UNCERTAINTY GUIDED CLASSIFIER GATED SEGMENTATION FOR EFFICIENT CAMERA LENS DEFECT LOCALIZATION

Camera-lens contamination degrades visual input in autonomous and surveillance systems. This paper proposes an uncertainty-guided architecture in which a lightweight classifier activates a U-Net only when localization is required. A three-state gate distinguishes clean, uncertain, and defective frames. Architecture analysis gives 37,177 parameters and 5.99 million MACs for the classifier, compared with 1,931,233 parameters and 2.607 billion MACs for the reference U-Net. Across illustrative activation scenarios, calculated computational savings range from 24.4% to 89.7%; these values are analytical rather than measured runtime results.