基于结构重参数化的复杂背景下天然草地植物图像轻量级分类识别方法  

Lightweight Classification and Recognition Method of Natural Grassland Plant Image under Complex Background Based on Structure Reparameterization

作　　者：王亚鹏 曹姗姗 李全胜[1] 孙伟[2,3] WANG Yapeng;CAO Shanshan;LI Quansheng;SUN Wei(Xinjiang Agricultural University,Computer and Information Engineering College,Urumqi Xinjiang 830052,P.R.China;Agricultural Information Institute,Chinese Academy of Agricultural Sciences,Beijing 100081,P.R.China;National Agriculture Science Data Center,Beijing 100081,P.R.China)

机构地区：[1]新疆农业大学计算机与信息工程学院,新疆乌鲁木齐830052 [2]中国农业科学院农业信息研究所,北京100081 [3]国家农业科学数据中心,北京100081

出　　处：《西部林业科学》2023年第4期144-153,共10页Journal of West China Forestry Science

基　　金：国家自然科学基金项目(31860180、32271880,32060321)。

摘　　要：野外环境下天然草地植物种类的准确快速识别对草地资源调查、科学实验和教学科普等应用场景至关重要,目前多采用人工现场判别等方式,耗时耗力且受限于专家经验。以新疆干旱区天然草地植物为研究对象,构建自然复杂背景下的整株天然草地植物图像数据集。引入非对称卷积并结合结构重参数化方法优化RepVGG网络,提出并验证了一种兼顾识别精度、并行度和效率的自然复杂背景下天然草地植物图像轻量级分类识别模型(RepVGG_ACB),并与主流的经典网络模型(VGG系列和ResNet系列)以及轻量级模型(MobileNetV2和ShuffleNetV2)的识别效果进行对比分析。结果显示:(1)结构重参数化的RepVGG_ACB系列模型A0_ACB、A1_ACB和B0_ACB对天然草地植物的识别准确率为90.7%、92.4%和95.6%,模型有效且识别效果显著。(2)优化后的RepVGG_ACB网络在训练阶段采用多分支结构,识别准确率提高了1.9%~4.2%,提高了网络的泛化能力;在推理阶段采用并行度更高的单路结构,减少了FLOPs和参数量,降低了模型复杂度。(3)与经典网络模型相比,在准确率相当的情况下推理速度提升了1.3~3倍;与轻量级模型相比,推理速度虽略不及但准确率提高了2.1%~3.2%。结果表明:RepVGG_ACB系列网络在识别精度、并行度和效率方面取得均衡,具有其他网络所不具备的优势,可应用于无人机机载传感器网络或智能手持终端等边缘计算环境,为野外植物自动化高精度智能分类识别提供新方法。The accurate and rapid identification of natural grass plant species in the wild is crucial to the application scenarios of grass resource survey,scientific experiments and teaching science,etc.At present,most of the methods such as manual field identification are time-consuming and limited by experts’experience.Taking natural grassland plants in the arid region of Xinjiang as the research object,a dataset of entire natural grassland plant images under natural complex backgrounds was constructed.The RepVGG network was optimized by introducing asymmetric convolution and structural reparameterization methods.A lightweight classification and recognition model(RepVGG_ACB)for natural grassland plant images under natural complex backgrounds was proposed and validated,which takes into account recognition accuracy,parallelism,and efficiency,And compare and analyze the recognition performance with mainstream classic network models(VGG series and ResNet series)and lightweight models(MobileNet V2 and ShuffleNet V2).The results showed that:(1)the structural reparameterization of RepVGG_ACB series models A0_ACB,A1_ACB and B0_ACB were 90.7%,92.4%and 95.6%accurate in the identification of natural grassland plants,and the models were effective and significant in identification.(2)The optimized RepVGG_ACB network adopts a multi-branch structure in the training phase,which improves the recognition accuracy by 1.9%-4.2%and improves the generalization ability of the network;it adopts a single-way structure with higher parallelism in the inference phase,which reduces the number of FLOPs and parameters and decreases the model complexity.(3)Compared with the classical network model,the inference speed is improved by 1.3-3 times with comparable accuracy;compared with the lightweight model,the inference speed is slightly less but the accuracy is improved by 2.1%-3.2%.The RepVGG_ACB series network achieves a balance in recognition accuracy,parallelism and efficiency,and has advantages that other networks do not have.It can be applied to

关 键 词：草地植物分类 自然复杂背景 植物图像识别 结构重参数化 轻量级网络模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] Q94[自动化与计算机技术—计算机科学与技术]