Detecting marine oil spill regions in synthetic aperture radar (SAR) images has always been posed as a segmentation problem in terms of minimizing a certain energy function(al). As most energy minimization problems do not have analytical solutions, minimizing an energy function(al) is usually achieved in an iterative numerical manner. In this scenario, one key factor that affects the segmentation accuracy is the initialization for starting or constraining the numerical iterations. To guarantee accurate segmentation, a proper initialization that characterizes the marine oil spill layouts in a SAR image is required. However, marine oil spill regions are always complicatedly shaped, and it is inefficient to manually devise precise initializations for capturing various marine oil spill shapes. In order to address this problem and render efficient and robust segmentation, we develop a one dot fuzzy initialization strategy. In contrast to the normal practice of manually labeling a large amount of pixels (possibly lines or cycles of pixels subject to strict spatial conditions) as initialization, our strategy just requires one arbitrary pixel within a marine oil spill region as the initial dot. The intuition of our strategy is that the fuzzy connectedness between an arbitrary initial dot and the rest pixels enables the derivation of a physically homogeneous region which is consistent for initializing the energy minimization. In the light of this observation, we develop schemes for exploiting the one dot derived region to initialize both level sets for minimizing continuous energy functionals and graph cuts for minimizing discrete energy functions. Experimental results validate the robustness of our one dot fuzzy initialization strategy.
- marine oil spill segmentation
- energy minimization
- one dot fuzzy initialization
- fuzzy connectedness
Ren, P., Xu, M., Yu, Y., Chen, F., Jiang, X., & Yang, E. (2018). Energy minimization with one dot fuzzy initialization for marine oil spill segmentation. IEEE Journal of Oceanic Engineering. https://doi.org/10.1109/JOE.2018.2842538