基于卷积神经网络和机器视觉的辣椒检测与识别

发布时间：2018-04-05 20:07

本文选题：辣椒　切入点：卷积神经网络　出处：《天津理工大学》2017年硕士论文

【摘要】：农业采摘机器人一直以来都在机器人研究领域占有重要地位。随着社会老龄化的加剧,使得人力资源缺乏,研发农业采摘机器人来代替人类在田地里劳作迫在眉睫。在国外,针对农业采摘机器人的研究已经取得一定成果,已步入商业化发展阶段。而在国内,相关研究起步较晚,特别对于辣椒的智能研究更是缺乏,仅仅只涉及根据辣椒的形状特征进行品级分类,还没有针对成熟辣椒采摘问题的研究。本文从实际情况出发,针对自然环境下的成熟辣椒的识别与定位进行研究。因为辣椒在生长过程中受到外界自然条件的影响使得采集到的图像很复杂,所以要先对图像进行一定的预处理。根据成熟辣椒的特征量较少和复杂的外界环境的特点提出了改进的卷积神经网络算法对辣椒进行识别。然后本文设计了基于机器视觉的辣椒定位系统,对成熟辣椒进行立体空间定位。具体研究内容如下:首先是在自然环境下对室外田地辣椒进行图像采集,对采集到的辣椒图像经过图像灰度化、图像增强、二值化和形态学处理等一定的预处理,得到了含有目标特征的二值图像。为了满足批量预处理和后期的识别研究,开发了基于MATLAB的GUI界面。然后根据辣椒的自身特点和外界复杂的自然环境,选取了有深度学习机制的卷积神经网络算法对成熟辣椒进行识别。为了提高识别率对神经网络结构进行了改进进而建立识别系统,并与传统BP神经网络进行比较。表明此识别系统具有良好的鲁棒性、容错性、泛化性和较高的识别率。最后是对成熟辣椒的立体定位研究,先阐释了双目立体测距的几何原理,并对相机标定、立体匹配、空间定位做了深入分析。本文设计了基于双目相机平行配置结构的辣椒定位系统,最终得到对目标物的立体定位的实验数据,并对其进行了定性分析。得出双目视觉系统能够相对准确的对辣椒的立体空间位置进行定位。
[Abstract]:Agricultural picking robot has been playing an important role in the field of robot research.With the aggravation of the aging society, human resources are scarce, so it is urgent to develop agricultural harvesting robot instead of human working in the field.In foreign countries, some achievements have been made in the research of agricultural harvesting robots, and they have entered the stage of commercialization.In China, the related research started relatively late, especially for the intelligence of pepper is lack, only involved in the classification according to the shape characteristics of pepper, there is no research on the problem of picking mature pepper.In this paper, based on the actual situation, the identification and location of mature pepper under natural environment were studied.Because the chili pepper is affected by the natural conditions in the growth process, the collected image is very complicated, so the image must be preprocessed first.An improved convolution neural network algorithm is proposed to recognize the mature pepper according to the characteristics of less characteristic quantity and complex external environment.Then, this paper designs a pepper positioning system based on machine vision to locate the mature pepper in three-dimensional space.The specific research contents are as follows: first of all, the outdoor field pepper image acquisition in the natural environment, the collected pepper image grayscale, image enhancement, binarization and morphology processing and other certain preprocessing,A binary image with target features is obtained.A GUI interface based on MATLAB is developed to satisfy batch preprocessing and later recognition research.Then, according to the characteristics of pepper and the complex natural environment, the convolution neural network algorithm with deep learning mechanism is selected to recognize the mature pepper.In order to improve the recognition rate, the neural network structure is improved and the recognition system is established, and compared with the traditional BP neural network.It is shown that this recognition system has good robustness, fault tolerance, generalization and high recognition rate.In the end, the geometric principle of binocular stereo ranging is explained, and the camera calibration, stereo matching and space location are deeply analyzed.In this paper, a pepper positioning system based on the parallel configuration of binocular camera is designed. Finally, the experimental data of the stereo positioning of the object are obtained and analyzed qualitatively.It is concluded that the binocular vision system can locate the spatial position of pepper relatively accurately.
【学位授予单位】：天津理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】