A Hybrid Optical-Neural Network Approach for Micro-Vibration Image Shift Estimation in Space Cameras

High-resolution space cameras require continuous knowledge of micro-vibration-induced image shift for image quality assessment and potential future compensation. This paper presents a simulation-based feasibility study of an optical measurement-assisted neural regression method that estimates imaging channel image shift without re-solving a full finite element model during inference. A high-frame-rate measurement channel shares the primary optics with the imaging channel; a lightweight, INT8-quantized and structurally pruned conventional neural regressor estimates the observable correction formed by the two measurement branch mirrors from rendered measurement frames, a calibrated sensitivity relation converts synchronized mirror-6 angular metrology into the same units, and both are added to the tracked measurement channel displacement. On the simulated dual-channel dataset, the proposed estimator reduces the correction RMSE from 0.396 px (centroid-only tracking) and 0.236 px (a specified structural-mismatch surrogate) to 0.130 px (0.133 px after INT8 quantization), which coincides with the structural scale of 0.132 px set by the common-path compensation residual; the remaining end-to-end image shift error of 0.398 px is dominated by the 0.378 px displacement tracking term rather than by the regressor. Its reported 0.0150 ms figure is a compute-only throughput estimate based on an assumed 120 GOPS INT8 accelerator; it excludes sensor exposure and readout, data transfer, synchronization, centroid tracking, and other pipeline overheads. A compute-only estimate does not establish the achievable end-to-end frame rate. The analytical error budget decomposes the measured shift error into displacement tracking, auxiliary metrology, and correction-scale terms and reproduces the measured value to within 1%. The modeled measurement branch does not observe independent motion of the imaging-only mirror; auxiliary state information or a validated mechanical constraint is required for a physically realizable estimator. The reported synthetic accuracy does not establish this observability; independent ray-tracing validation, optical bench experiments, hardware-in-the-loop tests, edge-hardware profiling, and on-orbit validation have not been performed.