Sex Detection of Chicks Based on Audio Technology and Deep Learning Methods

Area of Science:

• Poultry Science
• Bioacoustics
• Machine Learning

Background:

• Early sex detection in poultry is crucial for improving breeding efficiency and economic returns.
• Chick vocalizations exhibit sex-specific differences that can be leveraged for sex identification.
• Existing methods for chick sexing often require invasive procedures or specialized equipment.

Purpose of the Study:

• To design and evaluate a non-invasive sex detection method for one-day-old chicks based on their vocalizations.
• To compare the effectiveness of different audio features and deep learning models for chick sex classification.
• To assess the performance of the developed method across three distinct chick breeds.

Main Methods:

• Utilized deep learning models including Convolutional Neural Networks (CNN), Gated Recurrent Units (GRU), Convolutional Recurrent Neural Networks (CRNN), TwoStream, and ResNet-50.
• Extracted audio features such as Spectrogram, Cepstrogram, and MFCC+Logfbank from chick calls.
• Employed short-time zero-crossing rate for automatic endpoint detection of chick calls and majority voting for final sex determination.

Main Results:

• The ResNet-50 model with MFCC+Logfbank features achieved 83% test accuracy for three-yellow chicks.
• Highest sex detection accuracies were 95% for three-yellow chicks (ResNet-50 with Spectrogram), 90% for native chicks (GRU/CRNN with Spectrogram), and 80% for flaxen-yellow chicks (TwoStream/ResNet-50).
• Performance varied significantly across breeds, with the method being most effective for three-yellow chicks.

Conclusions:

• Chick calls contain sex-differentiating information, but significant diversity exists between breeds.
• The developed deep learning-based method shows promise for chick sex detection, particularly for breeds similar to those used in training.
• Further research is needed to optimize the method for broader applicability across diverse poultry breeds.