基于深度学习的猫狗细粒度识别方法  

A Fine-Grained Recognition Method for Cats and Dogs Based on Deep Learning

作　　者：石鑫雨 方虹苏 SHI Xin-yu;FANG Hong-su(School of Information Engineering,Chang'an University,Xi'an 710061,China)

机构地区：[1]长安大学,陕西西安710061

出　　处：《山东工业技术》2024年第6期49-56,共8页Journal of Shandong Industrial Technology

摘　　要：随着国内宠物市场不断发展扩大,细粒度猫狗识别对于宠物管理与监控等方面具备重要的研究意义。基于此,本文提出了一种基于YOLOv8与改进ShuffleNet-V2的猫狗细粒度识别方法。首先,利用YOLOv8算法的Backbone对猫狗图像进行特征提取,以实现对猫狗品种的准确分类,可达到较高的品种分类准确率。其后,提出了一种结合ECA注意力的ShuffleNet-V2轻量级孪生网络,通过特征学习可准确捕捉猫狗个体之间的细微差异,以达到对猫狗个体的准确识别。为验证算法的有效性,通过公开数据集及网络搜集的方式构建了猫狗品种分类及个体识别数据集,同时通过多个量化指标来对算法进行评估。实验结果表明,YOLOv8分类网络的Top1准确率可达88.1%,Top5准确率可达98.6%;孪生网络在结合ECA注意力机制后,在训练集、验证集及所构建的测试对上的准确率分别为90.8%、92.8%及92.6%。最后,通过猫狗细粒度识别可视化并结合用户友好UI界面,进一步论证了所提出方法的有效性。With the continuous development and expansion of the domestic pet market,fine-grained cat and dog recognition has important research significance for pet management and monitoring.Based on this,this article proposes a fine-grained cat and dog recognition method based on YOLOv8 and improved ShuffleNet-V2.Firstly,the YOLOv8 algorithm's Backbone is used to extract features from cat and dog images,in order to achieve accurate classification of cat and dog breeds and achieve high breed classification accuracy.Afterwards,a ShuffleNet-V2 lightweight twin network combined with ECA attention was proposed,which can accurately capture subtle differences between cat and dog individuals through feature learning,achieving accurate recognition of cat and dog individuals.To verify the effectiveness of the algorithm,a cat and dog breed classification and individual recognition dataset was constructed through public datasets and network collection,and the algorithm was evaluated through multiple quantitative indicators.The experimental results show that the Top1 accuracy of YOLOv8 classification network can reach 88.1%,and the Top5 accuracy can reach 98.6%;After combining the ECA attention mechanism,the twin network achieved accuracies of 90.8%,92.8%,and 92.6%on the training set,validation set,and constructed test pairs,respectively.Finally,the effectiveness of the proposed method was further demonstrated through fine-grained recognition visualization of cats and dogs,combined with a user-friendly UI interface.

关 键 词：深度学习 猫狗识别 YOLOv8 ShuffleNet 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]