Flow Filed Singularity Analysis For The Application On Image Dust Particle Detection

With the increasing of surveillance video clusters, video information has beenwidely used in daily production and human life. The upper-layer applications, such asdecision-making process, that base on the video information, the applied quality directlydepends on video quality. However, due to video-capture devices are exposed over along-dated, dust distortion can cause video quality degradation. Therefore, the videodust distortion detection becomes imperative. In this article, an effectivelynon-reference dust distortion detection scheme is presented based on flow fieldsingularity analysis.Firstly, the formed reasons of dust distortion are analyzed by means of the opticalmodel. According to the spatial and temporal invariance of dust, the relation betweendust and optical flow singularity is preliminary explored. Then, comparative analysis ofseveral common optical flow algorithms is introduced, and a suitable optical flowalgorithm for further research is determined in terms of accuracy and time-consuming.Combining with Matlab programming, a scheme that reduces time-cost is presented.Secondly, two generalized flow field singularity models: point singularity modeland line singularity model, are presented by fusing the flow field and singularity theory.Then, according to the polarization of the magnitude and direction of singularity, simpleanalysis of flow field singularity is described under some classical image processingalgorithms, which provides a foundation for the research of the dust distortion detection.Finally, comparative analysis of the magnitude and direction of optical flow ispresented, the direction is determined as the guidance information for dust detection interms of its higher integrity, and the fallible directions are corrected based on localneighborhood direction. On this basis, this article combines the aforementioned flowfield singularity analysis with the corrected direction to achieve the rough dust detection.And then, dust’s accuracy detection is realized through the introduction of temporalvoting mechanism and dust’s color space features.The proposed schemes have a good performance through experimental analysis.