Image-based preprocessing for wide-field survey telescope active optics

Active optics systems require accurate and efficient identification of stellar sources for wavefront correction. For the wide-field survey telescope (WFST), we redesign the preprocessing pipeline by integrating a MobileNetV2-based classifier to distinguish isolated stars from blended or spurious detections in curvature sensor images. Combined with signal-to-noise ratio-guided selection strategy, the system achieves over 99% precision for isolated sources and background while reducing false detections from overlaps. With sub-second CPU latency, the method was deployed in the WFST control loop and validated through six nights of observations, confirming real-time applicability and scalability for future wide-field active optics systems.