Hyperspectral Image Classification With Capsule Active Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional hyperspectral image classification methods struggle to effectively extract joint spectral and spatial features, and existing active learning methods fail to consider the diversity of actively selected samples and the cost of expert labeling, limiting their performance with few labeled samples.
Innovation Solution
A collaborative active learning method based on capsule networks using two base classifiers, CapsViT and CapsGLOM, which learn local and non-local cross-channel spectral dependencies, and incorporate uncertainty and diversity considerations in sample selection, along with expert labeling costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional classification methods are used, then the methods are simple and classical, but they have limited abilities to extract and represent joint features of spectrum and space
Solution Approach 1:
The patent introduces capsule networks as an intermediary architecture between conventional classification methods and deep learning models. The capsule network structure with its dynamic routing mechanism serves as a mediator that effectively extracts joint spectral-spatial features while maintaining computational efficiency, resolving the contradiction between simplicity and feature extraction capability
Solution Approach 2:
The patent segments the feature extraction process into distinct spectral and spatial components that are processed separately and then integrated through the capsule network. This segmentation allows the model to capture joint features more effectively while maintaining the simplicity of the overall framework
2Device complexity
If active learning methods only consider uncertainty, then the query strategy is simple, but they fail to consider the diversity of actively selected samples and the cost of expert labeling
Solution Approach 1:
The patent merges multiple query strategies (uncertainty-based, diversity-based, and cost-based) into a unified collaborative active learning framework. This combination allows the system to select samples that are not only uncertain but also diverse and cost-effective, improving sample selection quality while maintaining reasonable complexity through integrated processing
Solution Approach 2:
The collaborative query strategy serves multiple functions simultaneously: it evaluates uncertainty, ensures sample diversity, and considers expert labeling costs. This multi-functionality resolves the contradiction by making a single comprehensive query strategy that addresses all three criteria without requiring separate complex systems
3Reliability
If deep learning models are used with few labeled samples, then the models have high potential, but it is difficult for various deep models to give full play to their potential
Solution Approach 1:
The patent applies preliminary action by using semi-supervised learning to pre-train the capsule network with both labeled and unlabeled data before fine-tuning with active learning. This preliminary training allows the model to leverage its high potential with limited labeled samples, improving training efficiency while maintaining model reliability
Solution Approach 2:
The patent implements feedback mechanisms where the model's predictions on unlabeled samples are continuously evaluated and used to guide active learning sample selection. This feedback loop allows the model to gradually improve its performance with few labeled samples, resolving the contradiction between model potential and training efficiency
Data Source
AI summary
A collaborative active learning classification method for hyperspectral images based on capsule networks is provided in the application, the method trains base classifiers CapsViT and CapsGLOM using an initial training set; calculates BvSB values of candidate samples using the CapsViT; uses the CapsGLOM to predict labels of the candidate samples; sorts the candidate samples according to the BvSB values; puts the candidate samples after sorting into corresponding collectors according to category labels estimated by the CapsGLOM; labels information samples; updates the initial training set and candidate sample set and retrains the CapsViT and CapsGLOM; and obtains the classification results based on CapsViT and CapsGLOM after iteration.