基于改进EfficientNet的植物图像分类算法  

Plant image classification algorithm based on improved EfficientNet

作　　者：光金正 梁鉴如[1] 刘义生 GUANG Jinzheng;LIANG Jianru;LIU Yisheng(School of Electronic and Electric Engineering,Shanghai University of Engineering Science,Shanghai 201620,China;School of Mechanical Engineering,Suzhou University of Science and Technology,Suzhou 215000,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620 [2]苏州科技大学机械工程学院,江苏苏州215000

出　　处：《传感器与微系统》2022年第8期136-139,共4页Transducer and Microsystem Technologies

基　　金：上海自然科学基金资助项目(19ZR1421700)。

摘　　要：针对目前嵌入式设备受计算能力和存储容量的限制,难以运行模型体积较大的高精度网络的问题,提出了一种基于改进EfficientNet的植物图像分类算法。该算法是结合神经网络搜索技术,对网络的深度、宽度和分辨率按照特定的比例进行平衡放缩;同时兼顾了速度和精度,并将EfficientNet的激活函数更改为Mish激活函数,进一步提升了精度。实验结果表明:改进EfficientNet在自制植物数据集上分类准确率为97.2%,比原EfficientNet的96.8%提高了0.4%,但比MobileNetV2的94.1%提高了3.1%。在Oxford 102 Flowers数据集上,改进EfficientNet和DenseNet169的分类准确率均为97.7%,但改进EfficientNet有着更小的模型体积和计算量。因此,改进EfficientNet很适合应用在嵌入式设备末端部署。Aiming at the problem that embedded devices are limited by computing power and storage capacity,it is difficult to run a high-precision network with large model size,a plant image classification algorithm based on improved EfficientNet is proposed.The algorithm combines neural network search technology to balance the depth,width,and resolution of the network according to a specific ratio while taking into account speed and accuracy and changing the activation function of EfficientNet to Mish activation function to further improve the accuracy.The experimental results show that the classification accuracy of improved EfficientNet on the self-made plant dataset is 97.2%,which is 0.4%higher than the 96.8%of the original EfficientNet,but 3.1%higher than the 94.1%of MobileNetV2.On the Oxford 102 Flowers dataset,the classification accuracy of the improved EfficientNet and DenseNet169 are both 97.7%,but the improved EfficientNet has smaller model size and computational quantity.Therefore,the improved EfficientNet is very suitable for deployment at the end of embedded devices.

关 键 词：植物分类 图像识别 EfficientNet 迁移学习 嵌入式设备 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP23[自动化与计算机技术—计算机科学与技术]