Learned split-spectrum metalens for obstruction-free broadband imaging in the visible

Obstructions such as raindrops, fences, or dust degrade captured images, especially when mechanical cleaning is infeasible. Conventional solutions to obstructions rely on a bulky compound optics array or computational inpainting, which compromise compactness or fidelity. Metalenses composed of subwavelength meta-atoms promise compact imaging, but simultaneous achievement of broadband and obstruction-free imaging remains a challenge, since a metalens that images distant scenes across a broadband spectrum cannot properly defocus near-depth occlusions. Here, we introduce a learned split-spectrum metalens that enables broadband obstruction-free imaging. Our approach divides the spectrum of each RGB channel into pass and stop bands with multi-band spectral filtering and learns the metalens to focus light from far objects through pass bands, while filtering focused near-depth light through stop bands. This optical signal is further enhanced using a neural network. Our learned split-spectrum metalens achieves broadband and obstruction-free imaging with relative PSNR gains of 32.29% and improves object detection and semantic segmentation accuracies with absolute gains of +13.54% mAP, +48.45% IoU, and +20.35% mIoU over a conventional hyperbolic design. This promises robust obstruction-free sensing and vision for space-constrained systems, such as mobile robots, drones, and endoscopes.

💡 Research Summary

The paper tackles a long‑standing limitation of diffractive metalenses: the depth‑wavelength symmetry that ties a lens’s chromatic focal shift to an equivalent shift in object depth. Because of this symmetry, a metalens that is broadband‑focused for distant scenes inevitably also focuses near‑depth occluders, making obstruction‑free imaging impossible in a single thin element. The authors first derive an analytical model (z = λ_d f / (λ − λ_d)) that quantifies how a wavelength deviation Δλ maps to a depth deviation Δz, providing a solid theoretical foundation for subsequent design.

To break the symmetry, they introduce a “split‑spectrum” approach. Each RGB channel of the sensor is divided into a high‑transmission pass band (Λ_pass) and a zero‑transmission stop band (Λ_stop) using a multi‑band spectral filter. The metalens is then learned to focus light from far objects only within the pass bands. When the same lens receives light from a near‑depth occluder, the depth‑wavelength symmetry shifts the focal wavelength into the stop band, where the filter blocks it entirely. Consequently, the occluder’s contribution is physically removed before reaching the sensor, while the far‑field image remains sharp across the visible spectrum.

The learning framework treats the orientation map θ(x, y) of geometric‑phase meta‑atoms as the design variable. Two differentiable modules are built: (i) a PSF simulator f_psf(θ, z, c) that outputs the point‑spread function for any depth z and color channel c, and (ii) an image simulator f_img(θ, I_clean, I_obs) that synthesizes the captured RGB image from a clean far‑field scene and a near‑depth obstruction map. The loss combines an image fidelity term L_img (encouraging the captured image to match the clean target) and a PSF term L_psf (promoting sharp, efficient far‑depth PSFs). Importance sampling weighted by the sensor’s spectral sensitivity reduces the computational burden of spectral integration. Training uses high‑resolution DIV2K images and randomly generated occluders, optimizing θ via first‑order gradient descent.

Fabricated meta‑atoms are made of silicon nitride (SiN_x) with geometry (period = 395 nm, height = 700 nm, width = 305 nm, length = 125 nm) optimized by rigorous coupled‑wave analysis to achieve conversion efficiencies of 78.4 % (457 nm), 72.9 % (530 nm), and 66.9 % (628 nm). Across the three pass‑band wavelengths the average efficiency is 67.3 %. Three lenses of identical physical dimensions (f = 4 mm, aperture = 2.516 mm) are fabricated: (1) the learned split‑spectrum lens (proposed), (2) a learned broadband lens without spectral splitting, and (3) a conventional hyperbolic lens designed for 532 nm.

Experimental PSF measurements from 430 nm to 645 nm confirm that the split‑spectrum lens maintains a sharp focal spot for far‑depth objects within the pass bands while exhibiting strong defocus for near‑depth objects, whose shifted wavelengths fall into the stop bands. The broadband lens improves chromatic performance but fails to blur near‑depth occluders; the hyperbolic lens shows severe chromatic aberration and no obstruction suppression.

Imaging tests involve capturing scenes with and without artificial occlusions, followed by reconstruction with a lightweight neural network trained separately for each lens. The split‑spectrum system achieves a PSNR of 20.94 dB under obstruction—a 32.29 % improvement over the hyperbolic baseline and 11.45 % over the broadband baseline. In unobstructed conditions it reaches 23.41 dB (23.90 % over hyperbolic).

Beyond raw image quality, the authors evaluate downstream computer‑vision tasks using off‑the‑shelf models (no fine‑tuning). On drone aerial imagery (VisDrone) the split‑spectrum lens yields an mAP of 0.1704 versus 0.0350 (hyperbolic) and 0.0292 (broadband). For medical endoscopic segmentation (KvasiR‑SEG) it achieves IoU 0.8317 versus 0.3472/0.5950, and for autonomous‑driving segmentation (Cityscapes) it reaches mIoU 0.6701 versus 0.4666/0.4601. These gains (up to +48 % IoU) demonstrate that physically removing occluder light translates into substantially better perception performance without algorithmic compensation.

The discussion highlights that the depth‑wavelength analytical model and split‑spectrum concept are broadly applicable to other diffractive‑optics domains such as color holography and depth‑encoded displays. Future work may explore multi‑focal designs, dynamic spectral filtering, and real‑time neural‑network reconstruction to further enhance compact, obstruction‑free imaging platforms. In summary, the paper delivers a single‑shot, thin‑lens solution that simultaneously achieves broadband imaging and near‑depth obstruction suppression, opening new possibilities for compact robotic vision, aerial surveillance, and minimally invasive medical imaging.