基于改进MobileNetV3的笼养蛋鸡声音分类识别方法

Method for Sound Classification and Recognition for Caged Laying Hens Based on Improved MobileNetV3

摘要:

为实现笼养蛋鸡声音的准确分类，实现蛋鸡健康、情绪、生产状态等信息的智能化、非接触式检测，提出了一种基于改进MobileNetV3的笼养蛋鸡声音分类识别方法。以欣华二号蛋鸡为研究对象，采集蛋鸡在笼养条件下发出的热应激声、惊吓声、产蛋声以及鸣唱声，经过声音预处理将一维声音信号转化为三维梅尔频谱图，建立了包括8541幅梅尔频谱图的蛋鸡声音数据集。通过在MobileNetV3中引入高效通道注意力(Efficient channel attention, ECA)模块，提高了笼养蛋鸡声音分类准确率。试验结果表明，MobileNetV3-ECA模型准确率、召回率、精确率以及F1分数分别达到95.25%、95.16%、95.02%、95.08%，相比原始模型分别提高1.99、2.08、2.00、2.04个百分点。通过与分别引入坐标注意力(Coordinate attention, CA)、卷积块注意力模块(Convolutional block attention module, CBAM)的模型对比，引入ECA模块后模型准确率分别提高2.11、2.03个百分点，其他指标同样有更明显的提高。与ShuffleNetV2、DesNet121和EfficientNetV2模型相比，MobileNetV3-ECA准确率分别提高1.99、2.03、2.50个百分点。本文提出的基于MobileNetV3-ECA的蛋鸡声音分类识别方法，能够有效且准确地实现对包括热应激声在内的不同种类蛋鸡声音分类识别，为蛋鸡规模化养殖中的自动化、智能化声音检测提供了算法支持，为禽舍巡检机器人功能优化提供了参考，同时为规模化笼养蛋鸡热应激预警开辟了思路。

Abstract:

In order to achieve accurate classification of caged laying hens-sounds and intelligent, non-contact detection of laying hens-health, emotion, production status and other information, a caged laying hens-sound classification and recognition method based on improved MobileNetV3 was proposed. The heat stress sound, fright sound, egg-laying sound and singing sound produced by laying hens under cage conditions were collected from Xinhua No.2 laying hens as research object, the one-dimensional sound signals were transformed into three-dimensional Mel-spectrograms after sound pre-processing, and the laying hens’sound data set consisting of 8541 Mel-spectrograms was established. The accuracy of sound classification for caged laying hens was improved by introducing the efficient channel attention (ECA) module in MobileNetV3. The experimental results showed that the MobileNetV3-ECA model achieved 95.25%, 95.16%, 95.02% and 95.08% of accuracy, recall, precision and F1 score, representing an enhancement of 1.99, 2.08, 2.00 and 2.04 percentage points, respectively, in comparison with the original model. Comparing the models with the introduction of coordinate attention (CA) and convolutional block attention module (CBAM) respectively, the accuracy of the model was improved by 2.11 and 2.03 percentage points with the introduction of the ECA module. Significant improvements were also seen in other metrics. The accuracy of MobileNetV3-ECA was improved by 1.99, 2.03 and 2.50 percentage points compared with that of ShuffleNetV2, DesNet121 and EfficientNetV2. The MobileNetV3-ECA based sound classification and recognition method for laying hens proposed provided algorithmic support for automated and intelligent sound detection in the large-scale breeding of laying hens, and also provided a reference for the function optimization of poultry house inspection robots, and opened up a way of thinking for heat stress early warning of large-scale caged laying hens.