Enhancing UNet Architectures for Remote Sensing Image Segmentation with Sinkhorn Regularization in Self-attention Mechanism

This paper proposes a novel approach to enhance self-attention mechanisms in UNet architectures for remote sensing image segmentation tasks by incorporating Sinkhorn regularization to refine the ability of the network to capture local relationships and handle critical features within the data. Experimental evaluation on the ISPRS Vaihingen dataset demonstrates the effectiveness of the proposed method, which is competitive with state-of-the-art models in terms of average F1 score and overall accuracy. The enhanced attention allows for more efficient information flow and improved feature identification, leading to more accurate segmentation of small objects such as buildings, roads, trees, and cars.