Sharp or Blur : A Fast No-Reference Quality Metric for Realistic Photos

摘要

There is an increasing demand on identifying the sharp and the blur photos from a burst ofseries or a mass of collection. Subjective assessment on image blurriness takes account of notonly pixel variation but also the region of interest and the scene type. It makes measuring imagesharpness in line with visual perception very challenging. In this paper, we devise a noreferenceimage sharpness metric, which combines a set of gradient-based features adept inestimating Gaussian blur, out-of-focus blur and motion blur respectively. We propose a datasetadaptivelogistic regression to build the metric upon multiple datasets, where over half of thesamples are realistic blurry photos. Cross validation confirms that our metric outperforms thestate-of-the-art methods on the datasets with a total of 1577 images. Moreover, our metric isvery fast, suitable for parallelization, and has the potential of running on mobile or embeddeddevices.