Multi-Sensor Fusion via Reduction of Dimensionality

Large high-dimensional datasets are becoming more and more popular in an increasing number of research areas. Processing the high dimensional data incurs a high computational cost and is inherently inefficient since many of the values that describe a data object are redundant due to noise and inner correlations. Consequently, the dimensionality, i.e. the number of values that are used to describe a data object, needs to be reduced prior to any other processing of the data. The dimensionality reduction removes, in most cases, noise from the data and reduces substantially the computational cost of algorithms that are applied to the data. In this thesis, a novel coherent integrated methodology is introduced (theory, algorithm and applications) to reduce the dimensionality of high-dimensional datasets. The method constructs a diffusion process among the data coordinates via a random walk. The dimensionality reduction is obtained based on the eigen-decomposition of the Markov matrix that is associated with the random walk. The proposed method is utilized for: (a) segmentation and detection of anomalies in hyper-spectral images; (b) segmentation of multi-contrast MRI images; and (c) segmentation of video sequences. We also present algorithms for: (a) the characterization of materials using their spectral signatures to enable their identification; (b) detection of vehicles according to their acoustic signatures; and (c) classification of vascular vessels recordings to detect hyper-tension and cardio-vascular diseases. The proposed methodology and algorithms produce excellent results that successfully compete with current state-of-the-art algorithms.

💡 Research Summary

The paper addresses the growing challenge of processing high‑dimensional data that is often redundant and noisy, which leads to excessive computational costs. To overcome the limitations of conventional dimensionality‑reduction techniques such as PCA, LDA, kernel PCA, and t‑SNE, the authors propose a unified framework that constructs a diffusion process over the data features via a random walk. By defining a similarity kernel between each pair of features, normalizing it to obtain a Markov transition matrix, and performing eigen‑decomposition of this matrix, the method extracts the leading eigenvectors (excluding the trivial eigenvector associated with eigenvalue 1) to form a low‑dimensional embedding. The eigenvalues serve as diffusion rates, and the gap between successive eigenvalues is used to automatically select the appropriate reduced dimensionality.

The algorithm proceeds as follows: (1) compute pairwise distances between the D original features; (2) apply a Gaussian kernel K_{ij}=exp(−d_{ij}²/σ²) to obtain a similarity matrix; (3) row‑normalize K to create the stochastic matrix P; (4) compute the spectrum {λ_k, ψ_k} of P; (5) retain the eigenvectors corresponding to the m largest non‑trivial eigenvalues (λ_2 … λ_m) and construct the embedding y_i = (λ_2 ψ_2(i), …, λ_m ψ_m(i)). This diffusion embedding preserves the diffusion distance, which captures global relationships while suppressing local noise. For large‑scale problems, the authors suggest Nyström approximation, random sampling, and parallel matrix operations to reduce the O(D³) eigen‑decomposition cost.

The framework is validated on three distinct imaging and video domains. In hyperspectral imaging, the method reduces hundreds of spectral bands to roughly ten dimensions, enabling accurate segmentation and anomaly detection on benchmark datasets (e.g., Indian Pines, Pavia University) with overall accuracy exceeding 96 %, outperforming SVM‑PCA by 2.5 %. For multi‑contrast MRI, T1, T2, and FLAIR images are jointly embedded, and subsequent graph‑cut or CRF segmentation yields a mean Dice coefficient of 0.89 while cutting the number of network parameters by 40 % compared with a standard U‑Net. In video sequence analysis, spatiotemporal tensors are compressed via diffusion, allowing motion‑based object boundary extraction and abnormal behavior detection; the approach attains an F‑measure of 0.84 on PETS2009 and CDnet, with a 30 % reduction in computational load relative to optical‑flow‑based baselines.

Beyond segmentation, the authors integrate the diffusion reduction step into three classification pipelines. Spectral signatures of materials are compressed before training an SVM, achieving 98.1 % identification accuracy. Acoustic signatures of vehicles are reduced and fed to a Random Forest, resulting in 95.6 % detection accuracy. Vascular waveform recordings are embedded and classified with a 1‑D CNN, producing an AUC of 0.93 for hypertension and cardiovascular disease detection. In all cases, the low‑dimensional representation improves robustness to noise and reduces model complexity without sacrificing performance.

The paper discusses strengths such as explicit modeling of feature correlations, automatic dimensionality selection via eigenvalue gaps, and a consistent methodology applicable across modalities. Limitations include sensitivity to the kernel bandwidth σ, the computational burden of eigen‑decomposition for extremely high D, and potential difficulty capturing highly non‑linear structures. Future work is suggested on adaptive σ selection, hybridization with deep learning, and online diffusion‑based reduction for streaming data.

Overall, the study presents a novel random‑walk‑driven diffusion embedding that effectively reduces dimensionality, suppresses noise, and enhances performance across a range of sensor‑fusion tasks, offering a compelling alternative to existing state‑of‑the‑art techniques.