A hybrid convolutional network for defocus detection in large-scale optical imaging spectrographs

Defocus and degradation in image quality commonly occurs in astronomical optical spectrograph systems due to environmental factors (e.g., temperature and humidity), routine maintenance, and operational disturbances. Manual inspection of the image quality and focus is a subjective and time-consuming process. Conventional machine-learning systems are based on hand-crafted features and simple models and inherently lack robust multi-scale discriminative power and generalization in varying imaging conditions. In this work, we designed a hybrid convolutional network consisting of ConvNeXt and LSNet, and included a new HybridBlock, for the purpose of adaptively refining and a fusing multi-scale features. The HybridBlock refines the multi-scale features by selectively enhancing discriminative cues while suppressing noise induced by environmental and operational disturbances, and offers selection of local detail versus an associated pattern in a broader surrounding context—thereby improving robustness without significantly increasing computation. The model was evaluated on a defocus dataset from LAMOST, in which the model achieved a highest 91.66% recognition accuracy rate (approx. 3.4% increase compared to the ConvNeXt baseline). Overall, the results demonstrate that the method could be automated and provide reliable focus-quality monitoring of large-scale astronomical and optical imaging systems, and potentially generalized for a range of telescopes and imaging platforms.