文章摘要
陈勇,陈键,陈燚,裴植,易文超,王成,张文珠.基于深度学习的水体色度分类算法研究[J].高技术通讯(中文),2023,33(1):15~28
基于深度学习的水体色度分类算法研究
Research on water color classification algorithm based on deep learning
  
DOI:10. 3772/ j. issn. 1002-0470. 2023. 01. 002
中文关键词: 水质检测; 色度分类; 深度学习; 边缘检测; 神经网络
英文关键词: water quality detection, chromaticity classification, deep learning, edge detection, neural network
基金项目:
作者单位
陈勇 (浙江工业大学机械工程学院杭州 310023) 
陈键 (浙江工业大学机械工程学院杭州 310023) 
陈燚 (浙江工业大学机械工程学院杭州 310023) 
裴植 (浙江工业大学机械工程学院杭州 310023) 
易文超 (浙江工业大学机械工程学院杭州 310023) 
王成 (浙江工业大学机械工程学院杭州 310023) 
张文珠 (浙江工业大学机械工程学院杭州 310023) 
摘要点击次数: 834
全文下载次数: 668
中文摘要:
      运用深度学习技术进行非接触、快速水体色度检测与分类,采用无人机采集水体图像,运用色度仪对标定的图像完成分类,建立数据集。采用图像归一化处理减少环境因素对分类结果的影响,设计多特征的分步边缘检测算法,检测水域图像边缘,剔除无关像素。对VGG 16、GoogleNet-V3和ResNet 18卷积神经网络进行水体色度分类模型构建与训练,后筛选Inception结构和残差结构为基本构建单元,设计专门用于水体色度分类的WCNet 15与WCNet 21神经网络模型。在训练集上训练参数并利用验证集完成对2个模型的准确率的比较,筛选准确率高的WCNet 21模型作为最终水体色度分类模型。WCNet 21模型的最优准确率可达97.8%,满足水体色度分类需求,可应用到具体的水体色度分类工作当中。
英文摘要:
      Deep learning technology is used for non-contact, fast water color detection and classification. To establish the data set, the unmanned aerial vehicle is applied to collect the water images and the colorimeter is used to classify the calibrated images. Then images are normalized to reduce the impact of environmental factors on the classification results. A multi feature step-by-step edge detection algorithm is designed to detect the edge of the water image and eliminate irrelevant pixels. The proposed algorithm applies VGG 16, Googlenet-V3 and Resnet 18 convolutional neural networks to construct and train the water color classification model, and selects the concept structure and residual structure as the basic building units of WCNet 15 and WCNet 21 neural network models specifically for water color classification. The model parameters are trained on the training set, and the accuracy of the two models is compared by using the validation set. WCNet 21 is more suitable for water color classification as it has higher accuracy. The optimal accuracy of model WCNet 21 can reach 97.8%, which meets the standards of water color classification and can be applied to real-life water color classification.
查看全文   查看/发表评论  下载PDF阅读器
关闭

分享按钮