Pixel-Superpixel Contrastive Learning and Pseudo-Label Correction for Hyperspectral Image Clustering

• DOI: 10.48550/arxiv.2312.09630
• Publication Date: 2023-01-01
• Authors: Renxiang Guan; Zihao Li; Xianju Li; Chang Tang
• Keywords: FOS: Computer and information sciences; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition

Hyperspectral image (HSI) clustering is gaining considerable attention owing to recent methods that overcome the inefficiency and misleading results from absence of supervised information. Contrastive learning excel at existing pixel level super HSI tasks. The pixel-level contrastive method can effectively improve ability model capture fine features but requires a large time overhead. utilizes homogeneity reduces computing resources; however, it yields rough classification results. To exploit strengths both methods, we present pseudo-label correction (PSCPC) for clustering. PSCPC reasonably domain-specific fine-grained through pixels comparative small number within pixels. performance pixels, this paper proposes module aligns pseudo-labels super-pixels. In addition, are used supervise clustering, improving generalization model. Extensive experiments demonstrate effectiveness efficiency PSCPC.