UDHF2-Net: An Uncertainty-diffusion-model-based High-Frequency TransFormer Network for High-accuracy Interpretation of Remotely Sensed Imagery

• DOI: 10.48550/arxiv.2406.16129
• Publication Date: 2024-06-23
• Authors: Pengfei Zhang; Chang Li; Yongjun Zhang; Rongjun Qin
• Keywords: FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition

Remotely sensed image high-accuracy interpretation (RSIHI), including tasks such as semantic segmentation and change detection, faces the three major problems: (1) complementarity problem of spatially stationary-and-non-stationary frequency; (2) edge uncertainty caused by down-sampling in encoder step intrinsic noises; (3) false detection imagery registration error detection. To solve aforementioned problems, an uncertainty-diffusion-model-based high-Frequency TransFormer network (UDHF2-Net) is proposed for RSIHI, superiority which following: a spatially-stationary-and-non-stationary high-frequency connection paradigm (SHCP) to enhance interaction stationary non-stationary frequency features yield high-fidelity extraction result. Inspired HRFormer, SHCP remains stream through whole encoder-decoder process with parallel high-to-low streams reduces loss downsampling operation; mask-and-geo-knowledge-based diffusion module (MUDM) improve robustness noise resistance. MUDM could further optimize uncertain region result gradually removing multiple geo-knowledge-based semi-pseudo-Siamese UDHF2-Net task reduce pseudo error. It adopts architecture extract above complemental adaptively reducing differencing, recover besides noises. Comprehensive experiments were performed demonstrate UDHF2-Net. Especially ablation indicate effectiveness