Authors

Abstract

In response to the inefficiency and low accuracy issues of traditional detection algorithms in detecting the tray seeding process of green onions, this paper proposes a machine vision-based quality inspection system for green onion seeding. Considering the color characteristics of green onion seeds and the substrate soil, the original RGB images are converted into HSV images. The HSV color filtering algorithm is employed to separate green onion seeds from complex soil backgrounds. Image noise is removed using erosion-dilation operations and small-area methods. The projection method is utilized to determine the detection area of the tray and the position of the holes. Information about connected regions and their convex hulls is extracted, and eight feature parameters including perimeter, area, shape factor, perimeter ratio, area ratio, shape factor ratio, concave defect distance ratio, and error variance are used to establish a BP neural network for single and adherent seed classification. A concave point segmentation algorithm is used to separate adherent green onion seeds and count the number of green onion seeds in each hole to obtain seeding quality parameters of the seeder. Experimental results show that the average relative error of the system qualification rate is 2.24%, with maximum and minimum relative errors of 3.22% and 1.10%, respectively. The average absolute errors of the reseeding rate and void rate are 1.31% and 0.71%, respectively. The absolute error of the average number of particles is 0.025 particles, and the average processing time per image is 0.91 s. The research results provide reference data for precision seeding operations of green onion seeders.

Abstract in Chinese

针对传统的检测算法在检测大葱秧盘播种过程中，效率低下，检测精度低等问题，本文提出了一种基于机器视觉的大葱播种质量检测系统。针对大葱种子颜色和基质土壤的颜色特征，将原始RGB图像转换成HSV图像。通过HSV色彩过滤算法将大葱种子从复杂的土壤背景中分离。使用腐蚀膨胀操作以及小面积法去除图像噪声。使用投影法确定秧盘检测区域和穴孔位置。提取连通区域及其凸包的信息，使用了周长、面积、形状因子、周长比、面积比、形状因子比、凹缺陷距离比、误差方差8个特征参数，建立BP神经网络单粒种子与粘连种子分类模型。使用凹点分割算法将粘连的大葱种子分离，并统计每穴大葱种子数量，得到排种器的播种质量参数。试验结果表明，系统合格率的平均相对误差为2.24%，最大和最小相对误差分别为3.22%和1.10%；重播率和空穴率的平均绝对误差分别为1.31%和0.71%；平均粒数的绝对误差为0.025粒、每幅图像平均处理时间为0.91 s。研究结果为大葱排种器精密播种作业提供了参考数据。