CrCD: Multidirection-MLP-Based Cross-Contrastive Disambiguation for Hyperspectral Image Partial Label Learning
- Publisher:
- Institute of Electrical and Electronics Engineers (IEEE)
- Publication Type:
- Journal Article
- Citation:
- IEEE Transactions on Geoscience and Remote Sensing, 2024, 62, pp. 1-14
- Issue Date:
- 2024-01-01
In Progress
| Filename | Description | Size | |||
|---|---|---|---|---|---|
| 1779817.pdf | Published version | 3.96 MB |
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is being processed and is not currently available.
Due to the intricate nature of the hyperspectral image (HSI) and the constraints imposed by annotators limited prior knowledge, the collection of HSI data poses significant challenges. Although there has been growing attention toward the issue of transfer learning and few-shot learning, partial label learning (PLL) has received scant attention in HSI classification. PLL refers to a scenario where training samples are associated with a set of candidate labels, among which only one is the correct label. Consequently, PLL holds significant practical importance for HSI classification, as it can alleviate the costs associated with HSI labeling. In this article, a cross-contrastive disambiguation (CrCD) method is proposed based on multidirection multilayer perceptrons (MLPs) for HSI PLL, which includes two main components representation learning for feature extraction and label disambiguation strategy for PLL. First, we design a heterogeneous network framework composed of 2-D encoder and 3-D encoder, and introduce a multidirection MLP and a multiscale attention for long-range spatial-spectral information. Second, apart from enforcing consistency in feature representation, we pioneer the establishment of a consistency constraint on semantic prediction probabilities with contrastive learning. Furthermore, the cross-label disambiguation strategy is introduced to provide reliable guidance for network training. Extensive experiments demonstrate that CrCD outperforms several current state-of-the-art (SOTA) approaches in HSI PLL and achieves results comparable to fully supervised learning. Code https //github.com/Nemo96yu/CrCD.
Please use this identifier to cite or link to this item: