【作者】 金燕；

【导师】 石雪晖； 熊兴耀；

【作者基本信息】 湖南农业大学 ， 果树学， 2009， 博士

【摘要】 根据葡萄生长发育规律建立的栽培技术体系是一个复杂的系统。熟悉这个系统的专家在产业的建立和发展过程中起着极其重要的作用。但是,在产业发展实践中,往往由于专家数量不够或服务不能及时到位,影响产业的健康发展。运用计算机技术建立专家系统是解决该问题的重要途径。本论文以葡萄在南方的整个生长过程作为研究对象,对其栽培管理技术、病害诊断方法,以及葡萄果实整个生长过程中内质参数与其图像信息的关系等进行了分析和研究,设计了基于神经网络和机器视觉的南方葡萄专家系统。重点围绕南方葡萄专家系统的构建、栽培管理咨询及辅助决策、病害诊断和葡萄成熟度采前无损检测四个方面进行了研究,主要研究结果如下:1.设计和构建了基于神经网络和机器视觉的南方葡萄专家系统平台。利用面向对象、.NET技术和数据库等技术设计构建了葡萄专家系统平台,同时在此平台的基础上设计了南方葡萄专家系统的体系结构、功能结构和模块设计。2.构建了栽培管理子系统。针对葡萄生产技术推广的需求设计了栽培专家决策和栽培管理信息咨询模块功能。该子系统将为用户提供实用便捷的生产管理指导和智能化的生产管理决策服务。3.设计和构建了葡萄病害诊断子系统。该子系统以18种典型的葡萄病害作为研究对象,设计了葡萄知识库和事实数据库,利用RBF神经网络模型设计和构建了葡萄病害诊断神经网络知识库和推理机。分析和论证了网络结构和训练方法,通过验证表明了该诊断模型的有效性。4.设计和构建了自然环境下葡萄成熟度采前无损检测子系统。以湖南农业大学葡萄教学实验基地的2年生欧亚种葡萄品种红宝石无核和红地球为试材,收集和测试了葡萄生长周期中图像信息和内质参数。研究了针对自然环境中的葡萄图像信息的图像分割技术和特征提取方法,并结合对应的内质参数设计了葡萄成熟度采前无损检测神经网络组合模型,通过训练后建立的葡萄成熟度采前无损检测知识库及对应的推理机,系统可以有效和准确的对红宝石无核和红地球进行无损检测。本文的创新点如下:针对南方葡萄栽培的特点,结合葡萄生产实际需要,利用领域专家知识及先进的计算机技术,构建了基于神经网络和机器视觉的南方葡萄专家系统原型。针对葡萄病害诊断复杂性和实时性问题,提出和建立了一种实时性好,诊断精度高的病害诊断模型。该模型根据诊断病害的典型和非典型设计了不同的推理规则,从而提高了系统的实时性。针对自然环境中图像信息特征提取问题,提出了一套基于边缘检测的图像分割和基于色度频度值的特征提取方法。由于葡萄背景和葡萄颜色色度相差较小,通过边缘检测和改进Hough变换的方法可以有效地分割图像中葡萄信息,同时可以通过葡萄色度信息的统计频度值来实现对图片拍摄角度和大小的依赖性,有效地实现了特征的提取。针对葡萄无损检测的要求,提出了基于频度序列和内质参数的人工神经网络组合模型。通过对内质参数的相关性分析,设计了由葡萄图像信息预测色素内质参数的人工神经网络及由果皮色素参数预测果肉总酸、总糖、可溶性固形物的人工神经网络模型,通过两个模型组合来实现根据葡萄图像信息进行采前无损检测的目的。通过独立训练和组合测试,实验准确度和精度均达到理想的要求。更多还原

【Abstract】 Cultivation technique system based on growth and development regular pattern of grape is a complicated system. The expert who is farmilar with this system plays a critical role in the establishment and development of agriculture industrialization. However, during the development of agriculture industrialization, the healthy development of industry will be strongly impacted by the shortage of these experts or providing the service behind schedule. The expert system designed through computer technology is a very important channel to solve this problem.This paper took the whole process of grape development in south as the research object, did the analysis and research for the cultivation management, grape diseases diagnostic method and the relationship of endoplasm parameters and images information for the whole process of grape development, had designed a south grape expert system based on neural network and machine vision. Mainly focused on the structure of south grape expert system, consultation and decision aids of cultivation management, grape diseases diagnostic and nondestructive measurement before harvest. The purpose was in the service of the cultivation and production of grapes better through these design system in the south.The main research contents included:1. Designed the south grape expert system platform based on neural network and machine vision, and made used of object-oriented method, .Net and database technology to constructed the expert system platform, meanwhile, designed the architecture, functional structure and module designs based on this platform.2. Constructed the subsystem of cultivation management, and designed the grape cultivation management expert decision system and information consultation functions. This subsystem provided the users with convenient and practical management guidance and intelligent decision-making of cultivation management services.3. Designed and constructed the grape diseases diagnostic subsystem. This subsystem took eighteen typical grape diseases as the research object, designed the grape knowledge base and factual database. With the RBF artificial neural network model, designed the diseases diagnostic knowledge database and inference engine. Have done the analytic demonstration of network architecture and training method. The validity of this diagnostic module was validated by experiments4. Designed and constructed the nondestructive measurement before harvest subsystem of grape under natural environment. The experimental materials were the two years grape cultivars of Ruby seedless and Red Globe, which belong to V. Vinifera. Collected and tested the images information and endoplasm parameters during the whole progress of grape development. Have done the research for the images segmentation and feature extraction method under the natural environment, and combined with correspondent endoplasm parameters. The combined neural network model was designed for nondestructive measurement before harvest. After training for this combined neural network, the subsystem could effectively and accurately do the nondestructive examination for Ruby seedless and Red Globe grape.This article has done innovative work includes the following.To consider the characteristics of grape in south, combined with the practical needs of grape production, designed the prototype of the south grape expert system based on neural networks and machine vision technology. The domain expert knowledge and advanced computer technology were used in this prototype system.A good real-time, diagnostic accuracy of disease diagnosis model was proposed, to fix the complex and real-time problem for grape disease diagnostic. In this module, different inference rules were designed for typical and atypical grape diseases, so it has improved the real-time capabilities.For the feature extraction of images information issue under natural environment, a method which was based on the edge detection of images segmentation method and the frequency of color hue was proposed. Since the grape color was very similar with the background, the edge detection and the improved Hough transform method were selected to effectively partition image. Made use of the frequency statistics of the color hue of grape image to reduce the impact of angle-dependent and size during photographing and make it is easier for feature extraction.For the nondestructive measurement of grape, a combined artificial neural network module was proposed based on frequency sequence of images information and endoplasm parameters. After the pertinence analysis of endoplasm parameters, designed an artificial neural network which built the training sample with grape images information and pericarpial pigment parameters and another artificial neural network which built the training samples with pericarpial pigment parameters and total acid, total sugar, soluble solids of sarcocarp. Using this combined module, realized the nondestructive measurement according to grape images information. The experimental accuracy and precision met the requirements of an ideal after independent training and combined test.更多还原