基于神经网络架构搜索的细粒度花卉图像分类方法研究  

Fine⁃Grained Flower Image Classification Based on Neural Network Architecture Search

作　　者：郑兴凯 杨铁军[1,2] 黄琳 ZHENG Xingkai;YANG Tiejun;HUANG Lin(College of Computer Science and Engineering,Guilin University of Technology,Guilin 541004,China;College of Intelligent Medicine and Biotechnology,Guilin Medical University,Guilin 541199,China)

机构地区：[1]桂林理工大学计算机科学与工程学院,广西桂林541004 [2]桂林医学院智能医学与生物技术学院,广西桂林541199

出　　处：《河南农业科学》2024年第5期164-171,共8页Journal of Henan Agricultural Sciences

基　　金：国家自然科学基金项目(62266015);广西自然科学基金项目(2022GXNSFAA035644)。

摘　　要：为了提升深度卷积神经网络设计的自动化程度,并进一步提高细粒度花卉图像的分类准确率,提出了一种改进的基于DARTS的神经网络搜索方法,用于自动构建细粒度花卉图像分类模型。首先,通过构建注意力-卷积模块,形成全注意力-卷积搜索空间,增强网络对可判别特征的关注度。其次,通过构建具有更多浅层特征输入节点的密集连接缩减单元(DCR cell),保留更多的浅层特征信息,减少可判别特征信息的损失并促进多尺度特征融合。最后,在堆叠最佳cell时调整DCR cell的位置,构建参数量大小不一的网络模型,以便在更多的终端设备上部署。结果表明,该方法耗时4.5 h搜索到了最佳神经网络模型,在Oxford 102和Flower 17上的分类准确率分别为96.14%和94.12%。与AGNAS等方法相比,在Oxford 102上提高了1.40百分点,在Flower 17上提高了3.09百分点。To enhance the automation of deep convolutional neural network(CNN)design and improve fine⁃grained flower image classification accuracy,an advanced neural network search approach based on differentiable architecture search(DARTS)was proposed.This method automatically constructed fine⁃grained flower image classification models.Initially,an attention⁃convolution module was constructed to create a comprehensive attention⁃convolution search space,thereby increasing the network’s focus on discriminative features.Subsequently,a densely connected reduction cell(DCR cell)with more shallow feature input nodes was developed to retain additional shallow feature information,reducing the loss of discriminative feature information and promoting multi⁃scale feature fusion.Lastly,the positions of DCR cells were adjusted when stacking the best cells to create network models of varying parameter sizes,enabling deployment on a broader range of terminal devices.The results showed that this method took approximately 4.5 hours to find the optimal neural network model,achieving classification accuracies of 96.14%on the Oxford 102 dataset and 94.12%on the Flower 17 dataset.Compared with methods like AGNAS,it improved accuracy by 1.40 percentage points on Oxford 102 and 3.09 percentage points on Flower 17.

关 键 词：神经网络架构搜索 卷积神经网络 注意力机制 细粒度花卉分类 

分 类 号：S126[农业科学—农业基础科学]