【作者】 潘家志；

【导师】 何勇； 陈丽能；

【作者基本信息】 浙江大学 ， 生物系统工程， 2007， 博士

【摘要】 我国的农业生产已经取得了很大成就，但面临着一些重大问题，如农药使用过量而造成的环境污染、机械化程度不足、投入高效率低等等。精细农业是我国农业发展的必然趋势。精细农业是集成了电子、计算机、信息处理技术与智能机械等技术的现代农业技术和体系。机器视觉是其中的关键性基础技术，是作业机械的感觉器官。针对作物早期管理中的自动喷洒作业的需要，重点研究了作物与杂草的识别问题。田间作物与杂草的自动识别，根据研究内容和方法的不同，可以分为以下几类：1)根据作物与杂草的位置信息的区分；2)根据光谱反射率的区分；3)根据形状的区分；4)根据颜色的区分；5)根据纹理的区分。本文的主要研究内容与结果如下：(1)采用Vis／NIR光谱技术，区分了苗期的大豆与牛筋草、空心莲子草、凹头苋等几种南方地区常见的植物。用325-1075nm波段的光谱反射率，经过DBN小波在三层分解后，可以将光谱样本压缩到114个数据。光谱样本分两期共采集了360个样本。然后从经过小波变换后的结果中选取250个样本作为输入数据建模，包括了两个阶段的样本数据，剩余的110个样本用于校验。采用了径向基函数神经网络模型，结果表明，识别正确率达到了97.3％，只有3个牛筋草样本被错误识别。所以，应用Vis／NIR光谱技术区分作物与杂草，是一种准确率高、速度快的高效方法。(2)研究了基于多光谱成像仪的图像的颜色空间的变换。多光谱成像仪的三个图像信道分别是Gn、Ir、Rd。其Ir信道的图像质量很高，特别适合用来做作物与杂草的区分。可以将多光谱成像仪的图像与其它的颜色空间进行转换(HSV，OHTA，CIE XYZ，CIEL~*A~*B~*，CIEL~*U~*V~*)。比较了原始图像、CIE XYZ颜色空间、CIE LUV颜色空间的图像质量和灰度直方图分布，结果是在CIE XYZ颜色空间中，三个图像分量的灰度直方图和图像质量都有改善，在CIEL~*U~*V~*颜色空间中的V~*空间的灰度直方图类似原始图像中的IR信道，图像更清晰。说明了颜色空间变换对于多光谱图像的处理，可以作为一种重要的预处理手段。(3)在图像增强的方法上，分空域增强和频域增强两部分内容。在空域增强部分，研究了平滑、中值滤波、维纳滤波、对比度增强滤波等方法。在频域增强部分，重点研究了基于matlab的数字滤波器处理方法，并分别用IIR、FIR的数字高通、低通滤波器处理模糊的作物与杂草的近红外通道图像，并比较了处理的效果。结果表明，数字滤波器的设计灵活，可以根据需要的幅频响应函数去设计滤波器参数，得到明确的幅频、相频响应函数。在选取截止频率的问题上，提出了用图像的傅立叶变换后的图像的半径来确定截止频率的方法，并研究了其使用效果。证明了基于数字滤波器的处理方法是一种有效的图像增强方法，并且可以用来处理模糊图像，增强作物、杂草与背景的对比和清晰程度，为目标识别作预处理。(4)运用阈值分割、数学形态学方法和图像分析等，结合先验知识研究了作物与杂草的识别问题。根据含有牛筋草、空心莲子草和豆苗的多光谱近红外信道的图像，首先用阈值分割将图像上的土壤背景去除。然后，用数学形态学方法，经过连续的腐蚀与膨胀操作，将豆苗与两种杂草分割开。对于仅剩下两种杂草的二值图像，用图像分析工具统计杂草对象的特征，包括长、短轴，面积，实心度，周长等等因子。然后，根据先验知识确定两条简单的规则，识别出两种杂草。实验表明这是一种简单有效的方法。(5)研究了多光谱图像的边界提取和图像分割问题，比较了在不同的边界提取算子，包括Roberts、Prewitt、Sobel、Laplacian等边界算子作用下的效果。同时也比较了用数学形态学方法提取边界的效果。研究了基于分水岭模型的图像分割方法和基于邻域的分割方法。(6)统计了MS3100多光谱成像仪和普通相机所拍摄的图像特征。分三个通道统计象素的平均亮度、均方差等因子，数据表明，多光谱成像仪的近红外图像分量的亮度适中，明、暗部分变化大，图像清晰。三个信道表现出较大差异，而普通相机的三个图像分量统计特征相似。说明了应用多光谱成像仪识别作物与杂草，有潜在的优势。更多还原

【Abstract】 Agriculture has developed greatly in our country, but it is facing advent problems now: such as theenvironmental pollution caused by over dosage of herbicide and pesticide, and the low level ofmechanization and efficiency. Precision farming is the inevitable trend of China agriculture. It is anintegration of electronics, computer techniques, information techniques and intelligent mechanics.Machine vision is the principle technique, which makes the agriculture machines can see things andobtain intelligence. To realize the need of early management of crop seedling, weed control applyingmachine vision or other method is one important task. Based on the different method and principle,these kinds of researches may be categorized into five classed: 1) using the location information torecognize weed and crop; 2) using the spectroscopy of weed and crop; 3) using the shape information; 4)using the color information; 5) using the texture information. The main content and result of thisresearch is as following:(1) Using Vis/NIR spectroscopy techniques differentiated the soybean seedling, Goose grass,Alligator Alternanthera and Emarginate amaranth, which are often living together in southernChina. By using a hand held type of spectrometer: FieldSpec Pro FR to record the reflectanceof the leaves in 325-1075 nm band. Then using DB12 wavelet analysis to reduce one samplefrom 751 numbers to 114 numbers. All these 360 samples were drawn in two terms. Selecting250 from them to build a radial basis neural network, and using the left 110 samples tovalidating the model. The result showed that only 3 Goose grass samples were classifiedincorrectly. So, using spectroscopy to recognize crop and weed is a high speed and efficientmethod.(2) Making research on the translation of multi color space. The three channels of multi-spectralimager are Gn, Ir and Rd, which were in the 400-1100nm band. So, translate the multi-spectralimages into these color spaces such as HSV, OHTA,CIE XYZ, CIEL~*A~*B~* and CIEL~*U~*V~*can obtain some advantages, including the enhanced difference or intensity, or the histogramquality. This demonstrated that the color space translation is a good method to process themulti-spectral images.(3) As regards the enhancement of image quality, this work was divided into two parts. Firstlyusing space domain enhancement methods, such as average, winner filter, median value filterand contrast enhancement filter. In frequency domain enhancement, this research used IIR andFIR digital high pass, low pass filters based on MATLAB to process the multi-spectral images,and compared the result with the traditional frequency space enhancement. The result showedthat these kinds of methods own many merits. The frequency and phase response could beobtained as wanted. And the processing quality is higher than the commonly used way.Therefore, digital filters could be used as a highly efficient preprocessing method.(4) Based on the combination of threshold segment, morphological operations and image analysis,guided by experimental knowledge, this paper raised new method to identify soybean seedlingsand Goose grass and Alligator Alternanthera. Using threshold of the histogram to divide theplant object and soil background. Then, using continuous dilation and erosion morphologicaloperations to extract the small size weeds(Goose grass and Alligator Alternanthera) from thelarge sized soybean seedlings. Then, using the image analysis tool to compute the long axis,short axis, the centroid, the eccentricity and other characters of the objects in the imagecontained only two weeds. Guided by two experimental rules, these objects could be divided into two classes. The result showed that 90% of the weed blocks could be identified correctly.So, this method is simple and efficient to identify soybean and other two weeds.(5) Making research on the edge extraction and object segmentation problem. Comparison of resultbetween different operators, such as Roberts, Prewitt, Sobel and Laplacian was made. Also, thesegmentation result of watershed model and neighbor area were compared.(6) Making statistics on multi-spectral images and images taken by common digital camera. Theresult showed that the pixel in infrared channel image is largely different from that of other twochannels. Its average value is medium and with larger RMSE, which means that its contrast andclarity are better than images from other channels. So, this new devices owns advantage toidentify crop and weed.更多还原