摘要
针对现有樱桃缺陷检测识别中存在的问题,为实现移动端智能化快速检测与精准识别,本研究提出了一种基于卷积神经网络对樱桃图像进行缺陷检测识别的轻量化模型,可为开发樱桃的移动端无损化智能检测系统奠定理论基础。首先,将采集到的完好樱桃、刺激生长樱桃、双胞胎樱桃和腐烂樱桃4类樱桃图像经预处理后按比例划分训练集、验证集和测试集。其次,基于迁移学习对比分析NASNet-Mobile、MobileNetV2、ResNet18、InceptionV3、VGG-16网络模型后,选择各方面性能表现良好的MobileNetV2为基线模型,通过微调构建I-MobileNetV2模型;然后在I-MobileNetV2基础上,嵌入坐标注意力(CA)模块,构建ICA-MobileNetV2模型,该模型平均准确率达到97.09%,相比于基线模型(90.02%)提高7.85%,比I-MobileNetV2模型(94.34%)提高2.91%。可见,ICA-MobileNetV2作为可部署移动端的轻量化模型,具有较高准确率和较少参数,适用于樱桃缺陷检测与多分类任务,为樱桃缺陷检测与品质分级研究提供了新思路。
In view of the existing problems in cherry defect detection and recognition,and to realize in-telligent rapid detection and accurate recognition,a lightweight defect detection and recognition model based on convolutional neural network was proposed for cherry images,which could provide a theoretical basis for developing lossless intelligent detection system carried on mobile terminal.Firstly,the collected images of in-tact cherries,growth-stimulated cherries,twin cherries and rotten cherries were preprocessed,and then were divided into training,validation and test sets in proportion.Secondly,after comparing the network models such as NASNet-Mobile,MobileNetV2,ResNet18,InceptionV3 and VGG-16 based on transfer learning,the Mo-bileNetV2 with good performance in all aspects was selected as the baseline model,and then the I-Mobile-NetV2 model was established after fine tuning.On the basis of I-MobileNetV2,the coordinate attention was embedded,and then the ICA-MobileNetV2 model was constructed.The average accuracy of ICA-MobileNetV2 model reached 97.09%,which was 7.85%higher than that of baseline model(90.02%)and 2.91%higher than that of I-MobileNetV2 model(94.34%).As a deployable lightweight model,ICA-MobileNetV2 had high-er accuracy and fewer parameters,so it was suitable for chery defect detection and multi-classification tasks,which provided a new idea for cherry defect detection and quality classification research.
作者
代东南
马睿
刘起
孙孟研
马德新
Dai Dongnan;Ma Rui;Liu Qi;Sun Mengyan;Ma Dexin(College of Animation and Media,Qingdao Agricultural University,Qingdao 266109,China;Institute of Intelligent Agriculture,Qingdao Agricultural University,Qingdao 266109,China)
出处
《山东农业科学》
北大核心
2024年第3期154-162,共9页
Shandong Agricultural Sciences
基金
山东省自然科学基金项目(ZR2022MC152)
山东省高等学校青创人才引育计划项目(202202027)。
关键词
樱桃
缺陷检测
卷积神经网络
坐标注意力机制
Cherry
Defect detection
Convolutional neural network
Coordinate attention mechanism
