土地利用动态监测中多源影像融合方法的比较研究--以陕北黄土丘陵沟壑区为例

在陕北黄土丘陵沟壑区,采用单一传感器的遥感影像提取土地利用信息,存在着识别的土地利用类别少、某些类别混分现象较严重、分类结果的精度较低等问题。以TM多光谱数据和SPOT全色光谱数据的融合为例,提出了适宜于该地区的两种影像融合方法:主成分变换法和乘积运算法,并从影像的光谱质量、纹理信息和目视效果等方面对其进行了对比与评价。结果显示,主成分变换法为较理想的融合方法。以陕北无定河流域为实验样区的土地利用自动分类结果表明,该方法的应用使土地利用各类别的提取精度都有不同程度的提高;水体、水田和城镇用地等面积较小的类别分类正确率提高达到10%以上;坡耕地与林草地的混分明显减少,分类精度均提高了5%以上;分类总精度从82.0%提高到89.2%,取得了良好的分类效果。此研究对于遥感影像融合技术的评价与应用进行了有益的探索,同时为该地区的土地利用动态监测提供了关键技术。     

融合光谱与图像信息的河套蜜瓜糖度在线检测试验系统

针对蜜瓜糖度在线检测的需求,设计了融合漫透射光谱与图像信息的河套蜜瓜糖度在线检测试验系统,该系统包括硬件平台和软件系统两部分。硬件平台主要包括蜜瓜输送装置、光谱采集装置、图像采集装置和控制系统4部分。软件系统基于Microsoft Visual C++6.0语言,结合Omni Driver软件、Fly Capture2及Open CV软件开发。系统可实现蜜瓜光谱与图像信息的自动采集、显示及保存,可实现对漫透射光谱预处理,获取糖度检测所需光谱数据,对图像预处理,提取蜜瓜外观特征图像信息(R、G、B颜色值和蜜瓜体积)。在此基础上,系统可通过融合漫透射光谱与图像信息的蜜瓜糖度检测模型,计算蜜瓜糖度。系统同时可实现检测个数统计,外观特征信息及糖度的实时显示、保存等功能。测试试验表明,该试验系统检测1个样品用时1.2 s,糖度检测均方根误差为1.22,可满足河套蜜瓜糖度在线检测试验需求,为进一步开展河套蜜瓜糖度在线检测研究奠定了基础。     

排种器性能检测中种子位置智能检测方法

根据基于计算机视觉的排种器性能检测的特点，提出了基于神经网络的种子位置智能检测方法。采用两个光轴互相垂直的摄像机监测投种规律，并把以矩法求取的质心图像坐标作为输入节点，构造检测种子质心实际坐标的两个神经网络，并行处理左右两幅图像，从而快速求得种子位置。实践表明，此方法不仅能求出种子位置，而且可全面监测种子的运动情况。     

回转筒内秸秆碎料回转周期和料层完全混合时间研究

基于试验和理论分析,研究了回转筒型固态发酵罐内秸秆粉碎料的料床回转周期,建立了数学模型,其计算准确度达95%以上;结合图像分析法和料床回转周期,发现筒内上下料层完全混合时料层所转圈数与筒体转速关系不大,与筒内料体积填充率呈线性关系。由此回归并推导出只需知道筒内料体积填充率和筒体转速,便可预测上下料层完全混合的公式,其计算准确度达90%以上。提出了以上研究成果在工业级发酵罐中应用的方法。     

基于对数相似度约束Otsu的自然场景病害果实图像分割

针对自然场景下,由于复杂背景以及多变环境,水果病害果实图像分割难的问题,提出了一种基于对数相似度约束Otsu和水平集活动轮廓的近椭圆形病害果实图像分割方法。考虑背景的复杂多变,提出对数相似度约束Otsu分割来区分病害果实与背景;由于水平集活动轮廓模型的局部最优性,提出采用自适应膨胀系数的改进距离规则水平集活动轮廓模型来精确演化轮廓。先对病害果实区域样本的颜色进行混合高斯建模,获得整个病害果实图像与样本模型的对数相似度;对对数相似度进行约束Otsu阈值分割以及形态学滤波;采用最小二乘法对滤波后的曲线轮廓进行椭圆拟合,对拟合后的椭圆采用自适应膨胀系数的距离规则水平集活动轮廓演化,得到病害果实完整轮廓。对18个不同场景的病害果实进行分割,平均误判率和漏判率分别为1.77%和1.6%,实验结果表明,该方法可以从复杂自然场景图像中分割出病害果实。     

基于非线性势函数法的手写体数字识别

提出了基于非线性势函数的一种几何分类方法,设计了利用电势理论来改造决策面判别函数和利用非线性势函数对手写数字进行分类的算法。以该算法为核心构造的分类器可改善针对不同人写字习惯的分类效果。用VC 实现了该算法,结果表明该算法能减小分类时的错误率,对手写体数字的识别有较好的效果,识别率能达到97%。     

缺株玉米行中心线提取算法研究

无人驾驶农机自主进行行驶路径检测和识别系统需要具备环境感知能力。作物行的中心线识别是环境感知的一个重要方面,已有的作物行中心线识别算法在缺株作物行中心线提取中存在检测精度低的问题。该研究提出了一种能够在缺株情况下提取玉米作物行中心线的算法。首先采用限定HSV颜色空间中颜色分量范围的方法将作物与背景分割,通过形态学处理对图像进行去噪并填补作物行空洞;然后分别在图像底部和中部的横向位置设置条状感兴趣区（Region of Interest,ROI）,提取ROI内的作物行轮廓重心作为定位点。在图像顶端间隔固定步长设置上端点,利用定位点和上端点组成的扫描线扫描图像,通过作物行区域最多的扫描线即为对应目标作物行的最优线;将获取的最优线与作物行区域进行融合填充作物行中的缺株部位;最后设置动态ROI,作物行区域内面积最大轮廓拟合的直线即为目标作物行中心线。试验结果表明,对于不同缺株情况下的玉米图像,该算法的平均准确率达到84.2%,每帧图像的平均检测时间为0.092 s。该研究算法可提高缺株情况下的作物行中心线识别率,具有鲁棒性强、准确度高的特点,可为无人驾驶农机在作物行缺株的农田环境下进行作业提供理论依据。     

Automated image analysis as a tool to quantify the colour and composition of rainbow trout (Oncorhynchus mykiss W.) cutlets

The goal of this paper is to propose and evaluate automated image analysis methods for describing muscle cutlets in rainbow trout. The proposed automated image analysis methods were tested on a total of 983 scanned images of trout cutlets, and included quality traits such as fat percentage, flesh colour and the size of morphologically distinguishable subparts of the cutlet. A sub-sample of 50 images was randomly selected for manual segmentation of the cutlet, the dorsal fat depot and the red muscle and regions. The identification of these regions by manual and automatic image analysis correlated strongly (r = 0.97, r = 0.95 and r = 0.91, respectively). The estimated fat percentage obtained from image analysis, based on the area of visible fat and the colour of the cutlet flesh, correlated well with chemical fat percentage measured by mid-infrared transmission spectroscopy (MIT) (r = 0.78). The automated image analysis methods are therefore a reliable means of predicting the fat percentage of trout cutlets. Principal component analysis (PCA) loading plots were used to identify subsets of variables from the image analysis of special significance for further studies; cutlet area, dorsal fat depot area, red muscle area, back height, cutlet width, and width of left and right abdomen wall were among the variables selected. PCA loading plots of different colour variables indicated that simple statistical coefficients such as percentiles and mean values can be used to quantify different aspects of flesh colour. In conclusion, the methods presented here provide a powerful toolbox for describing important morphological structures and quality traits of trout cutlets.     

Image processing-based study of soil porosity and its effect on water movement through Andosol intact columns

The soil pore network and marcoporosity are important factors affecting water and solute transport. The transfer of contaminants to water resources is of particular importance in the Valle de Bravo watershed as it provides 10% of the drinking water for the 20 million inhabitants of Mexico City. This watershed is composed mainly of Andosols with unique mineralogical and physical characteristics. Soil porosity is usually examined on thin sections, using various image analysis techniques. We propose a novel methodology combining image analysis and a displacement experiment to study relationships between soil structure and water tracer transport parameters. H₂¹⁸O displacement experiments were conducted through intact soil columns sampled at three depths from a representative cultivated Andosol profile. The soil structure and pore characteristics were obtained by image analysis on thin sections obtained from each column at the end of the displacement experiment. The total 2D porosity (for pores larger than 50 μm) varied from 80% of the total section area in the topsoil to around 60% in the subsoil. Tubular pores were the most abundant in the soil profile, but ploughing of the topsoil had destroyed sections of these pores and replaced them with packing pores. Water transport in the intact subsoil columns was always in physical non-equilibrium, showing the existence of preferential flow pathways. In the topsoil, one column out of three showed no preferential flow, demonstrating that soil ploughing also homogenised pore connections. Pore connectivity was larger in the ploughed topsoil than in their deeper soil horizon counterparts. Our methodology offers a 2D quantitative characterisation of the macroporous network at 50 μm resolution and the determination of water transport parameters on the same intact soil samples. 3D characterisation of soil porosity using X-ray computed tomography (CT) gives a better picture of pore connection but usually has lower spatial resolution and a larger cost.     

基于图像旋转投影的导航路径检测算

提出一种农业车辆视觉导航路径识别算法--旋转投影算法.该算法通过角度枚举对图像ROI实施旋转变换,由旋转后图像的列均值与枚举角度构成旋转投影矩阵R,对其行向量实施差分运算得到差分旋转投影矩阵Rd,由Rd的极值可确定图像导航路径,即航向偏差θ与航位偏差d,进而可以求得世界坐标系下的导航路径参数.同理可以对田头线进行检测.为了提高算法的实时性,提出设定合理的ROI、实施线性压缩、旋转角度先粗分再细分的二步法以及充分利用前帧信息4种处理方法,使处理一帧用时6.2ms左右.通过对不同条件下成熟小麦图像测试表明,该算法识别导航路径准确率达到95%.