基于音频的肉鸽雌雄智能鉴别模型研究

Intelligent Identification Model for Male-female Pigeons Based on Audio

摘要:

肉鸽雌雄鉴别是种鸽养殖、配对过程的关键工作。为实现肉鸽雌雄智能化检测，本研究提出了一种基于肉鸽鸣叫声的雌雄鉴别方法，并设计了基于EGLKA模块的肉鸽鸣叫声端点检测模型EGVAN_VAD和雌雄鉴别模型EGVAN。该方法将肉鸽一维音频信号转换为二维的梅尔语谱图作为EGVAN_VAD模型的输入数据，检测肉鸽鸣叫声的音频区域，进而使用噪声门限和巴特沃斯低通滤波器过滤环境中的稳态和非稳态噪声，再将降噪后音频的梅尔语谱图作为EGVAN模型的输入数据，采用EGVAN、LSTM、GRU、TDNN、VAN模型对不同鸽龄（包括1~3月、3~6月和大于6月）的肉鸽鸣叫声进行识别。结果显示，EGVAN_VAD模型的准确率、召回率分别为93.4%、94.3%，每检测4s时长的音频耗时10ms，较其他端点检测模型综合性能最优。EGVAN模型对肉鸽雌性和雄性鸣叫声的综合识别准确率最高，分别为90.7%、90.3%，每检测3s时长的音频耗时14.1ms。经系统测试，对3~6月鸽龄的雌雄鉴别成功率可达99.5%，对各个鸽龄段平均鉴别成功率为89.0%。研究表明，EGVAN对雌雄肉鸽鸣叫声具有优异的识别性能。本研究为利用音频技术实现其他单态鸟类性别的智能化精准识别提供了技术参考。

Abstract:

Sex identification of pigeons is a crucial task in the breeding and pairing process. To enable intelligent sex identification of pigeons, a method based on pigeon vocalizations was proposed and the endpoint detection model EGVAN_VAD and the sex classification model EGVAN, were developed both based on the EGLKA module. The method transformed one-dimensional audio signals into two-dimensional Mel spectrograms as input to the EGVAN_VAD model for detecting pigeon vocal segments, followed by noise reduction by using a thresholding method and a Butterworth low-pass filter to eliminate both steady-state and transient environmental noise. The denoised Mel spectrogram was then used as input for the EGVAN model, which was compared with LSTM, GRU, TDNN, and VAN models for recognizing pigeon vocalizations across different age groups (1~3 months, 3~6 months, and more than six months). Experimental results indicated that the EGVAN_VAD model achieved an accuracy of 93.4% and the recall of 94.3%, with processing time of 10ms per 4-second audio segment, outperforming other endpoint detection models in overall performance. The EGVAN model achieved the highest recognition precision for female and male calls, reaching 90.7% and 90.3%, respectively, with processing time of 14.1ms per 3-second audio segment. For pigeons aged 3~6 months, the identification success rate reached 99.5%, while the average success rate across all age groups was 89.0% following systematic testing. These findings demonstrated that the EGVAN model exhibited excellent performance in recognizing both male and female pigeon calls. The research result can provide a technical reference for applying audio-based methods to intelligent and accurate sex identification in other monomorphic bird species.