"Star Sensor Image"

"Data quality constitutes the upper bound of model performance in scientific vision tasks. Prevailing datasets frequently resort to simplified additive Gaussian noise models, which fail to encapsulate the non-stationary and spatially correlated noise characteristics intrinsic to on-orbit CMOS sensors (e.g., stray light noise, random telegraph noise, row\/column fixed pattern noise). To bridge this domain gap, we introduce HiFAHS, a dual-mode simulation framework designed to synthesize training data with extreme physical realism. The pipeline operates through two complementary streams: a physics-based optoelectronic modeling engine (Type-A) and a hybrid augmentation system utilizing authentic sensor substrates (Type-B). Unlike traditional denoising benchmarks that treat noise as a superficial additive layer, the Type-A pipeline simulates the complete photon-to-electron transfer chain. This process is rigorously structured into three sequential phases: Scene Initialization, Optical Rendering, and Optoelectronic Degradation."

在科学视觉任务中，数据质量是模型性能的理论上限。当前主流数据集多采用简化的加性高斯噪声模型，无法刻画在轨CMOS传感器固有的非平稳、空间相关噪声特性（例如杂散光噪声、随机电报噪声、行/列固定模式噪声）。为弥合这一领域差距，本文提出HiFAHS——一种旨在生成具备极高物理真实度训练数据的双模态仿真框架。该流程包含两个互补分支：一是基于物理的光电建模仿真引擎（A型），二是利用真实传感器基底的混合增强系统（B型）。与传统将噪声视为表层加性噪声层的去噪基准数据集不同，A型流程可模拟完整的光-电子转换链路。该过程被严格划分为三个依次执行的阶段：场景初始化、光学渲染与光电退化。