基于改进EfficientNetV2的苹果叶片病害识别模型

扫码查看

原文链接

国家科技期刊平台
NETL
NSTL
万方数据

中文摘要：斑点落叶病、褐斑病、灰斑病、花叶病和锈病是常见的5种苹果树叶部病害，严重影响苹果产量。针对实际生产中叶片病害识别准确率不高以及现有模型参数量大难以迁移到移动设备的问题，本研究基于Ef-ficientNetV2-b0模型，经过改进后提出了轻量级的EGV2-CA网络。该网络在保留EfficientNetV2-b0主干网络的基础上，一是引入GhostNetV2网络中的核心Ghost Module，并选用original分支替换第一层卷积结构，可以进一步优化网络的计算效率，减少冗余计算，从而在提高模型速度和可移植性的同时保持较高的识别准确率;二是将EfficientNetV2-b0模型中Fused-MBConv模块的SE注意力机制替换为更高效的坐标注意力(CA)机制，通过将空间信息编码为坐标信息以更好地捕捉和表达空间上的细粒度特征。实验结果显示，相较于原始网络，EGV2-CA网络的识别精确率提高2。94个百分点，召回率提高2。53个百分点，F1得分提高2。67个百分点，Top-1准确率提高2。47个百分点，而参数量仅为48。9 M，可迁移至移动设备上使用，为真实场景下苹果叶片病害的快速、准确识别提供了一种有效的解决方案。

外文标题：Apple Leaf Disease Recognition Model Based on Improved EfficientNetV2

外文摘要：Spotted leaf blight,brown spot,gray spot,Mosaic and rust are five common diseases on apple leaves,which have serious impacts on apple yield.Aiming at the problems of low recognition accuracy of leaf disease in practical production and large parameter size leading to hard to be migrated to mobile devices,this study proposed the lightweight EGV2-CA network based on the improving EfficientNetV2-b0 model.With the EfficientNetV2-b0 backbone network being reserved,the model first introduced the core Ghost Module of GhostNetV2 network and used the original branch to replace the first layer convolution structure,which could further optimize the computational efficiency of the network and reduce redundant calculations,thereby im-proved speed and portability of the model while maintaining high recognition accuracy.Then,the SE attention mechanism in the Fused-MBConv module of EfficientNetV2-b0 model was replaced with the more efficient CA attention mechanism,which was better to capture and express the fine-grained features through coding spatial information into coordinate information.The experimental results showed that compared with the original net-work,the precision rate of EGV2-CA network was increased by 2.94 percentage points,the recall rate was in-creased by 2.53 percentage points,the F1-score was increased by 2.67 percentage points,and the Top-1 accu-racy rate was increased by 2.47 percentage points,while the parameter size was only 48.9 M,so the model could be migrated to mobile devices.It provided an effective solution for apple leaf disease recognition under real scenarios.

外文关键词：

Apple leaf diseaseLightweight networkDeep learning algorithmEfficientNetV2Ghost moduleCoordinate attention mechanism

作者：

王浩宇、崔艳荣、胡玉荣、胡施威

展开 >

作者单位：

长江大学计算机科学学院,湖北荆州 434000

荆楚理工学院计算机工程学院,湖北荆门 448000

关键词：

苹果叶片病害轻量级网络深度学习算法 EfficientNetV2 Ghost模块坐标注意力机制

基金：

国家自然科学基金面上项目国家自然科学基金项目荆门市重大科技创新计划项目

项目编号：

62077018623733822022ZDYF019

出版年：

2024

DOI：

10.14083/j.issn.1001-4942.2024.09.017

山东农业科学

山东省农业科学院,山东农学会,山东农业大学

山东农业科学

CSTPCD北大核心

影响因子：0.578

ISSN：1001-4942

年,卷(期)：2024.56(9)