A novel non-motion-compensated method is proposed for dirt detection in archived film sequences. A confidence measurement extracted from raw differences between current frame and each of the previous and next frames is used to exploit the temporally impulsive nature of dirt impairments. Further evolution of the confidence signal enables the minimization of false alarms and the fine-tuning of detector sensitivity. After morphological filtering and consistency checks candidate regions of dirt emerge, enabling the computation of binary detection masks. Our experiments show that our method compares favorably with extended spike detection index (SDIp), rank order detector (ROD), and lower-upper-middle (LUM) approaches and provides efficient and robust detection performance for a wide range of archived film material.