多变环境下基于多尺度卷积网络的猪个体识别

Individual Identification of Pigs based on Multi-scale Convolutional Network in a Variable Environment

【目的】通过实现复杂多变环境下非接触式猪个体身份识别,提高畜牧行业的生产效率。【方法】以猪舍环境下猪的脸部图像为基础,提出了一种基于多尺度卷积神经网络在多变环境下的猪个体身份识别模型。利用改进的多尺度网络结构,该模型实现了深度和宽度的扩展,网络深度达到了86层。网络不仅使用了对称和非对称的两种方法拆分卷积核和多通道的方法并行提取猪脸特征,还利用网络融合技术和Batch Normalization结构过滤掉通道中的冗余信息。避免了深层网络参数激增,增强了模型对猪脸特征的提取能力并提高模型的识别速度。利用预处理后的11 695张猪脸数据集训练并验证模型,通过设置7组不同环境下的对比实验,分析改进的模型在复杂环境下的识别效果。【结果】86层的基于多尺度分类网络的识别模型权重大小和每轮样本的训练时间分别为498.4 M和66 s,比16层的VGG网络权重小1140 M,每轮训练速度快8 s。利用7组测试集的对比实验的结果表明,提出的模型在7种环境下的识别率都高于其他网络,尤其是在真实养殖环境下识别率高达99.81%。当猪脸图像中出现遮挡和旋转的情况时,提出的模型识别率皆高于92%。【结论】提出的针对脸部特征的猪个体身份识别模型是有效的,并在多变环境下具有较高的识别率和鲁棒性,为实现一体化管理及追踪溯源的研究提供参考。

[Objective]By improving the accuracy of non-contact pig individual identification in complex and changing environments,the production efficiency of the livestock industry is improved.[Method]Based on the pig’s face image in the pig house,a pig individual identification model based on multi-scale convolutional neural network in a variable environment was proposed.Using the improved multi-scale network structure,the model achieved the extension of the model depth and width,and the network depth reached 86 layers.The network not only used symmetric and asymmetric methods to split the convolution kernel and multi-channel method to extract pig face features in parallel,but also used network fusion technology and Batch Normalization structure to filter out redundant information in the channel.It avoided the sudden increase of deep network parameters,enhanced the model’s ability to extract pig face features,and improved the model’s recognition speed.The pre-processed 11 695 pig face images were used to train and validate the model.Through the comparison experiments under 7 groups of different environments,the recognition effect of the improved model in this study under complex environments was analyzed.[Result]The training results show that the weight of the multi-scale classification network-based recognition model on the 86 th layer and the training time of each round of samples are 498.4 M and 66 s,which is 1140 M less than the weight of the 16-layer VGG network,and the training speed is 8 s faster.The results of comparison experiments using 7 test sets show that the recognition rate of the model proposed in this paper is higher than that of other networks in 7 environments,especially 99.81% in the real breeding environment.When occlusion and rotation occur in pig face images,the recognition rates of the models proposed in this paper are all higher than 92%.[Conclusion]The pig’s individual identification model for facial features proposed in this paper is effective,and has a high recognition rate and high robustness in a changing environment,providing a reference for the research on integrated management and traceability.