Enhanced event recognition in video using image quality assessment

摘要

Extensive growing repositories of multimedia present significant challenges for storage, indexing, retrieval, and analysis. The ability to recognize events based on automated analysis of the video content would facilitate tagging and retrieval of relevant data from large repositories. The unconstrained nature of multi-media data means that metadata often associated with a video is not known. In addition, many clips exhibit poor quality due to lighting, camera motion, compression artifacts, and other factors. The variable and frequently poor quality of video data challenges the state of the art in computer vision. In the absence of sensor metadata, we present an approach that estimates various attributes of video quality based on the content and incorporates this information into the event classification. Using a set of canonical content detectors, we establish a baseline level of event classification performance. Guided by the quality assessment into the classification process, we can identify data quality problems automatically. This analysis is a first step in tailored processing that would adapt the content extraction method to the estimated quality level. We present the formulation of the image quality measures and a quantitative assessment of the methods.