储粮害虫图像识别中的特征抽取研究

【作者】 张红涛；

【导师】 陈铁军； 邱道尹；

【作者基本信息】 郑州大学 ， 控制理论与控制工程， 2002， 硕士

【摘要】 我国是世界上最大的粮食生产、储藏及消费大国，搞好粮食储藏是关系到国计民生的大事。近年来，我国粮食总储量高达5000亿公斤。为了确保粮食的安全储藏，每年国家用于粮食储备方面的补贴费用就有数百亿元，但仍有不少粮食因管理决策不善等原因而遭受损失，其中，国库储粮损失率在0.2％左右，损失十分惊人，而虫害是主要因素之一。我国《“十五”粮食行业科技发展规划》明确提出要实现粮仓虫害的自动化检测。目前国内外的扦样、声测、近红外等检测方法均不能准确地在线提供粮虫的种类、密度等信息。另外，随着储粮害虫抗药性的提高，它们的种类和密度近年来有上升的趋势，这给粮虫的自动检测提出了更高的要求。因此，开发科学实用、准确方便的储粮害虫在线检测系统是很有必要的，也是极为迫切的。 利用图像识别的方法在线检测储粮害虫，具有准确度高、价格低廉、效率高、无污染、劳动量小、便于和粮库现有的计算机粮情检测系统相连接等优点，有助于粮库管理人员进行科学的决策，以及时采取合理的防治措施，达到粮食保质、保量、保鲜的目的。若使储粮损失再降低0.05％，每年可为国家挽回经济损失2.5亿元。 本文在中科院模式识别国家重点实验室开放基金与河南省自然科学基金的资助下，对储粮害虫在线检测系统进行了研究，特别对粮虫特征抽取的三个组成部分：特征形成、特征选择和特征压缩进行了比较深入的研究。特征抽取环节是识别系统的关键，因此，该课题不仅具有重要的学术价值，而且有着广阔的应用前景，可创造可观的社会和经济效益。 本文主要完成了以下工作： 1．设计、制作了第三代粮虫取样装置 本工具式装置能定点、多层自动抽取粮食样本，提供均匀恒定的无影光照，保证粮食匀速单层通过，并获取到比较清晰的序列化粮虫图像，为后续的图像处理和识别打下良好的基础。 2．粮虫图像增强与分割 运用灰度形态学图像平滑和自适应图像增强对粮虫灰度图像进行增强处理。用直方图高斯阈值法和相对熵阈值法提取粮虫图像的最优阈值，并进行分割以形成适于后续处理的二值化图像。对多种分割方法 郑什】大学硕士学位论文 摘要比较分析，得到最优性能的分割方法。3．粮虫特征形成 提取出粮虫H值化图像的面积、圆形性、不变矩等17个形态学特征。针对粮虫的灰度图像，提取出基于灰度直方图、游程长度和灰度共生矩阵的27个纹理特征，并对提取的形态学特征进行归一化处理。4．粮虫特征选择 对模拟退火算法和遗传算法两种组合优化方法的算法提出、实现步骤、参数分析、具体实现进行了深入的探讨。对抽取出的44个特征进行分析，初步筛选掉可用于离线分析的纹理特征，并从17个原始粮虫形态学特征中，选择出面积、复杂度、等效圆半径等10个适于分类的特征。5．粮虫特征压缩 运用总体类内离散度矩阵K－L变换的特征压缩，包含在类平均向量中判别信息的最优特征压缩，基于距离可分性准则的特征压缩三种方法，将粮虫的10维特征向量压缩到5维，以提高识别系统的整体性能。6．12种9类粮虫识别分类 针对粮仓中危害严重的大谷盗、米象、谷蠢、锯谷盗、黑菌虫、长头谷盗等12种9类粮虫，根据135幅模拟现场的粮虫图像，建立九类粮虫均值、方差模板库，及相应的隶属函数，在模糊极大极小原则的基础上进行识别归类。以90幅粮虫图像进行识别检验，其识别率在95％以上。7．算法比较分析 分析比较图像增强、图像分割、特征选择、特征压缩等环节的多种算法，通过比较发现：灰度形态学图像平滑、相对嫡阈值法图像分割、模拟退火算法选择特征、基于距离可分性准则的特征压缩等，为各环节效果较好的算法，并将它们作为现场应用的识别系统的最终算法。8．在线检测系统实现 利用VisualC＋＋6．0开发的粮虫识别系统软件包，与研制的取样装置相配合，能以86．5％的正确率在线识别出粮仓中危害严重的9类粮虫，为整套系统的产品化奠定了良好的基础。 文中所设计的储粮害虫在线检测系统，在郑州、民权等国家粮食储备库进行了现场试验，得到了粮食储藏方面专家的好评。在第七届全国大学生“挑战杯”、河南省第一届大学生“挑战杯”、河南省大学生科技活动日等活动中，得到有关院士、专家学者的认可，并分别获得三等奖、一等奖、银奖等。由于时间和水平有限，还需要在粮虫的获取手段、藏匿于粮粒中的幼虫识别、种类扩展等方面进一步改进、完善和提高，以进一步提高系统的性能。更多还原

【Abstract】 China is the large country of grain production, storage and consumption in the world. Doing well stored-grain management is a very important thing about the national economy and the people’s livelihood. In recent years, the stored grain is more than 500 billions of kilograms in our country. The center government offers billions of RMB to the grain deports in order to managing storage well. But plenty of storage still was attainted because of the ill management. The stored-grain loss is very serious calculated by the current 0.2% loss ratio in the national grain deports. The stored-grain insect pest is one of the important facts. "The Fifth Grain Trade Science and Technology Development Programming " put forward definitely to realize automatic detection about stored-grain insect pests. The ways of the Sampling, the Sound Detecting, the Near Infrared and others in the word can’t supply well and truly the grain pests’ category, dense and other parameters. In addition, with the increase of the stored-grain pests’ drug fastness, their category and dense are increasing in recent years. As a result, developing a kind of scientific, precision, simple detection technology for stored-grain pests is very necessary and imperious.There are a series of advantages making use of the image recognition technology to detect the stored-grain pests, such as high precise, low price, high efficiency, no pollution, less labor, convenient connection with the computer grain detection in grain deports, and so on. It can help the grain managers to make scientific decisions; in order to they can take rational prevention-measures in time, the storage can be managed in quality, quantity and greenness. If the storage loss ratio can decrease 0.05%, it may retrieve the lost 2500 millions RMB for China every year.This paper is supported by the Opening Foundation of National Laboratory of Pattern Recognition, CAS and Natural Science Foundation of Henan Province. It researched the stored-grain pests on-line detection system using of image recognition technology. The feature extraction composed of the feature forming, feature selecting and feature compressing was studied more profoundly. The aspectof feature extraction is the key of the recognition system, therefore, there is not only the important science value but also extensive utility foreground. It can create the considerable society and economical benefit. The main work is’as follows:1. Designing and making the third grain-pests sampling device. The implemental device can sample the grain automatically in the settled dot and the fixed layer, offer equality and invariable shadowless lamp, ensure the grain monolayer passing, obtain clear grain-pests image. It can settle a good foundation for the rear image enhancement and recognition.2. Enhancing and segmenting grain-pests image. The grain-pests image is smoothed by the gray morphology, enhanced by the adaptive method, segmented by the optimization thresholding offered by the histogram gauss method and the relative entropy method. Obtaining the best way in the familiar segmentation algorithms through the in-depth analysis and compare.3. Forming the stored-grain pests’ features. Extracting 17 morphological features to stored-grain pests binary image, for example area, circle quality, invariable moment and so on. Extracting 27 texture features to stored-grain pests ’ gray image. The morphological features extracted are normalized.4. Selecting the grain-pest features. The simulated annealing algorithm and genetic algorithm are studied in-depth. The two kinds of optimization algorithms’ ideal offering, realization approach, parameter analysis and material realization. Through the compare of the 44 stored-grain pests’ features, the texture features used in out-line analysis are taken out. They are used to select the 17 features normalized. There are 10 better features, for example, area, complexity, equivalent circle radius, and so on.5. Compressing the grain-pest featu更多还原