[1]叶建华,唐辉,罗奋翔,等.基于改进MobileNet的咖啡豆缺陷检测[J].福建工程学院学报,2023,21(03):257-263.[doi:10.3969/j.issn.1672-4348.2023.03.009]
　YE Jianhua,TANG Hui,LUO Fenxiang,et al.Coffee bean defect detection based on improved MobileNet[J].Journal of FuJian University of Technology,2023,21(03):257-263.[doi:10.3969/j.issn.1672-4348.2023.03.009]
点击复制

基于改进MobileNet的咖啡豆缺陷检测()
分享到：

《福建工程学院学报》[ISSN:2097-3853/CN:35-1351/Z]

卷:
第21卷
期数:
2023年03期
页码:
257-263
栏目:
出版日期:
2023-06-25

文章信息/Info

Title:
Coffee bean defect detection based on improved MobileNet
作者:
叶建华; 唐辉; 罗奋翔; 徐欢; 徐帅龙
福建工程学院机械与汽车工程学院
Author(s):
YE Jianhua;  TANG Hui;  LUO Fenxiang;  XU Huan;  XU Shuailong
School of Mechanical and Automotive Engineering, Fujian University of Technology
关键词:
图像识别; 卷积神经网络; 深度学习; MobileNet; 缺陷检测; 咖啡豆
Keywords:
image recognition;  convolutional neural network;  deep learning;  MobileNet;  defect detection;  coffee bean
分类号:
TP391.4
DOI:
10.3969/j.issn.1672-4348.2023.03.009
文献标志码:
A
摘要:
针对现有咖啡豆缺陷检测方法鲁棒性、实时性不高的问题,提出基于改进MobileNet 的咖啡豆缺陷检测方法,构建以MobileNet 为核心的轻量级检测网络。对咖啡豆检测样本进行采集和增强,构建包含5 种类型的咖啡豆缺陷检测数据集。通过卷积通道数和卷积模块的优化调整,压缩模型参数以匹配分类任务和满足边缘设备的部署要求。引入Mish 激活函数和学习率的自适应调整方法,提升模型的收敛性能。利用迁移学习的方式优化模型参数,进一步提升模型识别准确率。实验表明,改进模型在自建咖啡豆分类检测数据集上的平均准确率为96.13%,较原模型平均准确率的93.17%提升了2.96%;模型参数则从3.21×106 减少到了0.15×106,不到原模型的5%。相比于VGG16_bn、Res?Net50、SqueezeNet 和MobileNet,改进模型准确率分别高出0.65%、1.39%、2.39%、2.96%。在提升精度的同时,该方法在参数量、内存占用量和浮点运算量上表现也为最优,能为相关农产品的缺陷检测提供参考。
Abstract:
A coffee bean defect detection method based on improved MobileNet was proposed aiming at the problems of low robustness and poor real-time performance of existing coffee bean defect detection methods. A lightweight detection network based on MobileNet was constructed. A coffee bean data set with five defect types was constructed. The parameter numbers of the model were reduced by optimizing and adjusting its convolution channel number and convolution module. Therefore, it can match the classification task and meet the deployment requirements of edge devices. Mish activation function and adaptive adjustment method of learning rate was applied to improve the convergence performance of the model. Transfer learning was used to optimize the model parameters and further improve the model accuracy. Experiment results show that the average accuracy of the improved model on the self-built dataset was 96.13%, which was 2.96% higher than the 93.17% of the original model. The model parameters were reduced from 3.21×106 to 0.15×106, which was less than 5% of the original model. The accuracy of the improved model was 0.65%, 1.39%, 2.39% and 2.96% higher than that of VGG16_bn, ResNet50, SqueezeNet and MobileNet, respectively. While improving the accuracy, this method also has the best performance in terms of parameters, memory usage and floating-point operations. Therefore, it can provide a reference for the defect detection of related agricultural products.

参考文献/References:

［1］ GUNADI I A，ARTHA I K，CHRISTYADITAMA I P，et al. Detection of coffee bean damage in the roasting process based on shape features analysis［J］. Journal of Physics：Conference Series，2020，1503（1）：012001.［2］ ARBOLEDA E R，FAJARDO A C，MEDINA R P. An image processing technique for coffee black beans identification［C］∥2018 IEEE International Conference on Innovative Research and Development （ICIRD）. Bangkok，Thailand: IEEE，2018： 1-5. ［3］ GARCA，CANDELO-BECERRA，HOYOS. Quality and defect inspection of green coffee beans using a computer vision system［J］. Applied Sciences，2019，9（19）：4195.［4］ OLIVERI P，MALEGORI C，CASALE M，et al. An innovative multivariate strategy for HSI-NIR images to automatically detect defects in green coffee［J］. Talanta，2019，199：270-276.［5］ CHEN S Y，CHANG C Y，OU C S，et al. Detection of insect damage in green coffee beans using VIS-NIR hyperspectral imaging［J］. Remote Sensing，2020，12（15）：2348.［6］ DE OLIVEIRA E M，LEME D S，BARBOSA B H G，et al. A computer vision system for coffee beans classification based on computational intelligence techniques［J］. Journal of Food Engineering，2016，171：22-27.［7］ 刘阳，高国琴. 采用改进的SqueezeNet模型识别多类叶片病害［J］. 农业工程学报，2021，37（2）：187-195.［8］ 马本学，李聪，李玉洁，等. 基于残差网络和图像处理的干制哈密大枣外部品质检测［J］. 农业机械学报，2021，52（11）：358-366.［9］ HOWARD A G，ZHU M L，CHEN B，et al. MobileNets：efficient convolutional neural networks for mobile vision applications［EB/OL］. 2017：arXiv：1704.04861. https：∥arxiv.org/abs/1704.04861.［10］ DENG J，DONG W，SOCHER R，et al. ImageNet：a large-scale hierarchical image database［C］∥2009 IEEE Conference on Computer Vision and Pattern Recognition. Miami，FL，USA: IEEE，2009：248-255.［11］ NAIR V，HINTON G E. Rectified linear units improve restricted boltzmann machines［C］∥Proceedings of the 27th International Conference on International Conference on Machine Learning. New York：ACM，2010：807-814.［12］ MISRA D. Mish：a self regularized non-monotonic activation function［EB/OL］. 2019：arXiv：1908.08681. https：∥arxiv.org/abs/1908.08681.

更新日期/Last Update: 2023-06-25