NJU-PKU Team Achieves High-Fidelity Lensless Imaging of Dynamic Scenes in One Shot

Lensless cameras may soon become far more capable of recording moving biological samples, thanks to a new computational imaging framework developed by researchers at Nanjing University and Peking University in China. The method, called McLDI-INR, combines a physically designed optical mask with an artificial intelligence technique known as implicit neural representation to reconstruct sharp, high-resolution images from a single measurement. The approach is designed to address one of the most persistent problems in lensless imaging: how to recover clear images when the subject is moving and the sensor records only a blurred, overlapping diffraction pattern.

Conventional cameras form images through lenses that focus light onto a sensor. Although this architecture is highly effective, lenses add thickness, weight, and manufacturing complexity, and they can introduce optical aberrations, chromatic dispersion, and other imperfections. Lensless imaging removes the lens and instead allows a sensor to capture the diffraction pattern produced when light from an object interacts with a coded optical element. The recorded pattern does not resemble the original scene, so computational algorithms must mathematically reconstruct the image. This architecture can enable extremely compact cameras, but it also makes the reconstruction process particularly demanding when the object changes position or shape during exposure.

Motion creates a fundamental difficulty because the sensor does not record a clean snapshot of a moving object. Instead, it captures a mixture of spatial information and temporal changes in a single diffraction measurement. The resulting pattern may contain motion blur, overlapping structures, and weakened high-frequency details such as edges, fine textures, and narrow contours. Traditional reconstruction techniques often attempt to solve this problem by collecting multiple measurements or imposing assumptions about how the object is expected to move. Such strategies are not always practical for live biological specimens, rapid mechanical events, or portable devices that must operate with minimal hardware and limited acquisition time.

The team addressed this challenge by constructing a mask-constraint lensless imaging system. Before reaching the sensor, the incoming optical field passes through a binary mask that modulates the light in a controlled way. This mask changes the diffraction pattern so that it contains information that can be more readily related to the object’s physical structure. In effect, the mask acts as an optical coding layer rather than a conventional focusing lens. The sensor still records a highly transformed measurement, but the transformation is deliberately designed to make the inverse reconstruction problem more informative and more stable.

McLDI-INR then models the dynamic object as a continuous function of both space and time. Spatial coordinates and a temporal coordinate are provided to a neural network, which learns to represent the complex amplitude distribution of the moving scene. Unlike a conventional image-processing model that produces a fixed grid of pixels, an implicit neural representation describes the scene continuously. This allows the reconstruction to estimate object information at different spatial locations and time points, potentially providing a more flexible description of motion and reducing dependence on the exact sampling pattern of the detector.

A central feature of the framework is its dual-domain collaborative loss function. During training and reconstruction, the predicted object is evaluated not only in the image domain but also in the frequency domain. The image-domain constraint encourages the recovered scene to remain consistent with the measured sensor data and the optical propagation model. The frequency-domain constraint examines how spatial frequencies are reproduced, placing particular emphasis on information associated with fine edges and textures. By forcing the reconstruction to satisfy both forms of evidence, the method seeks to prevent a visually smooth but physically inaccurate result, a common failure mode in computational imaging.

The frequency-domain component is especially important for moving objects. Motion blur tends to suppress high-frequency information, making boundaries appear soft and small structures difficult to distinguish. A reconstruction system that relies only on pixel-level agreement may reproduce the general shape of an object while losing the details that are most useful for identification and measurement. The physics-guided frequency constraint gives the network an additional mechanism for recovering these details while remaining linked to the optical model. According to the researchers, this collaboration between learned representation and physical propagation helps reduce artifacts and improves the temporal and spatial fidelity of the reconstructed images.

Simulation experiments showed that McLDI-INR maintained its performance across a range of dynamic scenes, including rapidly moving linear and nonlinear targets. Compared with conventional approaches, the method produced results closer to the known ground truth and more consistently preserved contours, edge structures, and high-frequency textures. The simulations were intended to test whether the framework could remain reliable when motion became complex rather than simply when an object translated at a constant speed. The reported results suggest that the dual-domain strategy can provide a useful balance between data-driven flexibility and the restrictions imposed by the underlying optical system.

The researchers also tested the technique using a physical mask-constrained lensless imaging system. In experiments involving a moving USAF resolution target, McLDI-INR improved edge sharpness and reduced the visible effects of motion blur. More demanding tests used freely swimming rotifers, microscopic organisms whose bodies can bend and whose tails contract and move independently. The reconstructed sequences captured non-rigid changes, including tail motion and displacement, rather than treating the specimen as a rigid object. This result is significant because living samples often combine translation, deformation, and irregular motion, creating conditions in which simple motion assumptions quickly break down.

The work, published in Intelligent Opto-Electronics, demonstrates how computational models can extend the capabilities of optical systems that would otherwise be limited by their hardware. By combining a coded mask, an implicit neural representation, and collaborative constraints in the spatial and frequency domains, the researchers have created a route toward single-shot dynamic imaging without conventional lenses or repeated acquisitions. Potential applications include miniaturized microscopy, portable biological detection, live-sample observation, and intelligent sensing devices. The method does not eliminate the need for careful optical calibration and computational resources, but it illustrates a broader trend in imaging: future cameras may depend as much on learned physical models as on glass, sensors, and illumination. If the approach can be made faster and more robust across different specimens and hardware configurations, lensless systems could move closer to practical use in compact medical and scientific instruments.

News Publication Date: 30-Jun-2026

Web References: https://doi.org/10.67704/ioe.2026.260002

References: “McLDI-INR: mask-constraint lensless dynamic imaging by dual-domain collaborative implicit neural representation,” Intelligent Opto-Electronics, DOI: 10.67704/ioe.2026.260002.

Keywords

Lensless imaging, computational imaging, dynamic imaging, implicit neural representation, deep learning, diffraction imaging, motion deblurring, optical sensing, microscopy, biomedical imaging, artificial intelligence