Traditional document digitization using specialized scanners is expensive, while manual camera photography is time-consuming. This study proposes a novel two-stage filtering framework for video-based document digitization using a fixed overhead camera to automatically extract static images of pages. The framework combines (1) temporal anomaly detection using the cosine similarity of lightweight CNN features between consecutive frames to identify page-turning events (PTEs) and (2) density-based clustering with OPTICS to group similar frames and eliminate the remaining PTEs as noise. The key innovation is a lightweight implementation that runs on CPUs using pretrained MobileNetV3 features and requires no GPU or additional training. This enables practical deployment in resource-constrained settings. The workflow separates recording from processing, allowing batch processing and parameter adjustments without the need for re-recording. Experiments on four real-world datasets achieved perfect recall (1.0), which means that no pages were lost while maintaining a practical precision. This framework offers a cost-effective alternative for libraries and archives that operate under budgetary constraints.