Lens adhering contaminant detection using spatio-temporal blur

In this paper, we present a hierarchical spatiotemporal blur-based approach to automatically detect contaminants on the camera lens. Contaminants adhering to camera lens corresponds to blur regions in digital image, as camera is focused on scene. We use kurtosis for a first level analysis to detect blur regions and filter them out. Next level of analysis computes lowpass energy and singular values to further validate blur regions. These analyses detect blur regions in an image efficiently and temporal consistency of blur is additionally incorporated to remove false detections. Once the presence of a contaminant is detected, we use an appearance-based classifier to categorize the type of contaminant on the lens. Our results are promising in terms of performance and latency when compared with state-of-the-art methods under a variety of real-world conditions.