基于CNN-Transformer的视觉缺陷柑橘分选方法

A CNN-Transformer-based method for sorting citrus with visual defects

针对产线分拣缺陷柑橘费时费力等问题,以柑橘加工生产线输送机上随机旋转的柑橘果实为研究对象,开发了一种基于卷积神经网络(CNN)的检测算法Mobile-citrus,用于检测和暂时分类缺陷果实,并采用Tracker-citrus跟踪算法来记录其路径上的分类信息,通过跟踪的历史信息识别柑橘的真实类别。结果显示,跟踪精度达到98.4%,分类精度达到92.8%。同时还应用基于Transformer的轨迹预测算法对果实的未来路径进行了预测,平均轨迹预测误差达到最低2.98个像素,可用于指导机器人手臂分选缺陷柑橘。试验结果表明,所提出的基于CNN-Transformer的缺陷柑橘视觉分选系统,可直接应用在柑橘加工生产线上实现快速在线分选。

Manual sorting of citrus fruit with visual defects on the production line is time-consuming and cost-expensive.This article proposes a sorting solution based on machine vision and CNN-Transformer.The system can be directly implemented on various citrus processing lines for online sorting.For the citrus fruits randomly rotating on the conveyor,a detection algorithm Mobile-citrus based on convolutional neural network(CNN)was developed to detect and temporarily classify the defective one.A tracking algorithm Tracker-citrus was used to record the classification information along the path.The real category of the fruit was identified using the historical information,with tracking accuracy of 98.4%and classification accuracy of 92.8%.In addition,a trajectory prediction algorithm based on Transformer was used to predict the future path of fruit with the average prediction error of 2.98 pixels,which can be used to guide the robot arm to sort defective citrus in real time.The results showed that the method proposed can be applied to citrus production lines for online sorting.