自然光照影响下基于深度卷积神经网络和颜色迁移的杂草识别方法  

Weed identification method based on deep convolutional neural network and color migration under the influence of natural illumination

作　　者：温德圣 许燕[1] 周建平[1] 樊湘鹏 刘洋[1] WEN Desheng;XU Yan;ZHOU Jianping;FAN Xiangpeng;LIU Yang(College of Mechanical Engineering, Xinjiang University, Urumqi 830001, China)

机构地区：[1]新疆大学机械工程学院,乌鲁木齐830001

出　　处：《中国科技论文》2020年第3期287-292,共6页China Sciencepaper

基　　金：国家级大学生创新创业训练计划项目(201810755079S)。

摘　　要：针对光照影响下杂草识别特征信息缺失和精度低等问题,提出了基于深度卷积神经网络和颜色迁移的杂草识别方法。采集不同自然光照影响下(包括阴天、反光、倒影和多云天气等)的杂草图像1334张和未受光照影响的杂草图像1436张。首先,利用优化后的Reinhard算法对受光照影响的样本图像进行颜色迁移处理,降低光照影响因素,并将杂草图像通过ExG指数灰度化及Otsu阈值图像分割得到灰度化图像;然后,通过特征提取和融合,利用卷积神经网络(convolutional neural network,CNN)构建的Inception-V3分类器进行训练,以Softmax计算识别率;最后,对比有无颜色迁移算法和ExG灰度化的杂草识别率以及光照环境对杂草识别的影响程度。结果表明,所提方法将光照环境影响下的杂草识别精度提高了15.6%,实现了90.01%的识别率。In order to solve the problem of that lacking of feature information and low accuracy of weed recognition under the influence of natural illumination,a weed recognition method based on depth convolution neural network and color migration was proposed.There were 1334 weed images collected under different natural illumination conditions(including cloudy,reflective,inverted and cloudy weather),and other 1436 images were collected without the influence of natural illumination.Firstly,the optimized Reinhard algorithm was used to process the color migration of the sample image affected by the natural illumination to reduce the influence factors of the light,and the weed image was grayed out by ExG index graying and Otsu threshold image segmentation.Then,through feature extraction and fusion,the Inception-V3 classifier constructed by convolution neural network was used for training,and the recognition rate was calculated by Softmax.Finally,the weed recognition rate with or without color migration algorithm and ExG grayscale and the influence of light environment on weed recognition were compared.The results show that the proposed method can improve the accuracy of weed identification by 15.6%and achieve a recognition rate of 90.01%.

关 键 词：光照影响的杂草 深度卷积神经网络 Inception-V3分类器 颜色迁移 

分 类 号：TP274.2[自动化与计算机技术—检测技术与自动化装置] TP391.4[自动化与计算机技术—控制科学与工程]