人工智能
计算机科学
高光谱成像
聚类分析
模式识别(心理学)
变压器
计算机视觉
遥感
特征提取
深度学习
图像分割
合成孔径雷达
图像处理
作者
Nan Tian Huang,Cheng Li,Qichao Liu,Jinpeng Li,Pengfei Liu,Zhizhong Zheng,Liang Xiao
标识
DOI:10.1109/tgrs.2026.3686883
摘要
Local window-based Vision Transformers (ViTs) have recently attracted significant attention in hyperspectral image (HSI) analysis for their balance between global modeling and efficiency. However, window-based self-attention in unsupervised HSI clustering still suffers from two key limitations: (i) implicitly restricting semantic correlations to local window, overlooking fragmented and noncontiguous yet semantically homogeneous regions; and (ii) using fixed grid windows, causing boundary mixing and weakening spectral–spatial discriminative representations. To overcome these issues, we propose a Semantic Transformer for Self-Supervised Deep Clustering of large HSIs (STSDC), which couples a Multi-Scale CNN Block (MSCB) with a Semantic Transformer under a unified self-supervised optimization module. Specifically, the MSCB branch employs parallel receptive fields with adaptive channel recalibration to preserve fine-grained structures and suppress spectral redundancy. The Semantic Transformer introduces superpixel-guided attention mechanisms, including Semantic Contiguous Attention (SCA), which selects tokens from intra- and adjacent superpixels to enhance spatial continuity, and Semantic Global Attention (SGA), which aggregates long-range superpixel representations to associate spatially disjoint but semantically homogeneous regions. The self-supervised module aligns the embedding representations from the MSCB and the Semantic Transformer via a clustering objective. Extensive experiments on benchmark HSIs demonstrate that STSDC achieves superior clustering performance compared with state-of-the-art self-supervised methods for large HSIs.
科研通智能强力驱动
Strongly Powered by AbleSci AI