自然场景下猕猴桃识别方法研究

发布时间：2018-07-04 10:11

本文选题：猕猴桃 + 分割　；参考：《西北农林科技大学》2017年硕士论文

【摘要】：中国是猕猴桃生产大国,单纯依靠人力来进行猕猴桃采摘费工费时,随着科学技术的不断发展,自动化采摘技术已经慢慢渗入到人们的生产、生活中。准确识别自然场景下猕猴桃成为采摘机器人的首要任务,因此研究猕猴桃果实识别方法变得十分必要。本文以自然场景下猕猴桃作为研究对象,首先对采集图像进行分割处理,然后对于人工设计特征和深度学习特征进行特征提取,最后将提取的特征向量作为参数输入,采用神经网络和支持向量机识别模型,实现对猕猴桃果实识别,本文的主要内容如下:(1)对采集的猕猴桃图像进行分析,采用基于色差法的Otsu阈值法和基于Renyi熵阈值分割法两种方法实现果实与背景的分割。对比并分析实验结果,本文最终选择0.9R-G,G-0.9R,G-R,R-G色差分量对猕猴桃图像中的果实、树叶、草地、天空和树干目标物进行阈值分割处理,针对部分背景复杂,且果实众多的图像,本文采用基于Renyi熵的阈值分割法。针对分割处理后的二值图像存在的残留物问题,利用面积阈值法,形态学处理算法进行残留物去除,为特征提取打好基础。(2)根据分割后的二值图像绘制目标物边缘轮廓,并利用最小外接矩形算法提取样本。针对猕猴桃自身颜色特征及特殊的表皮毛刺构成的纹理特征和背景样本(包括树叶、天空、草地及枝干)的纹理、颜色特征的分析,本实验拟提取包括HSV颜色特征、灰度共生矩特征、Tamura特征中的对比度、线性度和粗糙度共10个人为设计特征和基于PCANet的深度学习特征,为基于神经网络和支持向量机模型的果实识别做好准备。(3)根据提取的果实、树叶、天空、草地及枝干特征,建立基于神经网络和支持向量机的果实识别模型。分别对BP神经网络、PCANet深度学习模型和支持向量机(SVM)算法进行建模实验。实验结果表明,使用PCANet网络的识别效果最好,识别率达到94.92%,而支持向量机和BP神经网络的识别率分别为87.67%和65.09%。综上所述,通过对猕猴桃图像特征及外在环境分析,本文提出了人工设计特征和深度学习特征表示样本特征,利用神经网络和支持向量机模型进行猕猴桃果实识别,取得了较好的结果,实现了自然场景下猕猴桃果实的准确识别。
[Abstract]:China is a large country of kiwifruit production, relying solely on manpower to harvest kiwifruit takes time and time. With the development of science and technology, automatic harvesting technology has gradually infiltrated into people's production and life. Accurate recognition of kiwifruit in natural scene is the first task of picking robot, so it is necessary to study the recognition method of kiwifruit. In this paper, the kiwifruit is used as the research object in the natural scene. Firstly, the image is segmented, then the artificial design feature and the depth learning feature are extracted. Finally, the extracted feature vector is used as the parameter input. Using neural network and support vector machine to recognize kiwifruit, the main contents of this paper are as follows: (1) the image of kiwifruit is analyzed. Otsu threshold method based on color difference method and Renyi entropy threshold segmentation method were used to realize fruit and background segmentation. By comparing and analyzing the experimental results, we select the 0.9R-GG -0.9RG G-RGG color difference component to segment the fruits, leaves, grasslands, sky and tree trunks of kiwifruit images by threshold segmentation, aiming at some images with complex background and numerous fruits. The threshold segmentation method based on Renyi entropy is used in this paper. In view of the residue problem of binary image after segmentation, the area threshold method and morphological processing algorithm are used to remove the residue. (2) the edge contour of object is drawn according to the segmented binary image. The minimum external rectangle algorithm is used to extract samples. Based on the analysis of the color features of kiwifruit and the texture features of special epidermis burr and background samples (including leaves, sky, grass and branches), the color features of kiwifruit were extracted, including HSV color features. The contrast, linearity and roughness in the Tamura feature of the gray level symbiotic moment feature are 10 individuals for the design feature and the PCANet based depth learning feature, which are ready for the fruit recognition based on neural network and support vector machine model. (3) according to the extracted fruit, The fruit recognition model based on neural network and support vector machine (SVM) was established based on the characteristics of leaves, sky, grassland and branches. The deep learning model and support vector machine (SVM) algorithm of BP neural network are simulated. The experimental results show that the recognition rate of PCANet is the best, the recognition rate is 94.922.The recognition rate of support vector machine and BP neural network are 87.67% and 65.09% respectively. To sum up, by analyzing the image features and external environment of kiwifruit, this paper proposes artificial design features and depth learning features to represent sample features, and uses neural network and support vector machine model to recognize kiwifruit. Good results were obtained, and the recognition of kiwifruit in natural scene was realized.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S663.4;TP391.41

【相似文献】