摘要
苹果生长过程中容易受到病害影响而减产,造成经济损失。大型卷积神经网络可准确识别出苹果病害,但移动设备有限的计算资源限制了该类网络在其上的具体应用。轻量级卷积神经网络可运行在移动端,并能够实现病害的实时识别,但其识别精度往往不如前者。为解决该问题,在轻量级卷积神经网络ShuffleNet V2基础上,通过调整基本残差单元结构和网络宽度,同时引入卷积块注意模块(convolutional block attention module,CBAM),提出了改进型ShuffleNet#苹果叶部病害诊断模型。以苹果疮痂病、黑腐病、锈病、健康叶片为研究对象,收集简单和复杂背景图像各2000张,通过数据增广将其扩充至40000张,构建苹果叶部病害图像数据集,应用该数据集,对苹果叶部病害诊断模型进行训练和测试。以识别准确率、模型复杂度、平均推理时间等为判别标准,并与多个现有卷积神经网络模型进行比较。结果表明:改进后的模型能有效地识别出上述2种不同背景的4类图像,在测试集上识别准确率达到98.95%,移动端单张图像的平均推理时间为39.38ms。相较于大型的ResNet101网络,该模型在准确率上仅降低0.05%,但平均推理时间缩减87.94%,在识别速度和精度上获得了较好的平衡。基于该模型,开发了一款面向Android移动端的苹果叶部病害识别应用,测试结果表明,该应用能够满足果园内上述3种病害和健康叶片的实时识别需求,可为设计高效、轻量的病害诊断模型提供思路和参考。
Apple is easy to be affected by diseases during growing,resulting in the decrease of yield and economic losses.Large convolutional neural networks can accurately identify apple diseases,but the limited computing resources of mobile devices limit the specific application of such networks.Lightweight convolutional neural network can run on the mobile terminal and realize the real time disease recognition,but the accuracy is often not as good as the former.To solve this problem,based on lightweight convolutional neural network ShuffleNet V2,this study proposed an improved ShuffleNet#apple leaf disease diagnosis model by adjusting the basic residual unit structure and network width and introducing the convolutional block attention module(CBAM).With apple scab,black rot,rust and healthy leaves as the research objects,2000 simple and complex background images were collected,which were expanded to 40000 by data enhancement,and the image data set of apple leaf disease was constructed.The diagnostic model of apple leaf disease was trained and tested by using the data set.The recognition accuracy,model complexity and average inference time were taken as the discriminant criteria and compared with several existing convolutional neural network models.The results show that the improved model can effectively recognize the four types of images with the above two different backgrounds,the recognition accuracy is 98.95%on the test set,and the average reasoning time of a single image on the mobile terminal is 39.38ms.Compared with the large ResNet101 network,the accuracy of this research model is only reduced by 0.05%,but the average reasoning time is reduced by about 87.94%,and a good balance is achieved in recognition speed and accuracy.Based on the model,an apple leaf disease identification application was developed for Android mobile terminal,and the test showed that the application could meet the realtime identification requirements of the above three diseases and healthy leaves in the orchard.This study can provide ideas and
作者
张旭
周云成
刘忠颖
李昕泽
ZHANG Xu;ZHOU Yun-cheng;LIU Zhong-ying;LI Xin-ze(College of Information and Electronic Engineering,Shenyang Agricultural University,Shenyang 110161,China)
出处
《沈阳农业大学学报》
CAS
CSCD
北大核心
2022年第1期110-118,共9页
Journal of Shenyang Agricultural University
基金
辽宁省教育厅基础研究项目(LSNJC202004)
国家重点研发计划政府间合作项目(2019YFE0197700)。
