Attention-based credible evidential segmentation network for remote sensing ship segmentation

To address the challenges in ship segmentation arising from image quality degradation and edge blurring caused by adverse weather conditions and wave interference, this paper introduces the novel attention-based Credible Evidential Network, CEviCU-Net. Built upon Dempster-Shafer theory and the classical UNet architecture, CEviCU-Net consists of two main components: the feature extraction module and the evidential segmentation module. Specifically, the feature extraction module utilizes the U-shaped encoder-decoder framework. And this framework incorporates the Coordinate Attention Convolutional (CAC) module at the skip connection layers to extract semantic feature vector for each pixel. Subsequently, the evidential segmentation module in the CEviCU-Net employs a distance metric to compute the basic belief assignments (BBAs) between prototypes in the feature space and semantic feature vectors of each pixel. These BBAs are aggregated in the utility layer to facilitate semantic segmentation, accommodating the classification of ambiguous pixels and outliers. Finally, the model is trained end-to-end to jointly optimize all parameters of the proposed CEviCU-Net. Experimental results demonstrate the CEviCU-Net achieves the mIoU scores of 86.89% on the HRSID dataset and 90.43% on the Small ShipInsSeg dataset, representing improvements of 1.97% and 3.34% in average, respectively, compared to baseline models. These findings highlight the effectiveness of our proposed CEviCU-Net in the remote sensing ship segmentation.