面向NNBI可视层析模拟重建的SG-DnCNN后处理方法

An SG-DnCNN Post-Processing Method for Simulated NNBI Visible Tomographic Reconstruction

摘要: 有限视角与投影噪声降低负离子源中性束注入(negative-ion-based neutral beam injection, NNBI)可视层析质量。提出结构引导去噪卷积神经网络(SG-DnCNN)：感兴趣区域掩膜门控模块(MaskGate)调制ROI特征，可学习高斯结构先验(GSP)提供束斑包络弱偏置，后置投影残差一致性修正(PRCC)降低重投影偏差。基于既有960组模拟图像对开展训练和评价，并与Lee、TV、SART和SART + TV进行比较。PRCC采用独立构建的四方向替代投影模型，其投影域结果仅用于评价替代观测模型下的相对一致性改善。在48个测试样本上，MaskGate与GSP联合配置在当前模块消融中相较单模块变体取得较低的综合误差；3次PRCC使MAE、MaxAE和投影域MAE分别下降8.56%、4.71%和1.66%，PSNR约提高0.70 dB，SSIM略有下降。该方法在当前模拟协议下兼顾图像域误差与替代模型下的重投影残差，但仍需采用同源投影和真实NNBI数据进一步验证。

Abstract: Limited views and projection noise degrade visible tomography of negative-ion-based neutral beam injection (NNBI). We propose a structure-guided denoising convolutional neural network (SG-DnCNN): ROI-guided MaskGate modulates valid-region features, a learnable Gaussian structural prior (GSP) provides a weak beam-envelope bias, and post-hoc projection residual consistency correction (PRCC) reduces reprojection discrepancy. The models were trained and evaluated on 960 existing simulated image pairs against Lee, TV, SART, and SART + TV. PRCC used an independently constructed four-view surrogate projection model, and its projection-domain result was used only to assess relative reprojection improvement under the surrogate observation model. On 48 test samples, the combined MaskGate-GSP configuration showed lower overall error than the single-module variants within the current module ablation. Three PRCC steps reduced MAE, MaxAE, and projection-domain MAE by 8.56%, 4.71%, and 1.66%, respectively, increased PSNR by approximately 0.70 dB, and slightly decreased SSIM. Under the current simulation protocol, the method balances image-domain error and surrogate-model reprojection consistency; validation with source-matched projections and real NNBI observations remains necessary.