Marine Oil Spill Image Segmentation Based On Energy Minimization

Oil spill monitoring is an important task for dealing with oceanic pollution.Therefore,it is quite worthwhile to explore the oil spill segmentation methods for remote sensing images.In this thesis,we investigate three oil spill segmentation methods based on the energy minimization theoretic framework.Firstly,we propose an oil spill segmentation method based on dual smoothing regionscalable fitting level set method in order to increase the oil spill segmentation accuracy of the existing region-scalable fitting level set methods.Our method formulates an energy function consisting of an image smoothing term based on guided filter and a region fitting smoothing term based on active contours,and thus results in a dual smoothing scheme in dual levels,i.e.the pixel level and the region fitting level.Then we transform the energy function into a new level set function,which is minimized by using of steepest descent method.The optimal oil spill area contour is obtained by the energy minimization.In our scheme,guided filter smoothes noisy oil spill images and also preserves oil region edges.Therefore,our dual smoothing regionscalable fitting level set method,which integrates the guided filter,has better noise immunity than existing region-scalable fitting level set methods.Secondly,in order to reduce errors in elongated oil spill segmentation based on Graphcuts,we investigate an oil spill segmentation method based on the cooperative model.Our method constructs high order energy function for the cooperative model with strong boundary consistency.The region term and edge term of the energy function are formulated based on the gamma mixture model and the cooperative model,respectively.We then transform a high order term into pairwise terms and transform the energy function into s/t graph.Min-cut/Max-flow is employed to obtaining the min cut of s/t graph,which is the optimal solution of energy function minimization.The high order term of the cooperative model penalizes not only the length of the boundary but also the diversity of the boundary,which overcomes the weak boundary consistency of graph-cuts.Therefore,this scheme is more effective for segmenting elongated strip oil spills.Finally,in order to further improve the oil spill segmentation accuracy of the cooperative model,we propose a dual smoothing cooperative model framework for oil spill segmentation.Our new method integrates the rolling guidance filter to the region term of cooperative model energy function.The new energy function not only smoothes the noise in both label and image field but also reserves the boundary detail.Therefore,the dual smoothing cooperative model framework is more robust in segmenting oil spill areas,and has stronger ability to retain boundary details.