目的针对图像修复中空间异质性退化（同时包含中心区域均匀模糊与边缘区域复杂折射）的难题，现有方法因缺乏退化属性先验与频域建模，难以兼顾两类退化属性的修复需求。为此，提出一种基于空间和频域联合处理与不确定性引导的多尺度稀疏变换器。方法首先设计多尺度物理解耦模块，利用不同膨胀率与核大小的并行卷积分支分别侧重提取全局模糊、过渡细节与高频折射特征；进而引入空间-频域联合处理模块，通过空域深度可分离卷积与频域幅度相位调制互补增强特征表示；在此基础上构建不确定性估计模块，生成空间不确定性图以量化各区域修复难度，并据此设计不确定性引导的稀疏注意力机制，使模型在高不确定性区域保持精细纹理，在低不确定性区域抑制冗余计算。同时，采用融合重建损失与不确定性图正则化的复合损失函数，约束不确定性图聚焦于退化区域。结果在SPAD雨纹数据集和AGAN‑Data雨滴数据集上，所提方法的峰值信噪比（peak signal-to-noise ratio，PSNR）和结构相似性（structural similarity index，SSIM）均优于现有基线模型；在Dense‑Haze真实雾霾数据集上，PSNR达到最优。在雨滴去除任务中，相比对比方法中的最优方法，PSNR指标提升了0.30 dB。消融实验验证了多尺度解耦、空间-频域联合处理及不确定性引导机制对性能的贡献。结论本文提出的空频联合与不确定性引导的多尺度稀疏变换器，能够有效应对空间异质性退化，在雨滴、雨纹和雾霾去除任务上均取得优越的修复效果。实验代码与数据已开源发布（https：//doi.org/10.57760/sciencedb.j00240.00269）。

Scalable classroom process evidence can support teacher reflection and expert review, but manual observation is labor-intensive and direct eye tracking is often too intrusive for routine classroom use. This paper presents a single-camera classroom video analytics system that converts ordinary classroom recordings into reviewable process evidence for teacher reflection. The system combines three-class head detection, teacher association and tracking, six-dimensional head-pose estimation for frontal teachers, pose smoothing with a multiplicative extended Kalman filter, gaze-cone projection, student and region hit testing, four-class rule-based behavior segmentation, and evidence export. In the perception experiments, the three-class head detector achieved an mAP@0.50 of 0.96, while frontal-teacher head-pose estimation yielded a mean angular error of 7.55°. In behavior verification, the algorithm achieved 78.8% sample-wise accuracy and a weighted F1-score of 77.1% against adjudicated human labels. The complete pipeline processed a single 1080p video stream at 48.20 FPS under the reported setup. Because the system infers visual attention from head pose, its output should be interpreted as a visual-attention proxy rather than as a direct measure of eye gaze. Its behavior labels describe observable classroom process patterns and are intended to support human review, not to evaluate teaching quality or individual participants. The exported heatmaps, regional timelines, behavior timelines, annotated videos, and structured reports provide traceable evidence that can be checked against the source recording.