Fusing few-shot learning and lightweight deep network method for detecting workpiece pose based on monocular vision systems

作者：Zhang, Tie^*; Zheng, Jingfu; Zou, Yanbiao

来源：MEASUREMENT, 2023, 218: 113118.

DOI：10.1016/j.measurement.2023.113118

• Pose measurement
• Deep learning
• Depthwise separable convolution
• Rotating object detection
• Two -stage fine-tuning strategy

With the popularization of workpiece grabbing and sorting applications, the measurement accuracy and speed of workpiece pose become the key issues. Therefore, a lightweight workpiece pose measurement algorithm based on the YOLO framework is proposed. In the study, the backbone network of YOLOv4-Tiny is lightened by the MobileNetV3 module. Then, the rotation anchor and loss function are introduced into the network to measure the pose. Additionally, to improve the adaptability of the algorithm to novel workpieces, a two-stage fine-tuning strategy is integrated into the pose measurement network to realize few-shot learning. Compared with the general object detection network, the proposed algorithm can achieve high speed and precision pose measurement and has good adaptability to novel workpieces. Experimental results show that the measurement speed of the proposed method is eight times faster than baseline Rotation-YOLOv4-Tiny, the position error is 0.42 x 10-3 m, and the angle error is 0.016 rad.