Accurate contour preservation for semantic segmentation by mitigating the impact of pseudo-boundaries

Accurately extracting the contours of ground objects has been an important research topic in the field of semantic segmentation of remote sensing imagery. However, existing efforts have primarily focused on refining the boundaries of predictive masks, with little consideration given to pseudo boundaries caused by abrupt changes in surface textures. Therefore, this paper addresses this challenge with the contour preservation network (CPNet), a novel semantic segmentation network that effectively mitigates pseudo-boundary effects and produces more precise contours. The key of CPNet is the boundary-guided feature alignment module (BGAM). This module employs supervised boundary guidance to adaptively transfer the model's attention from salient areas to correct semantic boundaries. This adaptive attention transfer mechanism enables the model to suppress the impact of internal pseudo boundaries and refine contours. To further refine boundaries, a boundary point feature rectification module (ReBPM) is designed to rectify the classification of boundary points with neighbor features. Extensive experimental validations have demonstrated the effectiveness and flexibility of the proposed CPNet on ISPRS Potsdam and Vaihingen datasets. The results showed that our model outperforms other state-of-the-art methods in terms of boundary IoU, mean IoU, and mean F1-score, and it exhibits significantly superior contour preservation ability compared to other models, notably in the presence of pseudo-boundaries. The code is available at: https://github.com/angiecao/CPNet.