A hyperspectral image classification method and device combining random occlusion and BYOL structure
By combining the methods of random occlusion and BYOL structure, positive sample pairs are generated and self-supervised feature extraction is solved, and the problem of relying on label data in hyperspectral image classification is achieved, achieving high-precision and robust classification effects.
CN115115878BActive Publication Date: 2025-03-28ZHEJIANG UNIV +1
2 Cites 0 Cited by
Patent Information
- Application Number
- CN202210742447.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-27
- Publication Date
- 2025-03-28
- Estimated Expiration
- 2042-06-27
AI Technical Summary
Technical Problem
Existing deep learning methods rely on label data in hyperspectral image classification, and the classification effect is poor, especially when data is scarce and labeling is difficult.
Method used
Combining the hyperspectral image classification method of random occlusion and BYOL structure, positive sample pairs are generated through random occlusion, and self-supervised feature extraction is used using the BYOL algorithm, and the extracted features are finally inputted to the SVM classifier for training.
Benefits of technology
High-precision hyperspectral image classification in a small number of samples is realized, which reduces the dependence on label data and improves the robustness of classification effect.
✦ Generated by Eureka AI based on patent content.
Abstract
The present invention provides a hyperspectral image classification method and device combining random occlusion and BYOL structure. First, hyperspectral data is collected; the hyperspectral data is preprocessed, including PCA dimension reduction and patch cutting; each patch is randomly occluded twice, and other data enhancement methods are used as an auxiliary to obtain positive sample pairs; hyperparameters are set, all positive sample pairs are input into BYOL network training, and the features after training are saved; the extracted features are used to train the SVM model and evaluate the classification effect. The device includes two parts: an acquisition unit and a processing unit. The present invention solves the problem that supervised learning hyperspectral classification depends on label data and the classification accuracy of unsupervised learning is poor, and proposes a hyperspectral image classification method combining random occlusion and BYOL structure, and realizes a hyperspectral image classification method based on spectral preprocessing, random occlusion, BYOL spectral feature extraction and SVM classifier, and the accuracy can exceed most existing methods.
Need to check novelty before this filing date? Find Prior Art