基于梅尔谱图和改进ResNet34模型的鸭蛋裂纹识别算法  

Identification algorithm of duck-egg shell crack based on MEL spectrum and improved ResNet34 model

作　　者：康俊琪 肖德琴[1] 刘又夫 孔馨月 殷建军[1] KANG Junqi;XIAO Deqin;LIU Youfu;KONG Xinyue;YIN Jianjun(College of Mathematics and Informatics,South China Agricultural University/Key Laboratory of Smart Agricultural Technology in Tropical South China,Ministry of Agriculture and Rural Affairs,Guangzhou 510642,China)

机构地区：[1]华南农业大学数学与信息学院/农业农村部华南热带智慧农业技术重点实验室,广州510642

出　　处：《华中农业大学学报》2023年第3期115-122,共8页Journal of Huazhong Agricultural University

基　　金：国家现代农业产业技术体系建设专项(CARS-42-13)。

摘　　要：针对鸭蛋裂纹人工检测受主观性影响造成精度波动大等问题,利用ResNet34网络模型,提出1种基于梅尔谱图的鸭蛋裂纹识别算法。首先利用敲蛋装置收集敲蛋声音数据,再将音频转化成梅尔谱图,构建梅尔谱图数据集,然后搭建ResNet34模型,引入迁移学习机制训练模型,再通过Adam优化算法更新梯度,增加注意力机制模块并将卷积结构替换为深度可分离卷积以对网络模型进行改进,并调整参数进行优化,最后利用模型对鸭蛋裂纹进行识别。结果显示:改进的ResNet34DP_CA网络模型检测的平均准确率为92.4%,对比原始ResNet34网络模型,平均准确率提高5.5个百分点,参数量减少32%;对比其他网络模型VGG16、MobileNetv2和EfficientNet,平均准确率分别提高10.9、13.7、16.3个百分点,识别时间为21.5 ms。结果表明,所提出的基于梅尔谱图和改进ResNet34模型的鸭蛋裂纹识别算法,能够有效地对鸭蛋裂纹进行检测识别。During the production,operation and processing of duck egg,the egg shells are easily broken and microorganisms including bacteria tend to invade the egg from the shell cracks,which in turn affect the quality of the eggs and damage economic benefits of production.An identification algorithm of duck-egg shell crack based on MEL spectrum was established by using ResNet34 network model to solve the problem of the subjectivity and large fluctuation of accuracy in manual identification of duck egg shell cracks.First,the egg knocker was used to collect the sound data,and the audio was transformed into the MEL spectrum graph to construct the dataset of MEL spectrum graph.Then the ResNet34 model was built and the transfer learning mechanism was introduced to train the model.The gradient was updated by Adam optimization algorithm,the attention mechanism module was added,and the convolution structure was replaced by a deeply separable convolution to improve the network model.The parameters were adjusted for optimization and the duck egg shell cracks were identified with the model.The results showed that the average detection accuracy of the ResNet34DP_CA enhanced network model was 92.4%,which was 5.5 percentage points higher than that of the original ResNet34 network model.The quantity of parameter was reduced by 32%.Compared with other network models including VGG16,MobileNetv2 and EfficientNet,the average accuracy was improved by 10.9,13.7 and 16.3 percentage points,respectively.The recognition time was 21.5 ms.It is indicated that the established identification algorithm of duck-egg shell crack based on Mel spectrogram and the improved ResNet34 model can efficiently identify the duck-egg shell cracks.It will be of great significance to improve the economic benefits of production and to build an intelligent and modern poultry factory.

关 键 词：梅尔谱图 无损检测 深度学习 模型优化 卷积神经网络 鸭蛋裂纹识别 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] S879.3[自动化与计算机技术—控制科学与工程]