Dynamic Affinity Graph Construction for Spectral Clustering Using Multiple Features.

Area of Science:

• Computer Vision
• Machine Learning
• Data Mining

Background:

• Spectral clustering (SC) is vital for computer vision tasks, relying on robust affinity matrices for data partitioning.
• Conventional SC methods struggle with multi-feature integration and high-dimensional, redundant visual data.

Purpose of the Study:

• To develop a novel SC approach for learning robust affinity matrices from multiple features.
• To address challenges of multi-feature fusion and high-dimensional data in spectral clustering.
• To simultaneously optimize feature weights, projection matrices, and affinity weights for improved clustering.

Main Methods:

• A new spectral clustering approach is proposed to learn an affinity matrix using multiple features.
• Optimal weights for each feature are determined, alongside projection matrices for lower-dimensional spaces.
• Affinity weights are assigned per data pair based on distances in the inferred low-dimensional space, with joint optimization.

Main Results:

• The proposed method effectively learns affinity graphs and performs feature fusion simultaneously.
• Experimental results demonstrate superior clustering performance compared to existing techniques.
• The approach avoids sensitive parameter tuning like neighborhood size and Gaussian kernel bandwidth.

Conclusions:

• The novel spectral clustering method offers significant improvements in handling multi-feature and high-dimensional data.
• Joint optimization of feature fusion and low-dimensional representation leads to enhanced clustering accuracy.
• This approach provides a more robust and effective solution for complex computer vision clustering problems.