基于不完整信息背景下麦穗识别技术的研究

发布时间：2018-04-26 18:28

本文选题：不完整信息 + Gabor局部显著性　；参考：《北京林业大学》2016年博士论文

【摘要】：作为世界著名的植物表型研究中心-澳大利亚表型工厂和南澳大学生物信息与表型研究中心将以计算机视觉技术为基础的植物表型学应用在小麦育种中,培育具有环境适应性的优良小麦品种。经过多年的长期记录积累了丰富的小麦生长的多光谱表型图像。希望在现有小麦表型分析中添加麦穗的表型研究以实现培育高产小麦品种的目的。在小麦表型分析的图像中,小麦是图像的主体,其信息通过成像设备得到充分的表达。但麦穗作为小麦的组成部分,并不属于成像的主体,其信息的表达严重不充分,因此属于不完整信息。文章利用澳大利亚表型工厂和南澳大学生物信息与表型研究中心提供的小麦可见光表型图像,以不完整信息下的麦穗识别为目标展开相关技术研究。研究内容包括不完整信息下麦穗图像的处理方法；麦穗图像特征的提取；麦穗识别模型的建立。文章对现有计算机视觉中的物体识别技术进行了分析,结果表明由于麦穗的不完整信息及小麦生长的自然特性,是无法应用目前的物体识别技术。在对小麦表型图像的研究基础上,提出了通过纹理分析的方法实现麦穗的识别。文章从空间域和频率域两个方面对麦穗的识别进行了研究。研究分析的结果表明在不完整信息背景下,无法用单一判别模型实现高准确性的麦穗识别,而必须采用层次式的模型结构：基于弱分类器的麦穗区域判别模型和基于强分类器的麦穗识别模型。针对小麦生长中由于相互遮挡、混杂、品种变化等复杂条件而造成麦穗无法识别的情况,文章在分析人类视觉注意机制的基础上提出Gabor:城局部显著性方法实现复杂条件下的麦穗识别。结果表明Gabor域局部显著性方法可以很好的实现复杂条件下的麦穗区域判断要求。在确定使用Gabor域局部显著性作为第一级区域判断模型的基础上,文章又针对第二级强分类器模型进行研究。在分析研究目前区域特征描述方法的基础上,确定利用局部二值特征(LBP)作为区域特征。经过对传统机器学习方法的研究,选择支持向量机(SVM)作为强分类器模型。研究中将确定后的层次模型应用在所有测试图像上,最终达到了91.37%的正确性。同时指出由于不完整信息的影响,在当前的小麦表型图像中只有当麦穗长到6.5mm时才能被检测出来。相对于传统机器学习的浅层学习方法,文章提出将深层次的学习方法用于不完整环境下的麦穗识别中。用深度模型作为层次模型的强分类器模型。文章建立了三个不同深度的卷积网络模型并进行研究,提出了对于具有千万个以上未知参数的深度网络的训练方法(预训练与Fine-Tune相结合)、训练数据的均衡化的方法、训练图片预处理方法。最终仅使用八百多个数据完成了深度为13层,网络未知参数为1000万的网络训练并获得麦穗判别模型,同时该模型的准确率达到98%,在GPU的平台下单一区域处理时间为0.06ms。显示了深度学习强大的分析判断能力。
[Abstract]:As the world's famous plant phenotypic Research Center - the Australian phenotypic factory and the University of South Australia bioinformatics and phenotypic Research Center, plant phenotypes based on computer vision technology are applied to wheat breeding to breed excellent wheat varieties with environmental adaptability. After years of long-term records, rich wheat has been accumulated. The multispectral phenotypic image of the growing wheat is expected to be added to the phenotypic analysis of wheat phenotypes to achieve the purpose of cultivating high yield wheat varieties. In the image analysis of the wheat phenotype, wheat is the main body of the image, and its information is fully expressed through the imaging equipment. But the wheat ear is not a component part of the wheat. The main body of the image is not fully expressed, so it belongs to incomplete information. The article uses the Australian phenotypic factory and the University of South Australia bioinformatics and phenotypic research center to study the visible light phenotypic images of wheat under the incomplete information. The research includes incomplete information. The processing method of the image of the ear of wheat, the extraction of the feature of the ear of wheat and the establishment of the model of wheat ear recognition. The article analyses the object recognition technology in the existing computer vision. The results show that the incomplete information of the wheat ear and the natural characteristics of the wheat growth can not be applied to the present technology of object recognition. On the basis of the research, the recognition of wheat ear is realized by the method of texture analysis. The paper studies the recognition of wheat ear from two aspects of spatial domain and frequency domain. The results show that high accurate ear recognition can not be realized with a single discriminant model in the background of incomplete information, but the hierarchical type must be adopted. Model structure: wheat ear region discrimination model based on weak classifier and wheat ear recognition model based on strong classifier. In view of the situation that wheat ear can not be identified because of the complex conditions such as mutual occlusion, hybrid and variety change in wheat growth, the article puts forward the local saliency side of Gabor: city on the basis of the analysis of human visual attention mechanism. The method realizes the recognition of wheat ear under complex conditions. The results show that the local saliency method in the Gabor domain can well realize the requirement of wheat ear region judgment under complex conditions. On the basis of determining the local saliency in the Gabor domain as the first level regional judgment model, the paper also studies the second level strong classifier model. On the basis of the current regional feature description method, the local two value feature (LBP) is used as the regional feature. After the study of the traditional machine learning method, the support vector machine (SVM) is selected as the strong classifier model. The determined hierarchical model is applied to all the test images, and the correctness of the model is achieved at the same time, at the same time, 91.37%. It is pointed out that the current wheat phenotypic image can be detected only when the wheat ear is long to 6.5mm because of the influence of incomplete information. Compared with the traditional learning method, the deep learning method is applied to the ear recognition in the incomplete environment. The depth model is used as the strong classification of the hierarchical model. In this paper, three convolution networks with different depths are established and studied. A training method for deep networks with more than ten thousand unknown parameters (pre training and Fine-Tune), a method of balancing the training data, and the pre training method of the picture are trained. Finally, only more than 800 data are used to complete the depth of the network. At the same time, the accuracy of the model is 98%. The accuracy of the model is up to 98%. The single area processing time of the GPU platform shows a powerful analysis and judgment ability in depth learning under the GPU platform.

【学位授予单位】：北京林业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】