Unsupervised Transformer With Generative Label Optimization for Marine Aquaculture Segmentation

Supervised deep learning neural networks provide an accurate method for synthetic aperture radar (SAR) image semantic segmentation, but require training that relies heavily on high-quality labeled data. Meanwhile, existing unsupervised methods to obtain high-precision segmentation results are challenging in complex scenarios. Moreover, focusing on the long-term serial monitoring task of marine aquaculture, a significant amount of unlabeled SAR data remains underutilized. To address these problems, this paper proposes an unsupervised transformer with generative label optimization (UTGLO) for marine aquaculture segmentation, which utilizes unsupervised training of generative pseudo-label optimization to improve segmentation performance. First, a dual network architecture based on pre-trained self-supervised transformer (SST) and semantic segmentation transformer (SegT) is proposed to improve the quality of pseudo-labels using the dual network iterative update mechanism. Second, to learn the global semantics of targets and reduce the interference between them and the background, a global optimal pseudo-label generator is created to obtain the target’s discriminative features, compute the similar semantics globally, and generate the initial pseudo-labels. In addition, to alleviate the network performance degradation problem caused by the coarseness of pseudo-labels, the fine-grained feature discrimination and complement module is designed to optimize the target boundary features and continuity features of pseudo-labels. The superiority of the proposed method over state-of-the-art algorithms is demonstrated via experimentation on GF-3, Radarsat-2, and Sentinel-1 SAR datasets and three different aquaculture regions. Code and data are available in the GitHub repository https: //github.com/fjc1575/Marine-Aquaculture/tree/main/UTGLO.