A Robust CFAR Algorithm Based on Superpixel Merging Operation for SAR Ship Detection

An issue of existing superpixel-level constant false-alarm rate (CFAR) ship detector in synthetic aperture radar (SAR) images is that they often rely on the manual adjustment of the superpixel sizes to generate good SAR image segmentation results but inappropriate superpixel sizes often lead to wrong segments of large targets and low detection performance. To address this issue, a robust CFAR method based on superpixel merging is proposed. In the preprocessing stage of our method, initial superpixels of marine SAR images are generated using the classical simple linear iterative clustering (SLIC) strategy. Then, a global threshold based on the distributional features of outlier-contaminated clutter region is used to classify initial superpixel cells into background clutter ones and candidate target ones. Next, the sidelobes and ghosts are filtered out from the set of candidate target superpixels by the knee point method, and the rest of candidate target superpixels from large targets are selected for merging processing. In this way, the merged candidate target superpixels will participate in the subsequent local CFAR detection as a whole, which preserves the integrity of the target information. Experiments results on real SAR images show that the proposed CFAR method outperforms other state-of-the-art methods in terms of robustness regarding the superpixel segmentation.