Abstract: A segmentation method based on self-distillation edge refinement is proposed in this paper to tackle the challenges of edge feature loss and excessive parameter redundancy encountered during semantic segmentation of remote sensing images. Firstly, a backbone network is constructed using EfficientNetB4 as the foundation. Subsequently, a lightweight edge refinement module is integrated into the self-teacher network branch. This module is designed to capture local information from intermediate feature maps while retaining the intermediate edge details filtered by shallow neural networks, with the purpose to improve the accuracy of edge pixel segmentation in remote sensing images. Finally, an adaptive multi-view vector is created to serve as a novel knowledge guide for encoder network training. This is achieves by utilizing the binary category labels of each image as the prediction matrix. The adaptive multi-view vector provides a better description of intra-class and inter-class distributions, as well as fitting inter-layer and intra-layer relationships. On the public datasets DeepGlobe and Vaihingen, the proposed method achieves an average intersection ratio of 72.4% and 83.3%, respectively. Comparative experiments demonstrate that the method introduced in this study enhances edge features while maintaining a balance among segmentation accuracy, model parameters, and inference speed. It has good feature extraction ability while lightweighting the model.

Key words: self-distillation, edge refinement, remote sensing images, semantic segmentation, adaptive multi-view