Deep Learning-Based Gender Recognition in Cherry Valley Ducks Through Sound Analysis.

Area of Science:

• Agricultural Science
• Bioacoustics
• Machine Learning

Background:

• Gender identification in duck farming is crucial but currently relies on labor-intensive manual methods.
• Developing an automated, efficient system for duck gender recognition is essential for the industry.

Purpose of the Study:

• To propose a novel method for distinguishing male and female day-old ducks based on their vocalization characteristics.
• To evaluate the effectiveness of different machine learning models for automated duck gender identification.

Main Methods:

• Recorded and extracted effective sound data from day-old ducks using endpoint detection.
• Calculated 36-dimensional feature vectors using Mel-frequency cepstral coefficients (MFCCs) and their difference coefficients.
• Trained and evaluated three classification models: Backpropagation Neural Network (BPNN), Deep Neural Network (DNN), and Convolutional Neural Network (CNN).

Main Results:

• Training accuracies for BPNN, DNN, and CNN were 83.87%, 83.94%, and 84.15%, respectively.
• Prediction accuracies reached 93.33% (BPNN), 91.67% (DNN), and 95.0% (CNN).
• The Convolutional Neural Network (CNN) demonstrated the highest recognition accuracy at 95.0%.

Conclusions:

• Sound-based gender identification of day-old ducks is a highly accurate method.
• The proposed automated system, particularly using CNNs, can significantly support efficient gender identification in duck production, reducing manual labor.