using a diffuser as an encoding element. A diffuser-encoding light-field transmission model is developed. This model describes how the four-dimensional light field (spatial and angular information) is mapped onto a two-dimensional sensor image. This mapping is represented by a transmission matrix. The researchers devised a calibration strategy to determine this transmission matrix. The calibration process involves capturing images of a point source at various positions and using these images to determine the matrix elements. The transmission matrix then allows for computational reconstruction of the light field from the sensor image, offering adjustable spatio-angular resolutions. This process breaks the limitations imposed by the sensor's resolution. The experimental setup includes a diffuser (a thin transparent phase plate with a statistically varying surface), a light source (point source or a scene), and a sensor to capture the generated patterns. The algorithms for matrix calibration and light field reconstruction are detailed, enabling computational decoupling of the captured light rays to achieve the desired spatio-angular resolution.

The experiments demonstrate the effectiveness of the lensless light-field imaging system. Experiments were conducted using a point source, multiple point sources, and a USAF-1951 resolution target. For a single point source, different spatial and angular sampling resolutions were tested (e.g., 512x512x6x6, 512x512x12x12, 1024x1024x6x6). Digital refocusing was applied to the reconstructed light field to produce a focal stack, confirming the successful capture of depth information. The results showed that increasing angular sampling reduced the resolved size of the point source, while increasing spatial sampling did not significantly improve it but did increase computation time. Experiments with multiple point sources demonstrated the ability to distinguish and refocus on individual points at different depths. Finally, experiments with the USAF-1951 resolution target showed that increasing angular sampling did not enhance reconstruction for area objects. However, adjusting spatial sampling and using a rescaled calibration pattern yielded better reconstruction for objects closer to the system, showcasing the system's adaptability to different object depths. Quantitative results on runtimes and resolved sizes at various resolutions are presented in tables and figures, demonstrating the system's performance and scalability.

The results demonstrate the feasibility of using a diffuser for lensless light-field imaging, offering a significant departure from traditional lens-based systems. The computational reconstruction, enabled by the calibrated transmission matrix, allows for exceeding the inherent resolution limitations of the sensor, achieving high spatio-angular resolution. The ability to adjust spatio-angular resolution provides flexibility depending on the application's needs. The successful imaging of both point sources and extended objects showcases the system's versatility. However, the performance is sensitive to the position of objects in the scene. The methodology also requires computationally intensive steps, which is a limitation to consider when focusing on real-time applications. Future research could focus on optimizing algorithms and exploring different diffuser designs to improve computational efficiency and imaging performance.

This research presents a novel lensless light-field imaging method using a diffuser for encoding spatial and angular information. The developed diffuser-encoding light-field transmission model and calibration strategy enable computational light-field reconstruction with adjustable spatio-angular resolutions, surpassing sensor resolution limitations. Successful experimental results with point sources and extended objects demonstrate the efficacy of this approach. Future research could investigate alternative diffuser materials and optimization of the reconstruction algorithms to enhance efficiency and image quality.

The computational cost of reconstructing the light field is significant, potentially limiting real-time applications. The accuracy of the reconstruction is sensitive to the calibration process and the selection of the calibration pattern, particularly for extended objects at different distances. The current implementation uses a single diffuser, and future work could explore the benefits of using multiple diffusers or other scattering media to potentially improve the reconstruction quality and range of operation.