基于实例分割的柚子姿态识别与定位研究

Research on pomelo pose recognition and location based on instance segmentation

采用Mask R-CNN和YOLOv3算法对复杂背景场景下的柚子进行目标识别和分割,通过微调方法训练了2个实例分割模型,并应用于柚子相关目标的识别和分割。结果表明,YOLOv3模型的帧速率为18~20 FPS,Mask R-CNN模型的帧速率为0.5~2 FPS,YOLOv3模型的目标检出率比Mask R-CNN模型少30%以上,包围框定位精度比Mask R-CNN模型偏离10%以上。基于Mask R-CNN模型输出的Mask提出的外形估算方法,免标定像素尺寸,时间复杂度为T(n),便于轮廓对比。基于目标边框对柚子角度进行计算,角度动态范围为±5°。说明基于实例分割的柚子外形估算和品质评价方法,能适应复杂图像背景,具有较强的泛化能力。

In this paper,two instance segmentation fine-tune models for pomelo image object were trained,using to Mask R-CNN and YOLOv3 respectively to recognize and segment the related target of pomelos.The results show that the frame rate of YOLOv3 model is 18 to 20 FPS,and that of Mask R-CNN model is 0.5 to 2 FPS.The target detection rate of YOLOv3 model is more than 30%lower than that of Mask R-CNN model,and the positioning accuracy of bounding box is more than 10%lower than that of Mask R-CNN model.Based on the Mask R-CNN model,the shape estimation method proposed by mask is output,which is free of calibration pixel size and time complexity is T(n),which is convenient for contour comparison.The dynamic range of pomelo angle was calculated to be±5°based on the target frame indicating that the pomelo shape estimation and quality evaluation methods based on instance segmentation can adapt to complex image background and has stronger generalization ability.