[Paper Review] Fusion de classifieurs pour la classification d'images sonar

Arnaud Martin|arXiv (Cornell University)|Jun 12, 2008

Underwater Acoustics Research4 references6 citations

TL;DR

This paper proposes and evaluates high-level information fusion techniques—voting, possibility theory, and belief theory—for fusing decisions from multiple neural network classifiers in sonar image analysis. Using four texture-based features extracted from sonar images and classified via multilayer perceptrons, the study demonstrates that belief theory with distance modeling yields the highest classification accuracy, particularly for dominant sediment types like sand and rock.

ABSTRACT

In this paper, we present some high level information fusion approaches for numeric and symbolic data. We study the interest of such method particularly for classifier fusion. A comparative study is made in a context of sea bed characterization from sonar images. The classi- fication of kind of sediment is a difficult problem because of the data complexity. We compare high level information fusion and give the obtained performance.

Motivation & Objective

• To evaluate high-level information fusion techniques for improving sonar image classification performance.
• To address the challenge of classifying seabed sediments from complex, noisy sonar images with inherent uncertainty and ambiguity.
• To compare symbolic and numeric fusion approaches in a real-world remote sensing application.
• To assess the impact of class imbalance and data homogeneity on fusion performance.
• To determine the optimal fusion strategy for improving classifier reliability in sonar-based seabed characterization.

Proposed method

• Extract four distinct texture features from sonar images using different image processing techniques.
• Train four separate multilayer perceptrons, each using one of the four feature sets to classify sediment types.
• Apply three high-level fusion strategies: majority voting, possibility theory, and belief theory (Dempster-Shafer) to combine classifier outputs.
• Use a distance-based model within the belief theory framework to measure conflict and uncertainty between classifier decisions.
• Implement a fusion rule that assigns a final class based on combined evidence, with an 'uncertainty' class for conflicting decisions.
• Evaluate performance using classification rates per sediment type on a real sonar image dataset.

Experimental results

Research questions

• RQ1 Which high-level fusion strategy—voting, possibility theory, or belief theory—yields the highest classification accuracy for sonar image-based seabed characterization?
• RQ2 How does the fusion of multiple classifiers trained on different texture features improve overall classification performance compared to individual classifiers?
• RQ3 To what extent does class imbalance in the training data affect the performance of fusion methods?
• RQ4 How do symbolic (voting) and numeric (possibility and belief theory) fusion approaches compare in handling uncertainty and conflict in sonar image classification?
• RQ5 Can fusion techniques effectively mitigate the impact of rare sediment types (e.g., mud, pebbles) with low representation in the dataset?

Key findings

• The belief theory with distance modeling approach achieved the highest overall classification accuracy, outperforming voting and possibility theory.
• The highest individual classification rates were 87.3% for 'roche' (rock) and 91.3% for 'homogènes' (homogeneous), indicating strong performance on well-represented classes.
• The lowest rates were 0.9% for 'cailloutis' (pebbles) and 4.9% for 'vase' (mud), reflecting the impact of class imbalance and low sample size.
• The 'non homogènes' (non-homogeneous) class achieved 63.1% accuracy, suggesting challenges in classifying mixed or complex regions.
• The fusion of multiple classifiers significantly improved performance over individual classifiers, especially under conditions of data uncertainty and noise.
• The study highlights that data quality and base construction—particularly the presence of mixed-sediment regions—introduce significant uncertainty that affects fusion outcomes.