Related Experiment Videos
Unsupervised multistage image classification using hierarchical clustering with a Bayesian similarity measure
Sanghoon Lee1, Melba M Crawford
1Department of Industrial Engineering, Kyungwon University, Kyunggi-do 461-701, Korea. shl@kyungwon.ac.kr
Summary
This study introduces a new hierarchical clustering method for unsupervised image classification. The technique effectively segments and classifies images using spatial context, improving accuracy for smooth patterns in remote sensing data.
Area of Science:
- Computer Vision
- Image Processing
- Machine Learning
Background:
- Unsupervised image classification is crucial for analyzing large datasets.
- Existing methods may struggle with spatial contextual information and computational efficiency.
Purpose of the Study:
- To develop a novel multistage hierarchical clustering method for unsupervised image classification.
- To enhance segmentation accuracy by incorporating spatial contextual information.
Main Methods:
- A two-phase approach: segmentation via hierarchical clustering with spatial constraints, followed by sequential merging for classification.
- Utilizes Markov random fields for spatial context in the first phase and context-free similarity in the second.
- Employs a multiwindow, pyramid-like structure for computational efficiency.
Main Results:
- The proposed method demonstrates high effectiveness in unsupervised image analysis.
- The region-merging approach leveraging spatial context yields more accurate classification, especially for images with smooth spatial patterns.
- Experiments on simulated and remotely sensed data validate the method's performance.
Conclusions:
- The multistage hierarchical clustering method offers a robust solution for unsupervised image classification.
- Incorporating spatial contextual information significantly improves classification accuracy.
- The method is computationally efficient and suitable for various image data types.