Weakly Supervised SAR Ship Segmentation Based on Variational Gaussian G (A) (0) Mixture Model A Learning

In this study, we propose a hybrid weakly supervised segmentation learning approach which employs a ship detection network and a novel segmentation process. First, two robust training strategies, creating soft labels and adding an extra regularization about the predicted probability of ship existence are proposed to train the ship detection network, which can alleviate the phenomenon of DNNS over fitting noisy and missing annotations. Then, OTSU is used to get Gaussian-$\\mathcal{G}_A^0$ mixture model-driven results on the parameter maps which are estimated by the VAE network and data-driven results on the original ROI data. By merging the two kinds of results, we can take the advantage of pixel-level information which can consider more structural details but is easily influenced by the speckle noise and the advantage of model-level information which can smooth the effect of the speckle noise. Our results demonstrate the accuracy of our algorithms regarding experiments on real Gaofen-3 SAR data which includes different complex sea conditions.