基于YOLO v2的莲蓬快速识别研究

Research on Lotus Quick Recognition Based on YOLO v2

受莲蓬形状外观和生长环境影响,传统计算机视觉算法识别莲蓬存在效率与精度不佳的问题。本文研究采用YOLO v2算法进行莲蓬识别的方式,通过扩充莲蓬检测数据集、K-means维度聚类、深度可分离卷积网络结构和多分辨率图像对模型微调等方法实现提高识别精度、鲁棒性与识别速度。对比Darknet-19、Tiny Darknet与DS Tiny Darknet算法,结果表明,本文研究的识别方式可以达到102.1 fps的识别速率,可实现在复杂环境下对莲蓬的快速识别,满足莲蓬采摘机器人在采摘过程中对实时视觉信息的需求。

Affected by the shape and growth environment of lotus,the traditional computer vision algorithm has the problems of poor efficiency and precision.In this paper,the scheme of using YOLO v2 algorithm to recognize the lotus was proposed.Through expanding lotus detection data set,K-means dimension clustering,depthwise separable convolution network,multi-scale classified network fine-tuning and other methods to improve the recognition accuracy,robustness and recognition speed.Contrasting the actual performance of Darknet-19,Tiny Darknet and DS Tiny Darknet with the YOLO v2 algorithm,the results showed that the scheme could achieve the recognition rate of 102.1 fps,realize the quick recognition of the lotus in a complex environment,so as to meet the realtime vision demand for picking robot in picking process.