基于Mask R-CNN的复杂环境下辣椒识别方法研究

Research on pepper recognition method in complex environment based on Mask R-CNN

针对辣椒采摘机器人在真实场景中辣椒簇状、粘连和光照不均导致无法精准采摘辣椒的问题,提出一种基于Mask R-CNN实例分割网络模型的辣椒识别方法。以真实场景下的辣椒为研究对象,采集自然生长的辣椒图像4496张,对其中的4000张进行数据标注作为数据集,通过设置不同的学习率、训练周期和模型网络层对数据集进行训练。试验结果表明,Mask R-CNN网络模型对真实场景下辣椒的识别和分割效果较好,平均准确率达到90.34%,平均速度达到0.82 s/幅,为智能辣椒采摘机器人的辣椒分割识别和定位提供有力的技术支撑。

In order to solve the problem that pepper picking robots can not pick pepper accurately in real scenes due to pepper clusters,adhesion and uneven lighting,a pepper recognition method based on Mask R-CNN instance segmentation network model is proposed.With pepper in the real scene as the research object,4496 images of naturally growing pepper were collected,and 4000 of them were labeled as data sets.The data sets were trained by setting different learning rates,training cycles and model network layers.The experimental results show that the Mask R-CNN network model has a good effect on pepper recognition and segmentation in the real scene,with an average accuracy of 90.34%and an average speed of 0.82 s/frame,providing a strong technical support for pepper segmentation recognition and location of intelligent pepper picking robot.