基于轻量级神经网络的小尾寒羊面部识别  

Sheep Face Recognition Based on Lightweight Neural Network

作　　者：孙权 宣传忠[1] 张梦宇 张曦文 赵明辉 宋硕 郝敏[1] SUN Quan;XUAN Chuan-zhong;ZHANG Meng-yu;ZHANG Xi-wen;ZHAO Ming-hui;SONG Suo;HAO Min(College of Inner Mongolia Agricultural University,Hohhot Inner Mongolia 010018)

机构地区：[1]内蒙古农业大学机电工程学院,内蒙古呼和浩特010018

出　　处：《山东农业大学学报（自然科学版）》2024年第2期254-261,共8页Journal of Shandong Agricultural University：Natural Science Edition

基　　金：内蒙古直属高校基本科研业务费(BR221032)资助。

摘　　要：为实现羊只面部身份快速识别,本文以自建数据集为研究对象,提出了一种基于SSD的轻量化检测算法。首先该算法将SSD的主干网络VGG16替换成轻量级神经网络MobileNetv2,构建了一种轻量化混合神经网络模型。其次在特征提取网络参数量为1122×32的bottleneck层前端和72×160的bottleneck层后端分别引入CA、SE、CBAM和ECA注意力机制,实验结果表明72×160的bottleneck层后端引入ECA注意力机制是最优的。最后将smoothL1损失函数替换成BalancedL1损失函数。最优模型(SSD-v2-ECA2-B)模型大小从SSD的132MB减小到56.4MB,平均精度均值为81.16%,平均帧率为64.21帧/s,相较于基础的SSD模型平均精度均值提升了0.94个百分点,模型体积减小了75.6MB,检测速度提高了5.23帧/s。利用相同数据集在不同目标检测模型上进行对比试验,与SSD模型、Faster R-CNN模型、Retinanet模型相比,平均精度均值分别提升了0.36、2.40和0.07个百分点,与改进前的模型相比具有更好的综合性能。改进模型在大幅减少模型大小及其计算量的同时使模型性能保持在一个较高的水平,为畜牧养殖数字化和智能化提供方法参考,具有较高的应用价值。In order to achieve fast facial identity recognition of sheep,this paper proposes a lightweight detection algorithm based on SSD with a self-built dataset.First,the algorithm replaces the SSD backbone network VGG16 with the lightweight neural network Mobilenet v2,and constructs a lightweight hybrid neural network model.Second,CA,SE,CBAM,and ECA attention mechanisms were introduced in the front end of the bottomleneck layer with 1122×32 parameter counts and the back end of the bottomleneck layer with 72×160,the results show that the introduction of the ECA attention mechanism at the back end of the bottleneck layer with 72×160 parameters of the feature extraction network is the most effective,and finally the smoothL1 loss function is replaced by the BalancedL1 loss function.The optimal model(SSD-v2-ECA2-B)size is reduced from 132MB of the original SSD to 56.4MB,the average accuracy mean value is 81.16%,and the average frame rate is 64.21 frames/s,which is 0.94 percentage points higher than the average accuracy mean value of the base SSD model,the model size is reduced by 75.6MB,and the detection speed is improved by 5.23 frames/s.Compared with SSD model,Faster R-CNN model and RETINANET model,the average accuracy of the two models was increased by 0.36%and 2.40.07%,respectively,compared with the improved model,it has better comprehensive performance.So the improved model can keep the model performance at a high level with the reduced model size and computational effort,and provide reference method for digitization and intelligence of animal husbandry and breeding with high application value.

关 键 词：羊脸识别 SSD目标检测算法 MobileNetv2轻量级神经网络 

分 类 号：S826[农业科学—畜牧学]