搜索

x
中国物理学会期刊

基于非训练神经网络的傅里叶单像素高光谱边缘成像

Fourier Single-Pixel Hyperspectral Edge Imaging Based on an Untrained Neural Network

PDF
导出引用
  • 高精度边缘检测在工业制造中对表面缺陷识别与质量控制具有关键意义,然而传统成像方法在微尺度缺陷的解析能力上仍受到空间分辨能力不足的限制。本文提出一种基于物理驱动的生成对抗网络近红外傅里叶单像素高光谱边缘成像方法。该方法通过空间光调制器加载具有微分算子结构的测量基,将高斯–拉普拉斯算子的空间响应嵌入单像素调制与探测过程,使测量对目标空间强度的高阶变化更加敏感,从而在数据获取阶段直接增强边缘相关的高空间频率信息。在重构端,构建融合单像素积分测量模型的生成对抗网络,并引入物理一致性约束与光谱相关性约束,实现从512个近红外光谱通道的桶探测信号中反演高分辨率二维边缘图像。实验结果表明,在稀疏采样条件下,该方法在边缘分辨力与光谱保真度方面均显著优于传统重构策略,能够有效解析常规方法难以识别的微尺度表面划痕。该工作为高光谱单像素成像在材料表征、表面缺陷检测及精密测量等应用中提供了一种新的物理—数据协同成像方案

    High-precision edge detection plays a crucial role in surface defect identification and quality control in industrial manufacturing; however, conventional imaging methods still suffer from limited capability in resolving microscale defects due to constraints in spatial resolution and signal-to-noise ratio. In this work, we propose a physics-prior-driven generative adversarial network–based near-infrared Fourier single-pixel hyperspectral edge imaging method. By loading measurement bases with differential-operator structures onto a spatial light modulator, the spatial response of the Laplacian of Gaussian operator is embedded into the single-pixel modulation and detection process. This design enhances the sensitivity of singlepixel measurements to higher-order spatial intensity variations, enabling direct amplification of edge-related high spatial frequency information at the data acquisition stage. On the reconstruction side, a generative adversarial network incorporating the single-pixel integral measurement model is developed, with physical consistency and spectral correlation constraints imposed during optimization. This framework enables the recovery of high-resolution two-dimensional edge images from bucket signals across 512 near-infrared spectral channels. Experimental results demonstrate that, under sparse sampling conditions, the proposed method significantly improves edge resolution and spectral fidelity compared with conventional reconstruction strategies, allowing reliable identification of microscale surface scratches that are diffcult to resolve using traditional approaches. This work provides a new physics–data collaborative imaging paradigm for hyperspectral single-pixel imaging in applications such as material characterization, surface defect inspection, and precision metrology.

    目录

    返回文章
    返回
    Baidu
    map