基于深度视觉和蚁群避障算法的果园精准喷施路径规划方法

为了实现对Y-shaped果树的精准喷施,本文融合了彩色及深度视觉图像,提出了1种基于蚁群避障算法的果园最优行驶路径的规划方法。首先,对彩色图像进行图像分割处理,划分道路及果树树墙障碍区域,得出喷施设备的可行驶区域,提出了喷施行驶范围检测算法;然后,通过对深度图像和彩色图像融合的处理,将Y-shaped果树树冠边缘轮廓精准拟合形成栅格地图,并与蚁群避障算法相结合,提出了最优行驶路径规划算法。最后,对拟合曲线和Y-shaped果树树冠边缘轮廓进行检验,验证算法的拟合程度。实验结果证明,本文提出的路径规划算法可以准确地检测出果树区域,并实现对行驶路径的精准规划。     

Deep learning for real-time fruit detection and orchard fruit load estimation: benchmarking of ‘MangoYOLO’

The performance of six existing deep learning architectures were compared for the task of detection of mango fruit in images of tree canopies. Images of trees (n?=?1 515) from across five orchards were acquired at night using a 5 Mega-pixel RGB digital camera and 720 W of LED flood lighting in a rig mounted on a farm utility vehicle operating at 6 km/h. The two stage deep learning architectures of Faster R-CNN(VGG) and Faster R-CNN(ZF), and the single stage techniques YOLOv3, YOLOv2, YOLOv2(tiny) and SSD were trained both with original resolution and 512?×?512 pixel versions of 1 300 training tiles, while YOLOv3 was run only with 512?×?512 pixel images, giving a total of eleven models. A new architecture was also developed, based on features of YOLOv3 and YOLOv2(tiny), on the design criteria of accuracy and speed for the current application. This architecture, termed ‘MangoYOLO’, was trained using: (i) the 1 300 tile training set, (ii) the COCO dataset before training on the mango training set, and (iii) a daytime image training set of a previous publication, to create the MangoYOLO models ‘s’, ‘pt’ and ‘bu’, respectively. Average Precision plateaued with use of around 400 training tiles. MangoYOLO(pt) achieved a F1 score of 0.968 and Average Precision of 0.983 on a test set independent of the training set, outperforming other algorithms, with a detection speed of 8 ms per 512?×?512 pixel image tile while using just 833 Mb GPU memory per image (on a NVIDIA GeForce GTX 1070 Ti GPU) used for in-field application. The MangoYOLO model also outperformed other models in processing of full images, requiring just 70 ms per image (2 048?×?2 048 pixels) (i.e., capable of processing?~?14 fps) with use of 4 417 Mb of GPU memory. The model was robust in use with images of other orchards, cultivars and lighting conditions. MangoYOLO(bu) achieved a F1 score of 0.89 on a day-time mango image dataset. With use of a correction factor estimated from the ratio of human count of fruit in images of the two sides of sample trees per orchard and a hand harvest count of all fruit on those trees, MangoYOLO(pt) achieved orchard fruit load estimates of between 4.6 and 15.2% of packhouse fruit counts for the five orchards considered. The labelled images (1 300 training, 130 validation and 300 test) of this study are available for comparative studies.

     

Immature peach detection in colour images acquired in natural illumination conditions using statistical classifiers and neural network

Detection of immature peach fruits would help growers to create yield maps which are very useful tools for adjusting management practices during the fruit maturing stages. Machine vision algorithms were developed to detect and count immature peach fruit in natural canopies using colour images. This study was the first effort to detect immature peach fruit in natural environment to the authors’ knowledge. Captured images had various illumination conditions due to both direct sunlight and diffusive light conditions that make the fruit detection task more difficult. A training set and a validation set were used to develop and to test the algorithms. Different image scanning methods including finding potential fruit regions were developed and used to parse fruit objects in the natural canopy image. Circular Gabor texture analysis and ‘eigenfruit’ approach (inspired by the ‘eigenface’ face detection and recognition method) were used for feature extraction. Statistical classifiers, a neural network and a support vector machine classifier were built and used for detecting peach fruit. A blob analysis was performed to merge multiple detections for the same peach fruit. Performance of the classifiers and image scanning methods were introduced and evaluated. Using the proposed algorithms, 84.6, 77.9 and 71.2 % of the actual fruits were successfully detected using three different image scanning methods for the validation set.     

Feature extraction of hyperspectral images for detecting immature green citrus fruit

At an early immature growth stage of citrus, a hyperspectral camera of 369–1042 nm was employed to acquire 30 hyperspectral images in order to detect immature green fruit within citrus trees under natural illumination conditions. First, successive projections algorithm (SPA) were implemented to select 677, 804, 563, 962, and 405 nm wavebands and to construct multispectral images from the original hyperspectral images for further processing. Then, histogram threshold segmentation using NDVI of 804 and 677 nm was implemented to remove image backgrounds. Three slope parameters, calculated from the pairs 405 and 563 nm, 563 and 677 nm, and 804 and 962 nm were used to construct a classifier to identify the potential citrus fruit. Then, a marker-controlled watershed segmentation based on wavelet transform was applied to obtain potential fruit areas. Finally, a green fruit detection model was constructed according to Grey Level Co-occurrence Matrix (GLCM) texture features of the independent areas. Three supervised classifiers, logistic regression, random forest and support vector machine (SVM) were developed using texture features. The detection accuracies were 79%, 75%, and 86% for the logistic regression, random forest, and SVM models, respectively. The developed algorithm showed a great potential for identifying immature green citrus for an early yield estimation.     

A new algorithm for automatic Rumex obtusifolius detection in digital images using colour and texture features and the influence of image resolution

In Gebhardt et al. (2006) an object-oriented image classification algorithm was introduced for detecting Rumex obtusifolius (RUMOB) and other weeds in mixed grassland swards, based on shape, colour and texture features. This paper describes a new algorithm that improves classification accuracy. The leaves of the typical grassland weeds (RUMOB, Taraxacum officinale, Plantago major) and other homogeneous regions were segmented automatically in digital colour images using local homogeneity and morphological operations. Additional texture and colour features were identified that contribute to the differentiation between grassland weeds using a stepwise discriminant analysis. Maximum-likelihood classification was performed on the variables retained after discriminant analysis. Classification accuracy was improved by up to 83% and Rumex detection rates of 93% were achieved. The effect of image resolution on classification results was investigated. The eight million pixel images were upscaled in six stages to create images with decreasing pixel resolution. Rumex detection rates of over 90% were obtained at almost all resolutions, and there was only moderate misclassification of other objects to RUMOB. Image processing time ranged from 45 s for the full resolution images to 2.5 s for the lowest resolution ones.     

基于距离变换的改进分水岭算法在白细胞图像分割中的应用

在白细胞图像中,由于白细胞细胞核的存在,直接应用分水岭算法时,往往达不到较好的效果。本文提出一种结合EM聚类的改进分水岭算法。通过将EM聚类获得的图像中细胞核区域替换,然后使用基于距离变换的分水岭分割,确定白细胞区域。对距离变换后的图像采用形态学处理减少了细胞分割中的过分割现象。同时使用细胞核位置的先验条件,合并分水岭分割区域,进一步减小过分割的影响。本文方法提供一种新的将分水岭算法应用于白细胞分割的思路。同时实验证明,方法在分割精度上有着良好的表现。     

基于无人机影像的高郁闭度杉木纯林树冠参数提取

  目的  冠幅是树冠结构的重要特征因子，直接影响树木的生产力和生命力，郁闭度是反映森林冠层结构与密度以及评价森林经营管理采伐强度的重要指标之一。利用无人机可以云下飞行，易于获取图像，精度高，低成本等优势，研究无人机影像上提取树冠参数的方法，使无人机影像提取林木树冠参数的操作系统化，实现精准高效的森林资源清查和监测。  方法  以福建将乐林场杉木人工纯林为研究对象，采用四旋翼无人机影像为数据源，基于面向对象分类的方法，将杉木纯林的树冠参数从无人机影像中提取出来。面向对象分类的方法需要先利用ESP工具选取最优分割尺度，然后根据影像的分割结果将树冠对象聚为一类，进而统计每个树冠对象栅格像素个数计算出树冠冠幅面积以及林分郁闭度。  结果  面向对象分类有效地对高郁闭度林分进行了树冠的提取。在分割尺度为70时，单木树冠分割效果最好，树冠被单独分割出来，但也存在一定的过分割以及未分割的问题，以至于部分单木的丢失。分割结束后，对分割对象进行特征空间的优化，选取适当的分类特征，最终将研究区分为树冠和林隙两类。通过统计每个对象栅格点数，计算得出的林分因子包括林分郁闭度，树冠面积。以地面实测数据作为参考，冠幅面积提取精度为0.829 1，林分郁闭度测量精度为0.973 1。  结论  研究结果表明，基于无人机高分辨率影像的树冠参数提取在高郁闭度林分同样适用，能有效提高森林资源调查的效率并且能够满足森林资源调查的精度。      

The recognition of apple fruits in plastic bags based on block classification

The recognition of apple fruits in plastic bags is easy to be affected by reflected and refracted light. In order to weaken the influence of light, a method based on block classification is proposed. The method adopts watershed algorithm to segment original images into irregular blocks based on edge detection results of R–G grayscale images firstly. Compared with the watershed algorithm based on gradient images, the segmentation method can preserve fruits edge and reduce the number of blocks by 20.31%, because graying image method, R–G, filters most of leaves and edge detection operator insures that the edge of fruits are detected accurately. Next, these blocks are classified into fruit blocks and non-fruit blocks by support vector machine on the basis of the color and texture features extracted from blocks. Compared with the image recognition method based on pixel classification, the proposed method can restrain the interference of light caused by plastic bags effectively. The false negative rate (FNR) and false positive rate (FPR) of the method based on pixel classification are 21.71 and 14.53% respectively. The FNR and FPR of the proposed method are 4.65 and 3.50% respectively.     

基于国产GF-2遥感影像的大麻地块提取方法研究 ——以安徽省六安市苏埠镇为例

针对传统的基于像元的分类方法提取大麻地块结果存在较为破碎、精度较低的问题,以国产"高分二号"(GF-2)4m的多光谱遥感影像为数据源,在安徽省六安市苏埠镇选取了一个研究区,使用基于规则集的面向对象的方法实现了大麻地块的精确提取。首先,对研究区预处理过的GF-2遥感影像进行多尺度分割,在多尺度分割结果的基础上,确定提取大麻地块的最优分割尺度。其次,针对不同地物类型选取样本对象生成光谱曲线,分析大麻地块与其他地物类型的异同点,并基于光谱分析结果构建规则集最终实现大麻地块的提取。最后,将基于规则集的面向对象分类结果和基于像元分类(监督分类)的结果进行对比分析。结果表明,基于规则集的面向对象方法可以有效的提取出研究区内的大麻地块,精度可以达到91.09%,解决了传统基于像元分类方法提取大麻地块结果较为破碎的问题。     

Detection and counting of immature green citrus fruit based on the Local Binary Patterns (LBP) feature using illumination-normalized images

Early detection and counting of immature green citrus fruit using computer vision can help growers produce a predictive yield map which could be used to adjust management practices during the fruit maturing stages. However, such detecting and counting is difficult because of varying illumination, random occlusion and color similarity with leaves. An immature fruit detection algorithm was developed with the aim of identifying and counting fruit in a citrus grove under varying illumination environments and random occlusions using images acquired by a regular red–green–blue (RGB) color camera. Acquired citrus images included front-lighting and back-lighting illumination conditions. The Retinex image enhancement algorithm and the two-dimensional discrete wavelet transform were used for image illumination normalization. Color-based K-means clustering and circular hough transform (CHT) were applied in order to detect potential fruit regions. A Local Binary Patterns feature-based Adaptive Boosting (AdaBoost) classifier was built for removing false positives. A sub-window was used to scan the difference image between the illumination-normalized image and the resulting image from CHT detection in order to detect small areas and partially occluded fruit. An overall accuracy of 85.6% was achieved for the validation set which showed promising potential for the proposed method.     

Low and high-level visual feature-based apple detection from multi-modal images

Automated harvesting requires accurate detection and recognition of the fruit within a tree canopy in real-time in uncontrolled environments. However, occlusion, variable illumination, variable appearance and texture make this task a complex challenge. Our research discusses the development of a machine vision system, capable of recognizing occluded green apples within a tree canopy. This involves the detection of “green” apples within scenes of “green leaves”, shadow patterns, branches and other objects found in natural tree canopies. The system uses both thermal infra-red and color image modalities in order to achieve improved performance. Maximization of mutual information is used to find the optimal registration parameters between images from the two modalities. We use two approaches for apple detection based on low and high-level visual features. High-level features are global attributes captured by image processing operations, while low-level features are strong responses to primitive parts-based filters (such as Haar wavelets). These features are then applied separately to color and thermal infra-red images to detect apples from the background. These two approaches are compared and it is shown that the low-level feature-based approach is superior (74% recognition accuracy) over the high-level visual feature approach (53.16% recognition accuracy). Finally, a voting scheme is used to improve the detection results, which drops the false alarms with little effect on the recognition rate. The resulting classifiers acting independently can partially recognize the on-tree apples, however, when combined the recognition accuracy is increased.     

基于语义分割的芒果表皮缺陷识别

【目的】运用语义分割技术自动识别芒果及其表皮缺陷,实现芒果的质量评估及分选,为芒果质量快速无损检测提供参考。【方法】采集自然环境下的多场景芒果表皮缺陷图像用于模型的训练与测试,将联合上采样金字塔(Joint pyramid upsampling,JPU)结构替换DeepLabV3+中空洞空间卷积池化金字塔(Atrous spatial pyramid pooling,ASPP),将Atrous-ResNet模型替换DeepLabV3+中Xception模型,采用类别像素准确率(Class pixel accuracy,CPA)、平均像素准确率(Mean pixel accuracy,MPA)、平均交并比(Mean intersection over union,MIoU)作为模型的精度评价指标。【结果】采用JPU模块替换ASPP模块,在ResNet网络中运用扩张卷积有利于增大模型的感受野,总体上预测的边界更加平滑,且对细小缺陷的识别更精确;与SegNet、LinkNet算法的对比验证表明,Atrous-ResNet模型具备更高的精度,CPA小幅提升,MPA提升3.79个百分点,MIoU...     

一种融合聚类和分类算法的树木图像多目标优化分割方法

目的结合树木图像颜色和纹理特征,融合聚类和分类算法对树木图像进行多目标优化分割,从而提高自然背景下树木图像分割的准确性。方法首先,利用MSCC框架理论,解决聚类和分类目标函数同时依赖于聚类中心的问题。然后,分别选定聚类性能评价指标函数和分类性能评价指标函数。最后,采用多目标进化优化方法——NSGA-II算法进行优化,得到Pareto前端最优解集,并通过计算聚类有效性指数I的最大值,寻找最优解决方案。选择具有代表性的法国梧桐、侧柏、松树和杏树等自然背景下拍摄的4幅图像作为样本。分别采用K-means、Fuzzy C-means、对聚类目标函数进行单目标优化,采用MOPSO方法进行多目标优化,以及NSGA-II方法进行多目标优化等5种方法对样本图像进行分割比较。结果在聚类中心数量相同、种群大小相同、遗传代数相同的条件下,指数I的值表明本文提出的分割方法优势显著。对于4类不同样本图像分割的指数I值进行对比可知,以HF指数为单目标函数进行遗传优化的结果优于单一使用K-means和FCM算法;MOPSO多目标优化方法分割结果优于单目标优化结果;基于NSGA-II优化的多目标函数分割结果又优于MOPSO多目标优化结果。结论融合聚类和分类算法构建聚类性能评价指标函数和分类评价性能指标函数,并采用非支配排序遗传算法对多目标函数进行优化,能更好地保留树木图像的颜色和纹理特征,分割准确率显著提高。      

自然光照条件下苹果识别方法对比研究

针对自然光照条件下果园苹果识别效果不佳的问题,从苹果的颜色分割和形状提取2方面进行对比研究,提出一种自然光照条件下的苹果识别方法。利用错检率、漏检率和处理速度3个量化指标综合对比分析颜色阈值、SVM和BPNN 3种苹果颜色分割方法的处理效果。比较6种边缘检测算法对苹果区域图像的边缘检测效果,并使用Hough圆检测算法对苹果形状进行提取,以获得苹果的圆心和半径。试验结果表明:由BPNN的苹果颜色分割方法以及结合Log和Hough的苹果形状提取方法所构建的果实识别算法具有较高的鲁棒性和准确性,能有效克服果实遮挡、重叠和颜色变异等问题,果实平均识别率可达91.6%。     

The potential of automatic methods of classification to identify leaf diseases from multispectral images

Three methods of automatic classification of leaf diseases are described based on high-resolution multispectral stereo images. Leaf diseases are economically important as they can cause a loss of yield. Early and reliable detection of leaf diseases has important practical relevance, especially in the context of precision agriculture for localized treatment with fungicides. We took stereo images of single sugar beet leaves with two cameras (RGB and multispectral) in a laboratory under well controlled illumination conditions. The leaves were either healthy or infected with the leaf spot pathogen Cercospora beticola or the rust fungus Uromyces betae. To fuse information from the two sensors, we generated 3-D models of the leaves. We discuss the potential of two pixelwise methods of classification: k-nearest neighbour and an adaptive Bayes classification with minimum risk assuming a Gaussian mixture model. The medians of pixelwise classification rates achieved in our experiments are 91% for Cercospora beticola and 86% for Uromyces betae. In addition, we investigated the potential of contextual classification with the so called conditional random field method, which seemed to eliminate the typical errors of pixelwise classification.     

吊瓶精准输液技术对芒果产量的影响

采用吊瓶精准输液技术,在芒果树开花到成熟期按树体大小定量向树杆木质部注射营养液,以确定对芒果产量的增产效应,结果表明:芒果果数可增产23.85%,单果重可增产22.77%,总产量可增产52.06%,其产投比为3.38:1,是对照产投比(2.29:1)的1.48倍。通过定量分析和计算机模拟的方法推导出芒果精确施肥用药量的数学模型为y!=5.15383(D)3.09674(g),对吊瓶输液技术导致芒果抗旱丰产的原因进行分析及讨论。     

改进C-V模型在YCbCr空间下的羊体图像分割

针对羊体图像背景复杂、分割难以及不同光照条件干扰羊体图像的问题,采用一种基于YCbCr空间改进C-V主动轮廓模型的分割方法,对具有复杂背景的羊体图像分割进行研究。结果表明:1)根据羊体图像的颜色特点,对羊体图像进行从RGB空间到YCbCr空间的转换能克服拍摄环境中光照对羊体的影响;2)利用手动勾画羊体的粗略轮廓构造预处理水平集,对其内部、外部以及边界进行划分后可以演化羊体图像的轮廓。试验证明改进C-V模型能对复杂背景下的羊体图像进行准确分割,分割结果能够应用到后续羊体测量点的识别中。     

Immature green citrus fruit detection and counting based on fast normalized cross correlation (FNCC) using natural outdoor colour images

A fast normalized cross correlation (FNCC) based machine vision algorithm was proposed in this study to develop a method for detecting and counting immature green citrus fruit using outdoor colour images toward the development of an early yield mapping system. As a template matching method, FNCC was used to detect potential fruit areas in the image, which was the very basis for subsequent false positive removal. Multiple features, including colour, shape and texture features, were combined in this algorithm to remove false positives. Circular Hough transform (CHT) was used to detect circles from images after background removal based on colour components. After building disks centred in centroids resulted from both FNCC and CHT, the detection results were merged based on the size and Euclidian distance of the intersection areas of the disks from these two methods. Finally, the number of fruit was determined after false positive removal using texture features. For a validation dataset of 59 images, 84.4 % of the fruits were successfully detected, which indicated the potential of the proposed method toward the development of an early yield mapping system.     

基于G-MRF模型的玉米叶斑病害图像的分割

【目的】图像分割是作物病害自动识别系统实现的难点之一,前人研究大多采用基于阈值或聚类的分割算法,方法简单、易于实现,但分割精度较低。本文引入高斯模型的Markov随机场分割模型(G-MRF),对玉米叶部病斑图像进行分割试验,以期提高分割精度。【方法】在VC6.0下实现了G-MRF分割模型,G-MRF既利用了图像像素的灰度信息,又通过像素类别标记的Gibbs光滑先验概率引入了图像的空间信息,是能较好地分割含有噪声图像的算法。采用该算法对大斑病、小斑病、灰斑病和弯孢菌叶斑病等4种主要玉米叶部病害的图像进行了分割测试,并与基于阈值和基于Gauss模型的分割算法进行比较。【结果】基于G-MRF分割模型的分割,目标区域的一致性和边缘的清晰方面明显好于基于阈值和Gauss模型的分割算法,其平均正确分类率达96.35%,分别较基于阈值和基于Gauss模型的分割算法高出3.75%和4.03%,差异达到显著水平。【结论】基于G-MRF模型的分割算法鲁棒性高,能够有效地将病斑区域从叶片部分离,分割正确分类率达96.35%,可用于玉米叶斑类病害图像的分割。     

Hyperspectral waveband selection for automatic detection of floral pear buds

Thinning of fruit-tree blossoms is used to regulate the yearly tree bearing and to increase the fruit yield and quality. While this is still mostly done by hand, the increasing costs of manual labor have created a demand for mechanization. This has recently led to the development of several prototype thinning machines. The main disadvantage of these machines is that they are not selective, while the fruit bearing capacity of different floral buds is not equal. On-line information about the position and distribution of the floral buds on the tree can improve the efficiency of mechanized thinning. Therefore, the aim of this study was to identify the most informative wavebands to develop a multispectral vision sensor for detection of the floral buds of the pear cultivar Conference. Hyperspectral scans were taken from tree samples in five early phenological stages to create a database of reflectance spectra for the different tree features. A stepwise algorithm was then applied to this training set to select the best combination of wavebands having the highest discriminating power between the components of interest. Subsequently, canonical correlation analysis was used to create discriminant functions out of the selected wavebands. It was possible to correctly classify 95 % of the (pixel) observations with six selected wavebands. The discrimination performance was also tested as a function of the number of used wavebands. Analysis showed that when only the two most important wavebands were used, still over 90 % of the (pixel) observations could be correctly classified.