摘要
【目的】目前堆肥腐熟度主要采用复杂的化学、生物学方法进行判断,操作繁琐且效率低。卷积神经网络(convolutional neural network,CNN)模拟人类视觉,既可保留堆肥图像的颜色信息,也提取了轮廓、线条、粒度等更加具有代表性的特征,从而避免了因光照条件不同对堆肥腐熟度预测识别效果的影响。本文提出了通过堆肥图像判断堆肥腐熟度的方法,构建基于卷积神经网络的堆肥腐熟度预测模型,并验证了该模型进行堆肥腐熟度判断的准确度。【方法】供试堆肥样本采集自江苏、山东、浙江三省,堆肥原料分别为秸秆、尾菜和畜禽粪便,堆肥周期依次为50 d、45 d和60 d。在厂棚内的堆肥槽中,用海康威视摄像头(型号为C3W,焦段为广角,焦距2.8 mm,清晰度1080 p,夜间自动补光,摄像头距堆肥表面约1 m)拍摄不同腐熟时期的堆肥图像,图像格式为JPEG。分别取三种不同原料的堆肥图像样本构成三组图像数据集,将三种原料的图像按照尾菜∶秸秆∶畜禽粪便原料1∶1∶1构成第四组图像数据集。每组数据集中,80%的图像数据用于训练基于卷积神经网络模型,并建立预测模型参数。剩余20%的图像用于测试,验证模型的腐熟度预测效果。【结果】搭建的堆肥腐熟度预测模型由输入层、3层卷积层、3层池化层、2层全连接层和输出层构成。构建的腐熟度预测模型在秸秆、尾菜、畜禽粪便及三者堆肥混合图像数据集上的腐熟度预测平均准确率分别为98.7%、98.7%、98.8%和98.2%。与几种经典高效的图像特征提取、分类方法相比,较每个数据集上最优经典算法的平均准确率提升了3~14个百分点。通过CNN方法判断堆肥腐熟度,纹理特征比颜色特征更加有效。【结论】采用卷积神经网络的堆肥腐熟度预测模型能够提取堆肥图像外观特征,实现在可见光条件下直接通过堆肥图像准确、快速地识别堆肥腐熟度,可为堆肥企业生产实践提供指导。
【Objectives】The compost maturity is mainly judged by complex chemical and biological experiments,which is difficult to operate and inefficient.Convolutional neural networks simulate human vision,which can retain the color information of compost images,and extract representative features such as contour,lines and granularity at the same time,thus avoiding the influence of different illumination conditions on the prediction effect of compost maturity.This paper proposed and verified a prediction model combining images of compost and convolution neural network.【Methods】Composting samples were collected from Jiangsu,Shandong and Zhejiang provinces,and the composting materials in three provinces were straw,cauliflower residues and livestock manure,respectively,and the composting cycle was 50 days,45 days and 60 days in turn.In the factory shed,HIKVISION video camera(model C3 W)was used to take composting images of different maturing stages in JPEG format under automatic light and compensation at night.The used focal length was 2.8 mm,resolution was 1080 p,and the distance was 1 m from the compost surface.Except for the image data sets from the three composts,the fourth image data set was taken in the mixture of them three in ratio of 1∶1∶1.80%of the images from each data set was used to train CNN model and to establish prediction model parameters,and the remaining 20%to testify the prediction accuracy of the model.【Results】The compost maturity prediction model was composed of one input layer,three convolution layers,three pool layers,two full connection layers and one output layer.The accuracy of the predicted maturity from the compost images was averaged98.7%,98.7%,98.8%and 98.2%for the vegetables residues,straws,livestock manure and the mix of above,respectively.Comparing with the most optimal result of classical algorithm on each data set,the average accuracy of this method in image feature extraction and classification were improved by 3 to 14 percentage points.Texture feature was more effective than color feature in judging compost maturity by CNN method.【Conclusions】As the priority of convolutional neural network in extracting the appearance features of compost images,the compost maturity prediction model based on it could accurately and rapidly identify the compost maturity directly through compost images under natural light conditions.
作者
薛卫
胡雪娇
韦中
梅新兰
陈行健
徐阳春
XUE Wei;HU Xue-jiao;WEI Zhong;MEI Xin-lan;CHEN Xing-jian;XU Yang-chun(College of Information Science and Technology,Nanjing Agricultural University,Nanjing 210095,China;College of Resources and Environmental Sciences,Nanjing Agricultural University/Jiangsu Key Laboratory of Solid Organic Waste Utilization,Nanjing 210095,China)
出处
《植物营养与肥料学报》
CAS
CSCD
北大核心
2019年第11期1977-1988,共12页
Journal of Plant Nutrition and Fertilizers
基金
国家重点研发计划(2017YFD0800204)
关键词
堆肥
腐熟度
机器视觉
深度学习
卷积神经网络
compost
maturity
machine vision
deep learning
convolutional neural network