Fully-Connected Continuous Conditional Random Field With Stochastic Cliques for Dark-spot Detection In SAR Imagery

Dark-spot detection from synthetic aperture radar (SAR) imagery is a fundamental step in marine oil-spill detection and monitoring. However, to achieve robust and accurate detection is difficult due to SAR sensor limitations and the complex marine environment. To address this problem, the large-scale spatial-contextual information in SAR imagery has to be utilized to increase the class separability between the dark-spot and the background. A stochastic fully-connected continuous conditional random field (SFCCRF) approach to model SAR imagery and perform soft-label inference has been designed and built, leading to an efficient detection algorithm. Instead of treating all pixels in the imagery as being connected, SFCCRF determines the connectivity of two pixels in a stochastic manner based on their proximity in both feature space and image space. Since SFCCRF provides an efficient and effective way for modeling the largescale spatial correlation effect, the resulting soft-labels can resist the influence of speckle noise and highlight the difference between dark-spot and the background. Dark-spot detection is achieved by binarizing the soft-labels estimated by SFCCRF. The proposed algorithm is tested on both simulated and real SAR imagery. The results show that SFCCRF can delineate the dark-spot with low commission and omission error rates.