基于智能优化的苹果采摘机器人目标识别研究

发布时间：2018-04-01 16:42

本文选题：苹果采摘机器人　切入点：苹果图像　出处：《江苏大学》2016年博士论文

【摘要】：我国拥有世界上最大的苹果种植面积和产量,苹果的采摘作业属于高强度密集型。随着农业劳动力结构的转变,农业作业劳动力匮乏,为保证苹果的适时采摘,需提高采摘效率,因此实现苹果自动化采摘成为亟待解决的问题。现有的苹果采摘机器人样机因其采摘效率偏低,目前仍停留在实验室研究阶段。在计算机、信息技术的推动下,实现采摘机器人果园现场作业,提高采摘效率是关键,可从以下两个方面着手：一方面改善其自身性能；另一方面延长作业时间,实现全天候自动采摘。目标果实的识别效率直接制约着苹果采摘机器人的实时性和可靠性,而目标果实的精准识别作为采摘机器人视觉的技术瓶颈,进而影响着采摘效率的提高。本研究以苹果图像为研究对象,将提高采摘效率作为苹果采摘机器人的研究目标,围绕目标果实的精准识别展开相关研究。主要包括苹果夜间图像采集、夜间图像分析与降噪、图像的分割与特征提取、目标果实识别模型的建立等方面,详细探讨几种不同的夜间图像降噪算法、遗传神经网络目标果实识别算法、特殊样本建模等等。本研究的主要内容如下：1在各种人工光源辅助下采用定点标记的方式分别采集自然光、夜间苹果图像。通过对采集到的图像进行色彩分析可知,夜间图像目标果实的RGB颜色分量间的对比相对于自然光图像更为明显；发现白炽灯下的苹果夜间图像更接近自然光图像。从直观视觉观察,夜间图像较为模糊,且有椒盐噪声存在；通过对夜间图像进行差影法分析,判定其噪声类型是混合噪声,且是以高斯噪声为主,并伴有部分椒盐噪声。2针对夜间图像的噪声问题,分别提出三种智能优化的降噪算法,其中基于模糊阈值(Fuzzy threshold)改进的小波变换(Wavelet transform,WT)降噪算法(F-WT),利用模糊理论优化小波阈值潜在风险,得到的低噪图像的相对峰值信噪比(Relative peak signal-to-noise ratio,RPSNR)提高了19.69%；基于WT和独立成分分析(Independent component analysis,ICA)融合降噪算法(WT-ICA),优化后低噪图像的RPSNR提高了29.94%；基于粒子群优化(Particle swarm optimization,PSO)的ICA降噪算法(PSO-ICA),通过PSO算法求解分离矩阵,旨在改善ICA的运行效率,优化后低噪图像的RPSNR提高了21.28%。对比三种优化降噪算法,WT-ICA算法降噪能力最强,但运行效率偏低；PSO-ICA算法的降噪能力虽略低于WT-ICA算法,但运行效率却大幅度提高。白炽灯下的夜间图像,无论是原始图像还是处理后的低噪图像的RPSNR的值均最高。因此,可初步筛选白炽灯为苹果采摘机器人夜间作业的辅助光源。3对比在Lab颜色空间下采取K-means聚类算法,以及直接采用PCNN算法对苹果图像的分割效果,二者均取得较好的分割效果,PCNN的运行效率稍高于K-means聚类分割。并根据目标果实的特点,分别提取RGB和HIS颜色空间下的6个颜色特征,圆方差、椭圆方差、周长平方面积比、致密度等4个几何特征,以及7个Hu不变矩,这17个颜色和形状特征足以表征目标果实。4根据所提取到的特征向量设计分类器,将RBF、Elman两种神经网络(Neural network,NN)算法引入目标物识别,为克服它们存在的固有缺陷,利用遗传算法(Genetic algorithm,GA)进行优化,采取一种新优化方式,即连接权值和神经网络结构同时进化,分别建立GA-RBF-LMS和GA-Elman两种神经网络优化分类算法。由UCI数据仿真实验和苹果图像识别实验可知,两种识别模型的运行效率和识别精度得到大幅度提高,泛化能力也相应提高；目标果实的整体识别率均达95%以上。比较两种识别模型的算法性能,GA-RBF-LMS算法的优势在于运行效率高,而识别精度略低；GA-Elman算法的优势则在于识别精度高,而运行效率上稍微欠缺。5针对苹果识别建模过程中容易遇到的大样本问题,首先采用偏最小二乘(Partial least squares,PLS)降维算法缓解样本高维度难题,得到更有意义的低维数据。再利用谱聚类(Hierarchical cluster analysis,HCA)进行横向降维,将相近的样本划归为一类,细化样本,对每个类分别进行神经网络建模,建立PLS-HCA-NN大样本分类算法。由UCI数据测试可知,新建立的两种分类算法在其识别精度、运行效率、泛化能力均得到大幅度提高。本研究成果为苹果采摘机器人视觉系统改善、采摘效率的提高奠定了一定的理论基础,且所建立的智能优化算法也值得在其他领域进一步推广应用。
[Abstract]:China has the world's largest apple planting area and yield of apple picking is a high intensity. With the change of labor intensive agricultural structure, agricultural operation labor shortage, in order to ensure the timely picking apples, the need to improve the picking efficiency, thus achieving the apple picking automation has become an urgent problem. The existing apple picking because of the low efficiency of harvesting robot, is still in the stage of laboratory research. In the computer, the development of information technology, to realize the orchard picking robot field operations, improve the picking efficiency is the key, can proceed from the following two aspects: to improve its performance on one hand; on the other hand, prolonged operation time, all-weather automatic picking fruit. Target recognition efficiency directly restricts the real-time and reliability of apple picking robot, and accurate target identification for fruit picking machine The technical bottleneck in human visual perception, and affects the picking efficiency. Based on the apple image as the research object, will improve the picking efficiency of apple picking robot as the research target, the research on accurate target identification. The main fruits include apple night image acquisition, image analysis and noise reduction at night, segmentation and feature extraction the image, based on target recognition model, discusses several different nighttime image denoising algorithm, genetic neural network target recognition algorithm, special sample modeling and so on. The main contents of this study are as follows: 1 in a variety of artificial light assisted by fixed way of marking natural light were collected at night, apple image. Through the analysis of color images of the night, compared with the RGB color component images between the target relative to the natural light image is more Obviously; found in incandescent light Apple night image closer to natural light image. From the observation of visual image, the night is vague, and salt and pepper noise through the night; image subtraction method, determine the type of noise is mixed noise, and the noise is in Gauss, and with some salt and pepper noise.2 aiming at the problem of noise at night image, put forward three kinds of intelligent optimization algorithm, which is based on fuzzy threshold (Fuzzy threshold) improved wavelet transform (Wavelet transform WT) algorithm (F-WT), the potential risk of wavelet threshold fuzzy optimization theory, the low noise image relative to the peak signal-to-noise ratio (Relative peak signal-to-noise ratio, RPSNR) increased by 19.69%; and based on independent component analysis WT (Independent component analysis, ICA) (WT-ICA), fusion algorithm improves low noise image after optimization RPSNR 29.94%; based on particle swarm optimization (Particle swarm optimization, PSO) ICA (PSO-ICA) denoising algorithm, the separation matrix is solved through PSO algorithm, aimed at improving the ICA efficiency, optimized low noise image RPSNR improved three optimized denoising algorithm 21.28%. WT-ICA algorithm is compared, the strongest, but the efficiency is low; noise reduction ability of PSO-ICA algorithm was slightly lower than that of the WT-ICA algorithm, but the operating efficiency is greatly improved. The incandescent lamp under the night image, whether it is low noise image of the original image or after the treatment of RPSNR are the highest. Because of this, the preliminary screening of incandescent lamp for apple harvesting robot.3 auxiliary light contrast work at night to take K-means clustering algorithm in Lab color space, and the PCNN algorithm is directly used for apple image segmentation, two showed better segmentation effect, efficiency rate of PCNN was slightly higher than that of K-means Clustering segmentation. According to the characteristics of the target fruit, were extracted from 6 color features, RGB and HIS color space elements, ellipse square variance, perimeter area ratio, density of 4 geometric features and 7 Hu invariant moments, the 17 color and shape features are sufficient to characterize the target fruit according to.4 the feature vector classifier design, extracted from RBF, Elman two neural network (Neural network NN) algorithm in object recognition, in order to overcome the inherent defects of their existence, using genetic algorithm (Genetic algorithm GA) is optimized to adopt a new optimization method, namely the weights and structure of neural network at the same time the evolution of GA-RBF-LMS and GA-Elman, were established two optimized neural network classification algorithm. By UCI simulation results and experimental data shows that the apple image recognition, operation efficiency and recognition accuracy of two recognition model has been greatly improved, the generalization The ability also increased; the overall target fruit rate is above 95%. The performance comparison of two kinds of recognition algorithm, GA-RBF-LMS algorithm has the advantage of high efficiency, and the recognition accuracy is slightly lower; the GA-Elman algorithm is to the advantage of high accuracy and efficiency, a little lack of.5 for large sample problems of apple the identification of the modeling process, first, using partial least squares (Partial least, squares, PLS) dimension reduction algorithm to ease the sample high dimension problem, low dimensional data get more meaningful. Then using spectral clustering (Hierarchical cluster analysis, HCA) horizontal dimension, similar to the sample will be classified as a type of refinement, for each sample. Are neural network modeling, a PLS-HCA-NN classification algorithm. By UCI test data shows that two kinds of new classification algorithm on its recognition accuracy, efficiency and generalization The force has been greatly improved. The research results laid a theoretical foundation for improving apple picking robot vision system and improving picking efficiency, and the intelligent optimization algorithm is also worthy of further application in other fields.

【学位授予单位】：江苏大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41;TP242

【相似文献】