融合注意力机制的开集猪脸识别方法  

Open-set Pig Face Recognition Method Combining Attention Mechanism

作　　者：王荣 高荣华 李奇峰[1,3] 刘上豪 于沁杨[1,3] 冯璐 WANG Rong;GAO Ronghua;LI Qifeng;LIU Shanghao;YU Qinyang;FENG Lu(Research Center of Information Technology,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China;College of Information Engineering,Northwest A&F University,Yangling,Shaanxi 712100,China;National Engineering Research Center for Information Technology in Agriculture,Beijing 100097,China)

机构地区：[1]北京市农林科学院信息技术研究中心,北京100097 [2]西北农林科技大学信息工程学院,陕西杨凌712100 [3]国家农业信息化工程技术研究中心,北京100097

出　　处：《农业机械学报》2023年第2期256-264,共9页Transactions of the Chinese Society for Agricultural Machinery

基　　金：北京市自然科学基金项目(4202029);北京市农林科学院杰出科学家培育专项(JKZX202214)。

摘　　要：针对闭集猪脸识别模型无法识别训练集中未曾出现的生猪个体的问题,本文设计了一种融合注意力机制的开集猪脸识别方法,可实现开集猪脸图像识别,识别模型从未处理过的生猪个体。首先基于全局注意力机制、倒置残差结构和深度可分离卷积构建了轻量级的特征提取模块(GCDSC);然后基于高效注意力机制、Ghost卷积和残差网络设计C3ECAGhost模块,提取猪脸图像高层语义特征;最后基于MobileFaceNet网络,融合GCDSC模块、C3ECAGhost模块、SphereFace损失函数和欧氏距离度量方法,构建PigFaceNet模型,实现开集猪脸识别。实验结果表明,GCDSC模块可使模型猪脸识别的准确率提高1.05个百分点,C3ECAGhost模块可将模型准确率进一步提高0.56个百分点。PigFaceNet模型在开集猪脸识别验证中的准确率可达94.28%,比改进前提高1.61个百分点,模型占用存储空间仅为5.44 MB,在提高准确率的同时实现了模型轻量化,可为猪场智慧化养殖提供参考方案。To solve the problem that the closed-set pig face recognition model cannot recognize pig individuals that have not appeared in the training set,an open-set pig face recognition method that integrated attention mechanism was proposed,which can realize open-set pig face image recognition and recognize pig individuals that the model had never seen.Firstly,a lightweight feature extraction module(GCDSC) was constructed based on a global attention mechanism,inverted residual structure,and depth separable convolution.Secondly,C3ECAGhost module was designed based on efficient attention mechanism,Ghost convolution,and residual network to extract high-level semantic features of pig face images.Finally,based on the MobileFaceNet network,incorporating GCDSC module,C3ECAGhost module,SphereFace loss function,and Euclidean distance measurement method,the model PigFaceNet was constructed to realize open-set pig face recognition.The experimental results showed that the GCDSC module can improve the accuracy of pig face recognition by 1.05 percentage points,and the C3ECAGhost module can further improve the accuracy of the model by 0.56 percentage points.The accuracy of the PigFaceNet model in open-set pig face recognition verification can reach 94.28%,which was 1.61 percentage points higher than that before modification.The model proposed was a lightweight model with 5.44 MB parameters,which can improve the accuracy and provide a reference for intelligent breeding of pig farms.

关 键 词：猪脸识别 开集识别 注意力机制 模型轻量化 

分 类 号：S431.9[农业科学—农业昆虫与害虫防治] TP391.4[农业科学—植物保护]