A dual-channel network for cross-domain one-shot semantic segmentation via adversarial learning

The purpose of one-shot semantic segmentation is to classify all pixels in a query image with a labeled support sample. Its current approaches divide data into two disjoint parts, a training set and a testing set. When the training and testing data are from different domains, the migration of the trained model to the testing data for prediction tends to perform less efficiently. Therefore, through the introduction of a distribution alignment module, a new dual-channel network (DCNet) for one-shot segmentation is proposed in this paper. The purpose of this introduction is to alleviate the performance degradation caused by distribution differences between the datasets. Furthermore, a soft pseudo-mask based prototype refinement module with Gumbel-Softmax is designed to enhance the expressive ability of prototypes and improve segmentation performance. Experiments are made on four datasets with large distribution differences, proving that our approach outperforms the compared methods by a large margin.& COPY; 2023 Elsevier B.V. All rights reserved.