Metalenses for Generalizable Computer Vision Achieve Comparable Accuracy across Sensor Pixel Sizes

Optical neural networks represent a promising pathway to dramatically reduce the energy demands of machine learning, and researchers are increasingly focused on developing compact, efficient optical components to power these systems. Yubo Zhang, Johannes Fröch, and Jinlin Xiang, along with colleagues, investigate the potential of metalenses, ultra-thin optical components, to overcome a key limitation of current designs: a lack of generalizability to different visual tasks.

The team demonstrates that a carefully designed metalens, optimised for full-colour imaging, achieves image classification accuracy comparable to conventional sensors, and consistently outperforms alternative designs across a range of sensor sizes. This work not only advances the performance of optical neural networks, but also reveals that preserving spatial frequency information within the image is a critical factor in achieving high accuracy, offering valuable insight for the future design of these energy-efficient systems. Metalens and Neural Network Co-design for Imaging Researchers are pioneering a new approach to image classification by simultaneously optimizing both the optical and digital components of a vision system. This work focuses on Optimized Neural Networks (ONNs), where a metalens, a flat, nanoscale optical component, is co-designed with a neural network to maximize performance. Scientists demonstrate that by jointly optimizing the metalens’s properties and the neural network’s parameters, they can achieve superior image classification accuracy. A key finding is the importance of balancing the metalens’s ability to transmit different spatial frequencies, measured by the Modulation Transfer Function (MTF), across the visible spectrum.

The team developed a method that allows gradients, signals indicating how to adjust parameters, to flow from the digital neural network back to the physical properties of the metalens, enabling end-to-end optimization. They use sophisticated simulations to model how light interacts with the metalens and differentiable surrogates to facilitate this backpropagation process. Results consistently show that this co-design approach improves accuracy compared to using conventional lenses. Scientists discovered that balancing the MTF across red, green, and blue wavelengths is crucial for optimal performance. They found that the integral of the MTF, a measure of overall information preservation, is a more reliable indicator of performance than the specific phase profile of the metalens.

This research highlights the potential of co-designing optical and digital components for improved vision systems and provides valuable insights into designing optical systems that preserve information effectively. Metalens Design for Full-Color Image Classification Researchers are developing metalenses, flat, nanoscale optical components, to create more efficient optical neural networks. This work addresses limitations of previous designs, which suffered from reduced accuracy. Scientists demonstrate that a metalens optimized for full-color imaging achieves image classification accuracy comparable to high-end systems, consistently outperforming a conventional hyperboloid lens across a range of sensor pixel sizes. This achievement represents a significant step towards creating vision systems that are both accurate and energy efficient. The fabrication process involves depositing silicon nitride onto a quartz substrate, followed by electron-beam lithography to define a precise array of nanoscale scatterers. Reactive ion etching transfers this pattern into the silicon nitride layer, creating the functional metalens. Optical and scanning electron microscopy confirm the structural quality and uniformity of the resulting metasurface. Characterization using laser sources reveals that preserving spatial frequency information is critical for optimal performance of optical neural networks, providing both an interpretable understanding of the optimization process and practical guidance for designing high-performance optical encoders. The straightforward geometry and fabrication process are inherently compatible with scalable nanoimprint lithography, making this technology suitable for mass manufacturing.

This research paves the way for creating compact, energy-efficient vision systems for a wide range of applications.

Metalenses Enable Efficient Optical Neural Networks Scientists are achieving breakthroughs in optical neural networks (ONNs) by demonstrating the utility of metalenses for generalized vision tasks. Their work focuses on designing optical encoders that efficiently process images before they reach a digital neural network, dramatically reducing energy consumption while maintaining high classification accuracy. Researchers discovered that a metalens optimized for full-color imaging achieves image classification accuracy comparable to high-end lenses, consistently outperforming a hyperboloid metalens across a range of sensor pixel sizes. This demonstrates the potential of metalenses to create more efficient and powerful vision systems.

The team designed an end-to-end single-aperture metasurface for ImageNet classification and found that the optimized surface balances the modulation transfer function (MTF) for each wavelength. This balance, they discovered, is crucial, as the preservation of spatial frequency-domain information is a key factor in ONN performance. Measurements confirm a strong correlation between the volume under the MTF curve and task performance for incoherent imaging systems. Simulations reveal that the optimized optical encoder exhibits larger MTF integrals than a hyperboloid design across a broad range of spatial frequencies. Experiments using a 1-cm-aperture metalens, coupled with a digital neural network, demonstrate consistently high performance.

The team trained digital backends on data collected from both the optimized lens and a hyperboloid meta-lens, finding that the end-to-end optimized lens consistently outperforms its counterpart. These results collectively demonstrate the pivotal role of spatial frequency information preservation in determining ONN performance, offering a unified perspective for improving end-to-end optical-electronic co-designs.

Metalens Optimisation Maximises Image Modulation Transfer Function This work investigates hybrid optical-electronic neural networks, specifically focusing on the design of a single-aperture metalens coupled with a finite-pixel-size sensor for image classification tasks. Researchers identified a consistent optimization trend during end-to-end training: the optical component consistently maximizes the integrated modulation transfer function (MTF) within the limitations imposed by the sensor’s cutoff frequency. This suggests that preserving spatial frequency information is crucial for optimal performance of these networks. Experimentally, the team implemented a 1-centimeter-aperture metalens optimized using MTF maximization within a differentiable simulation environment. This lens demonstrates improved color rendition and a balanced RGB MTF at the sensor plane compared to a conventional hyperboloid lens, leading to higher classification accuracy across various digital back ends. Importantly, the performance advantage of this optimized lens remains consistent even with changes in detector resolution, indicating robustness to variations in pixel size. These findings support the hypothesis that emphasizing the in-band MTF integral during broadband operation helps preserve task-relevant spatial frequencies. While this study focused on a static, rotationally symmetric metalens for clarity, the authors note the framework readily extends to more complex optical designs. Future research will explore task-adaptive fine-tuning and programmable meta-optics capable of sustaining high MTF across broader spectral and angular bandwidths, paving the way for scalable and physically interpretable integration of meta-optical front ends with modern machine perception systems. 👉 More information 🗞 Applicability of Metalenses for Generalizable Computer Vision 🧠 ArXiv: https://arxiv.org/abs/2512.08109 Tags: