基于MobileNetV2的轻量级羊身份识别模型  

A Lightweight Sheep Recognition Model Based on Improved MobileNetV2

作　　者：郭依蓓 吴颖丹[1] 邵洋琳 徐久红 李炎 GUO Yibei;WU Yingdan;SHAO Yanglin;XU Jiuhong;LI Yan(School of Science,Hubei Univ.of Tech.,Wuhan 430068,China)

机构地区：[1]湖北工业大学理学院,湖北武汉430068

出　　处：《湖北工业大学学报》2025年第2期100-106,共7页Journal of Hubei University of Technology

基　　金：湖北省社会科学基金(19Q062)。

摘　　要：针对目前基于深度学习的羊身份识别模型中存在的参数量多、计算量大,难以真正应用于羊场的问题,提出一个基于MobileNetV2进行改进的融合注意力机制和优化激活函数的轻量级羊身份识别网络模型LAAMNet。为使模型轻量化的同时进一步提高识别精度,选取了轻量级网络MobileNetV2,并在其瓶颈结构的基础上,修改了激活函数,同时增加了轻量级注意力机制。实验结果表明,与MobileNetV2相比,LAAMNet不仅模型参数量减少了约40%,而且对羊个体身份的识别准确率达到98.78%,提高了0.18个百分点。该模型较好地兼顾了识别准确率和复杂度,为实际羊场部署提供了思路。In response to the problem that the sheep recognition model has a large number of parameters and is computationally intensive,which makes it difficult to be applied to actual sheep farms,a lightweight sheep recognition model LAAMNet is proposed based on MobileNetV2 with ancombiation of attention mechanism and optimized activation function.To improve the recognition accuracy,the activation function of MobileNetV2 bottleneck was modified and a lightweight attention mechanism was added to it;to lighten the model,both the expansion factor and the number of bottleneck of the MobileNetV2 were reduced.The experimental results show that compared with MobileNetV2,LAAMNet not only reduces the number of model parameters by about 40%,but also achieves an accuracy of 98.78%for the identification of individual sheep,an improvement of 0.18 percentage points.The model provides a better balance of recognition accuracy and complexity and provides ideas for the deployment in real sheep farms.

关 键 词：羊身份识别 MobileNetV2 轻量级 注意力机制 激活函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]