计算机科学
图像分割
人工智能
分割
网(多面体)
图像(数学)
计算机视觉
自然语言处理
遥感
地质学
数学
几何学
作者
Junming Chen,Anzhu Yu,Bing Liu
出处
期刊:
[SPIE]
日期:2025-06-03
卷期号:: 8-8
摘要
Using the semantic segmentation network to intelligently segment the categories of remote sensing images (RSI) is of great significance for urban planning, land use, and environmental monitoring. However, due to the complexity of the foreground and background of remote sensing images and the multi-scale characteristics of segmented objects, accurate segmentation of multi-class objects is still a difficult problem. Multi-source data fusion strategy can alleviate these problems by adding complementary information. Inspired by the strategy of multi-source data fusion and the powerful generalization ability of large models, we propose a improved U-Net that fuses the depth map of large model reasoning to improve the segmentation accuracy of RSI. In detail, we first use the pre-trained large model to reason the depth map, then leverage the feature enhancement module to fuse the depth map at the feature level, and finally use U-Net to realize end-to-end training and prediction. We conduct numerous semantic segmentation experiments on the Vaihingen dataset and the Potsdam dataset published by ISPRS. The results show that the proposed method achieves 73.23% and 74.95% mean cross-union ratio (mIoU) on Vaihingen and Potsdam datasets, respectively, which is 2.54% and 1.22% higher than U-Net. This proves the effectiveness of our proposed method.
科研通智能强力驱动
Strongly Powered by AbleSci AI