Blur detection for surveillance video based on heavy-tailed distribution

With the prevalence of video surveillance systems, the demand for video quality assessment in terms of blur is raised quickly. In this paper, a fast and effective method based on distribution of gradient magnitudes is proposed. The moving foreground regions are first extracted based on adaptive background mixture models. Detections of two types of blur, the global blur and the partial blur, are classified according to the accumulated area of the foreground regions. The gradient magnitudes distribution of background image is used as a reference, and judgment is made by comparing the value of current frame with it. In detection of global blur, the distribution of the whole current frame is used. But for the partial blur, the distributions of moving regions are used instead. Experimental results show that the proposed method can achieve high-accuracy and high-speed blur detection, and the global blur and partial blur can also be distinguished effectively.