Fusion of Scores in a Detection Context Based on Alpha Integration

Area of Science:

• Signal processing
• Machine learning
• Pattern recognition

Background:

• Combining information from multiple detectors is crucial for improving detection accuracy.
• Existing score fusion methods may not be optimal for all detection scenarios.
• Alpha integration offers a flexible framework for score fusion.

Purpose of the Study:

• To adapt alpha integration for fusing scores from multiple detectors in a two-hypotheses context.
• To present and evaluate three optimization methods for alpha integration: least mean square error, ROC curve maximization, and probability of error minimization.
• To demonstrate the effectiveness of the proposed alpha integration method through experiments with simulated and real-world data.

Main Methods:

• Developed an alpha integration method tailored for score fusion in detection tasks.
• Implemented gradient algorithms for optimizing alpha integration parameters based on least mean square error, ROC curve area maximization, and probability of error minimization.
• Validated the method using simulated data for two-detector scenarios and real-world data from multimodal biometrics and sleep analysis (EEG/ECG).

Main Results:

• Simulated data experiments demonstrated performance improvements of alpha integration over individual detectors.
• Real-world biometric data analysis showed effective maximization of detection probability for a given false alarm rate.
• Sleep analysis (EEG/ECG) demonstrated successful minimization of probability of error, matching expert detections.

Conclusions:

• Alpha integration provides a robust and adaptable method for fusing detector scores.
• Optimizing fusion parameters through methods like least mean square error, ROC maximization, and error minimization significantly enhances detection system performance.
• The proposed approach shows broad applicability in diverse fields requiring accurate detection and classification.