[Paper Review] Multisensor Images Fusion Based on Feature-Level

Firouz Abdullah Al-Wassai, N. V. Kalyankar+1 more|arXiv (Cornell University)|Aug 20, 2011

Advanced Image Fusion Techniques17 references16 citations

TL;DR

This paper proposes a feature-level multisensor image fusion framework using Principal Component Analysis (PCA), Segment Fusion (SF), and Edge Fusion (EF) to enhance image quality and information preservation. It evaluates fusion performance using metrics like SNR, NRMSE, and DI, demonstrating that feature-level fusion outperforms traditional pixel-level methods in preserving structural and textural details across remote sensing data.

ABSTRACT

Until now, of highest relevance for remote sensing data processing and analysis have been techniques for pixel level image fusion. So, This paper attempts to undertake the study of Feature-Level based image fusion. For this purpose, feature based fusion techniques, which are usually based on empirical or heuristic rules, are employed. Hence, in this paper we consider feature extraction (FE) for fusion. It aims at finding a transformation of the original space that would produce such new features, which preserve or improve as much as possible. This study introduces three different types of Image fusion techniques including Principal Component Analysis based Feature Fusion (PCA), Segment Fusion (SF) and Edge fusion (EF). This paper also devotes to concentrate on the analytical techniques for evaluating the quality of image fusion (F) by using various methods including (SD), (En), (CC), (SNR), (NRMSE) and (DI) to estimate the quality and degree of information improvement of a fused image quantitatively.

Motivation & Objective

• To explore feature-level fusion as an alternative to conventional pixel-level fusion in multisensor remote sensing.
• To improve information preservation and image quality by transforming input images into optimized feature spaces.
• To evaluate fusion performance using multiple quantitative metrics including SNR, NRMSE, and DI.
• To compare the effectiveness of three feature-based fusion techniques: PCA, Segment Fusion, and Edge Fusion.
• To provide a systematic analytical framework for assessing fusion quality in remote sensing applications.

Proposed method

• Employed Principal Component Analysis (PCA) to extract dominant features from multisensor images, reducing dimensionality while preserving variance.
• Applied Segment Fusion (SF) by dividing images into regions and fusing based on segmentation-based feature representation.
• Used Edge Fusion (EF) to preserve and enhance edge features through gradient-based feature extraction and fusion.
• Integrated feature-level fusion by combining outputs from PCA, SF, and EF using heuristic rules to maximize information retention.
• Utilized a suite of quantitative metrics—Standard Deviation (SD), Entropy (En), Correlation Coefficient (CC), Signal-to-Noise Ratio (SNR), Normalized Root Mean Square Error (NRMSE), and Quality Index (DI)—to evaluate fusion performance.
• Applied the fusion pipeline to multispectral and panchromatic remote sensing images to assess structural and textural fidelity.

Experimental results

Research questions

• RQ1 How does feature-level fusion compare to pixel-level fusion in preserving image quality and structural details?
• RQ2 Which feature extraction method—PCA, Segment Fusion, or Edge Fusion—yields the highest information retention in fused images?
• RQ3 To what extent do standard image quality metrics (SNR, NRMSE, DI) reflect the actual improvement in fused image quality?
• RQ4 Can heuristic-based fusion rules effectively combine multiple feature-level representations for optimal fusion results?
• RQ5 What is the impact of feature-level fusion on the preservation of edges and textures in remote sensing imagery?

Key findings

• The proposed feature-level fusion framework achieved higher SNR and lower NRMSE values compared to traditional pixel-level fusion, indicating improved signal fidelity.
• Edge Fusion (EF) demonstrated superior performance in preserving fine details and edges, as confirmed by higher DI and CC values.
• Segment Fusion (SF) showed strong performance in homogeneous regions, maintaining texture consistency and reducing noise.
• PCA-based fusion provided effective dimensionality reduction and preserved global image structure, though with moderate edge preservation.
• The combination of multiple feature-level techniques using heuristic rules led to a more robust and information-rich fused output.
• The evaluation metrics (SNR, NRMSE, DI) effectively quantified improvements, with DI values increasing significantly post-fusion, indicating enhanced image quality.