基于深度学习的储粮害虫检测方法研究

发布时间：2018-05-31 08:29

本文选题：深度学习 + 害虫检测　；参考：《河南工业大学》2017年硕士论文

【摘要】：我国是人口大国,粮食生产大国,也是粮食储藏大国。在储粮过程中,我国每年的粮食损失约为储粮总量的0.2%至0.5%,其中由储粮害虫危害带来的损失为50%左右。因此,为了有效减少储粮害虫所带来的损失,粮虫防治已成为我国粮食安全保障的关键技术问题,而粮食害虫的检测与识别已成为粮虫防治的首要环节和关键问题。纵观国内外储粮害虫的研究现状,有扦样、引诱、声测、近红外、图像识别等多种方法进行粮虫的检测与分类,而图像识别方法因其具有高识别率、操作简易、成本低廉等特点而成为粮虫防治领域的研究热点和主要技术手段。传统图像识别的特征提取大多数是以人工方式进行的,这种方法存在诸多局限性与不足。同时基于深度学习的图像识别与分类方法已成为国内外的研究热点,深度学习通过仿生的方法,用人工神经网络来模拟人类视觉系统,以无监督方式自动学习图像的特征,可显著提高图像识别的准确率。本文针对储粮害虫检测问题,探索基于深度学习的粮食害虫的检测与识别方法。主要研究工作如下:1.对比人工神经网络、BP神经网络等浅层学习的方法,深入系统的研究了稀疏自动编码器、限制玻尔兹曼机、深度信念网络、卷积神经网络等深度学习的方法,分析了卷积神经网络的模型、结构、算法及应用演变,为基于深度学习的粮虫图像的检测与识别提供了基础。2.进行甲虫类(象甲总科、谷盗科)和蝶类(花蝶和黑脉金斑蝶)粮虫数据的采集与相应数据库的制作。设计5层的卷积神经网络模型(1个输入层,2个卷积层,2个全连接层),以Sigmoid作为激活函数,以均方误差(MES)作为损失函数,进行粮虫图像的检测与识别实验。根据实验结果,分析了基于5层的卷积神经网络模型的粮食害虫的检测与识别方法所存在的问题与不足。3.针对小样本集训练的模型不具备泛化能力的问题,本文提出了基于图像扭曲技术的粮食害虫图像样本集构造方法。通过图像尺度变换、图像旋转、图像弹性扭曲三种图像增强技术实现粮虫图像训练样本集的构造,实验表明,加入图像扭曲技术的卷积神经网络,通过训练所得到的模型具有更强的泛化能力,检测与识别效果得到了显著提高。4.针对浅层卷积神经网络训练的模型不具备复杂特征表达能力的问题,本文提出了一种基于深度卷积神经网络模型的粮食害虫的检测与识别方法。设计7层的卷积神经网络模型(1个输入层,2个卷积层,2个池化层,2个全连接层),以ReLU作为激活函数,以softmax+cross-entropy作为损失函数,采用深度学习Caffe框架实现。粮虫图像的检测与识别实验表明,所提出方法在不增加训练代价的基础上,显著提高了复杂特征的获取能力,甲虫类的检测分类率高达95%,蝶类的识别率也提高了20%。
[Abstract]:China is a populous country, a large country of grain production, and also a large country of grain storage. In the process of grain storage, the annual grain loss in China is about 0.2% to 0.5% of the total grain stored, and the loss caused by stored grain pests is about 50%. Therefore, in order to effectively reduce the losses caused by stored grain pests, the prevention and control of grain pests has become a key technical issue of grain security in China, and the detection and identification of grain pests has become the most important link and key problem in the control of grain pests. According to the research status of stored grain pests at home and abroad, there are many methods, such as sampling, luring, sound measuring, near infrared, image recognition and so on, to detect and classify grain insects. However, the image recognition method is easy to operate because of its high recognition rate. The characteristics of low cost have become the research hotspot and main technical means in the field of grain pest control. The traditional feature extraction of image recognition is mostly carried out manually, which has many limitations and shortcomings. At the same time, image recognition and classification method based on depth learning has become a hot topic at home and abroad. Through bionic method, depth learning simulates human visual system by artificial neural network, and automatically learns image features in unsupervised way. The accuracy of image recognition can be improved significantly. In this paper, the detection and identification method of grain pests based on deep learning is explored. The main research work is as follows: 1. Compared with the shallow learning methods such as artificial neural network (Ann) and BP neural network (Ann), the methods of deep learning, such as sparse automatic encoder, restricted Boltzmann machine, depth belief network and convolution neural network, are studied systematically. The model, structure, algorithm and application evolution of convolution neural network are analyzed, which provides the basis for the detection and recognition of grain insect image based on deep learning. In this paper, the data collection of beetles (Elephantae, Graconidae) and butterflies (flower butterflies and black-necked butterflies) were carried out, and the corresponding database was made. A five-layer convolution neural network model (one input layer, two convolution layers, two fully connected layers) was designed to detect and identify the grain insect images using Sigmoid as the activation function and MSE as the loss function. Based on the experimental results, the problems and shortcomings of the method of detection and identification of grain pests based on the five-layer convolution neural network model are analyzed. Aiming at the problem that the training model of small sample set does not have generalization ability, a method of constructing image sample set of grain pests based on image distortion technique is proposed in this paper. Three kinds of image enhancement techniques such as image scale transformation, image rotation and image elastic distortion are used to construct the training sample set of grain worm image. The experiment shows that the convolution neural network is added to the image distortion technology. The model obtained by training has stronger generalization ability, and the detection and recognition effect is improved significantly. 4. Aiming at the problem that the training model of shallow convolution neural network does not have the ability to express complex features, a method of detecting and identifying grain pests based on deep convolution neural network model is proposed in this paper. Seven layers of convolution neural network model (1 input layer, 2 convolution layer, 2 pool layer, 2 fully connected layer, ReLU as activation function, softmax cross-entropy as loss function and depth learning Caffe framework) are designed. The experiments on the detection and recognition of grain insect images show that the proposed method can significantly improve the ability to acquire complex features without increasing the training cost. The detection and classification rate of beetles is as high as 95 percent, and the recognition rate of butterflies is also increased by 20 percent.
【学位授予单位】：河南工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S379.5;TP391.41

【参考文献】