Weighted Feature Fusion of CNNs Utilizing Magnitude and Phase Information for Chicken Health Detection

Recent studies have shown the potential of phase features in audio processing, providing distinct insights into signal structure beyond what magnitude features can capture. In chicken health detection, however, existing models often rely heavily on magnitude-based features, neglecting phase-related information that could enhance detection performance. This paper introduces a Weighted Feature Fusion of Convolutional Neural Networks (WFF-CNNs), which integrate magnitude features, namely Mel Frequency Cepstral Coefficient (MFCC), with phase features, specifically Modified Group Delay Cepstral Coefficient (MGDCC), to enhance the identification of chicken-health issues. By integrating phase information, the proposed WFF-CNNs address an acknowledged limitations in current single-stream approaches that predominantly utilize magnitude-only features, hence augmenting the capacity of the model for capturing acoustic signals in chicken vocalizations. Compared with conventional CNNs using only MFCC, our approach overcomes the benchmark challenge of insufficient phase-based feature representation, leading to more robust detection of chicken health states. The efficacy of the proposed method is validated using audio signals from a publicly available database. Experimental results demonstrate that CNNs using MGDCC outperform conventional CNNs using MFCC, highlighting the benefits of using phase information. Furthermore, the WFF-CNNs utilizing MFCC and MGDCC outperform single-feature CNNs and existing standard models, confirming the efficacy of integrating both magnitude and phase information for accurate chicken health assessment. We anticipate that these findings will be useful in the formulation of more complete guidelines for monitoring poultry health, considering both size and phase characteristics.