基于物理先验引导生成对抗网络的中子图像超分辨率重建

Neutron Image Super-Resolution Reconstruction Based on Physics-Informed Generative Adversarial Network

摘要: 中子照相技术在复杂装备的无损检测中不可或缺，但受限于探测器闪烁屏的光学散射与泊松–高斯混合噪声，原始图像常面临空间分辨率低、边缘模糊的困境。传统的图像恢复算法极易导致“过平滑”效应，而常规的深度生成模型在重建高频细节时，往往因脱离物理规律而产生严重的“AI幻觉”与结构失真。针对上述痛点，本文提出了一种基于物理先验引导的生成对抗网络(PI-GAN)超分辨率重建算法。该算法创新性地构建了基于中子点扩散函数(PSF)的可微分正向物理退化层，并引入物理一致性损失，将神经网络的纹理生成空间严格限制在物理法则允许的流形内。同时，结合残差通道注意力机制与相对论判别器，网络在极低信噪比下实现了端到端的高保真特征提取。以典型工业样品为测试对象的实验结果表明，本文算法不仅彻底抑制了强混合噪声，更从根本上杜绝了虚假波纹等伪影。与传统滤波方法相比，该算法在结构相似性(SSIM)和峰值信噪比(PSNR)上取得显著提升，调制传递函数(MTF)分析进一步证实其近乎无损地保留了极限空间分辨率。本文研究为突破中子成像硬件瓶颈、实现高精度无损检测提供了全新的理论框架与算法支撑。

Abstract: Neutron radiography is indispensable in the non-destructive testing (NDT) of complex equipment. However, constrained by optical scattering within the detector’s scintillator screen and high Poisson-Gaussian mixed noise, raw projection images often suffer from low spatial resolution and blurred edges. Traditional image restoration algorithms easily lead to an “over-smoothing” effect, while conventional deep generative models, when reconstructing high-frequency details, frequently produce severe “AI hallucinations” and structural distortions due to the lack of physical constraints. To address these bottlenecks, this paper proposes a Physics-Informed Generative Adversarial Network (PI-GAN) for super-resolution reconstruction. The proposed algorithm innovatively constructs a differentiable forward physical degradation layer based on the neutron Point Spread Function (PSF) and introduces a physical consistency loss, strictly restricting the neural network’s texture generation space to a manifold permitted by physical laws. Simultaneously, by combining a residual channel attention mechanism with a relativistic discriminator, the network achieves end-to-end high-fidelity feature extraction under extremely low signal-to-noise ratios. Experimental results on typical industrial samples demonstrate that the proposed algorithm not only thoroughly suppresses strong mixed noise but also fundamentally eliminates artifacts such as fabricated wavelets. Compared with traditional filtering methods, this algorithm achieves significant improvements in Structural Similarity (SSIM) and Peak Signal-to-Noise Ratio (PSNR). Furthermore, Modulation Transfer Function (MTF) analysis confirms that it preserves the ultimate spatial resolution almost losslessly. This research provides a novel theoretical framework and algorithmic support for breaking through the hardware bottlenecks of neutron imaging and realizing high-precision NDT.