Abstract
Traditional hyperspectral imaging (HSI) systems often struggle with the trade-off between temporal resolution and hardware complexity. Snapshot HSI using Diffractive Optical Elements (DOEs) offers a solution, but the gap between high-precision computational designs and discrete physical fabrication levels often degrades performance. This paper introduces Quantization-aware Deep Optics (QDO), a framework that integrates hardware manufacturing constraints directly into the end-to-end optimization loop. By employing quantization-aware training and an adaptive height-level mechanism, the proposed system achieves high-fidelity hyperspectral reconstruction even with low-precision (e.g., 4-level) fabrication, effectively bridging the "reality gap" in deep optics.
Introduction
Hyperspectral imaging (HSI) is a critical tool for capturing spatial and spectral information simultaneously, enabling applications in medical diagnostics, precision agriculture, and environmental monitoring. Unlike standard RGB cameras, HSI captures a 3D "data cube" across numerous narrow wavelength bands. However, conventional HSI systems typically require scanning or multiple exposures, which limits their use in dynamic scenes.
Snapshot HSI aims to capture the entire data cube in a single 2D sensor exposure. A common approach involves using a Diffractive Optical Element (DOE) to encode spectral information into a spatially multiplexed image, which is then decoded by a reconstruction algorithm. Modern "Deep Optics" techniques optimize both the DOE's physical profile and the reconstruction network end-to-end. However, a significant challenge remains: while optimization occurs in high-precision floating-point space, physical fabrication via lithography is restricted to a few discrete height levels. This "quantization mismatch" leads to substantial performance drops when theoretical designs are manufactured.
Methods
The researchers propose a Quantization-aware Deep Optics (QDO) framework to address the fabrication gap. The system consists of a diffractive imaging encoder, a computational decoder, and a quantization-aware DOE optimization module.
The framework introduces two primary innovations:
- Quantization-aware Training (QAT): Instead of post-hoc quantization, the model incorporates the discrete nature of the DOE during training. Using an alpha-blending technique, the system gradually transitions from a full-precision height map to a quantized one. This allows the reconstruction network (a Res-UNet architecture) to learn and compensate for the specific optical artifacts introduced by the discrete fabrication levels.
- Adaptive Quantization (QDO+A): Standard quantization typically assumes evenly spaced levels. However, lithography allows for custom heights for each step. QDO+A treats these physical heights as trainable parameters. By optimizing the specific height of each fabrication level alongside the DOE pattern, the system minimizes information loss and maximizes reconstruction accuracy.
Results
The QDO framework was evaluated using the ICVL hyperspectral dataset and validated with a physical prototype. In synthetic experiments with only 4 quantization levels, the QDO+A method achieved a Peak Signal-to-Noise Ratio (PSNR) of 35.32 dB, a significant improvement over the 31.68 dB achieved by standard Deep Optics approaches. Remarkably, this 4-level adaptive model performed nearly as well as the theoretical "Full" precision model (36.82 dB).
Qualitative results demonstrate that QDO+A produces reconstructions with fewer artifacts and more accurate spectral signatures compared to non-quantization-aware methods. Physical validation using a fabricated 4-level DOE confirmed these findings, with the prototype successfully reconstructing spectral curves that closely matched ground-truth measurements from a commercial spectrometer.
Conclusion
This work demonstrates that hardware limitations, such as discrete fabrication levels, do not have to be a bottleneck for Deep Optics. By treating these constraints as learnable parameters within a quantization-aware framework, researchers can design snapshot HSI systems that are both high-performing and physically realizable. Future work may extend this framework to other optical tasks, such as high-dynamic-range imaging and depth sensing, further narrowing the gap between computational design and physical manufacturing.
Summary updated