【作者】 于杨；

【导师】 董桂菊；

【作者基本信息】 东北农业大学 ， 农业电气化与自动化， 2010， 硕士

【摘要】 本文的研究目的是应用计算机对田间青椒进行自动识别,为将来实现田间青椒自动采摘机器人打下基础。本文的主要研究内容如下:1.提出了颜色因子与直方图阈值相结合对田间青椒与叶子(枝干)进行分割的方法。根据田间拍摄的青椒图像中的青椒、叶子(枝干)等颜色特征,进行了仔细地统计和分析,提出了颜色因子与直方图阈值相结合对田间青椒与叶子(枝干)进行分割的方法。2.为了准确提取青椒图像的形状特征,本文提出了对图像进行去噪、二值形态学、边缘检测等一系列的前期处理过程。由于图像在输入和处理过程中会产生干扰,形成噪声,降低图像质量,本文采用中值滤波对图像进行了去噪处理;对图像实行开启和闭合运算进行误判区域处理;通过Canny算子对图像进行了边缘检测处理,对目标图像的轮廓进行特征提取。3.计算并提取了青椒的五个形状特征,作为BP神经网络的输入。如果直接把预处理后的青椒图像作为神经网络的输入,会导致数据量大,识别速率比较慢,因此,本文提取了青椒的形状系数、伸长度、紧密度、长宽比、主轴周长比等五个归一化形状特征,作为神经网络的输入。4.构造BP神经网络图像识别系统。本文在实验的基础上确定了改进的BP神经网络模型及其对应参数,包括神经网络层数、各层结点数、激活函数、学习速率、动量因子等,并通过稳定的BP神经网络实现样本的识别测试,最终对识别结果进行统计、分析。5.本文利用最小二乘法进行圆的拟合以提取青椒的质心,从而实现青椒的准确定位。更多还原

【Abstract】 The purpose of this research is to apply computer to automatically identify green peppers in the fields, and for realization of research about the fields picking green peppers robot in the future to lay the foundation. The main research content of this article as follows:1. Proposed segmentation method for the field peppers and leaves (branches) to using of the combination of color factors and histogram threshold.Through carefully statistics and analysis to color features of green peppers, leaves (branches) and so on among the peppers images taken in the field, proposed segmentation method about the combination color factor and histogram threshold of the peppers and leaves (branches) in the field.2. In order to accurately extract the shape characteristics of peppers image, in the paper, image pre-processing, such as image de-noising, binary morphology, edge detection and so on, are proposed. In the paper, use of median filtering to remove the image noise, because they will interfere to form a noise, reduce the image quality, when the image input or processing; The use of open and closed operations to process misjudge the regional in the image; Use Canny operator to process the image edge detection and to extraction feature of the target image contours.3. As a BP neural networks input, calculated and extracted green pepper characteristics of the five shapes, If the pretreatment peppers image directly as a neural network input then that will increase the amount of data, and slow identification rate. Therefore, extracted five normalized shape feature of pepper as a neural network input, including the form factor, stretch length, close degrees, aspect ratio, spindle circumference ratio.4. Construction BP neural network image recognition system. Through experiments, this paper identified the improved BP neural network model and the corresponding parameters. Including neural network layers, layers of nodes, activation function, learning rate, momentum factor and so on, and use a stable BP neural network to realize sample identification test. Finally, statistic and analysis recognition results.5. In this paper, using Least-squares method to fit the circle for extract the pepper centroid, and accurately locate pepper.更多还原