A rotatable battery recognition method based on improved YOLOv5
Wenming Chen,Dongtai Liang,Wenhui Ding,Meng Wang,Zizhen Chen
DOI: https://doi.org/10.1504/ijsnet.2024.138913
2024-06-04
International Journal of Sensor Networks
Abstract:To realise end-to-end visual identification, positioning, and angle detection of cylindrical batteries, a rotated object recognition method based on YOLOv5 is proposed. Firstly, aiming at the problems of battery appearance scale variation and surface reflection, a recursive gated convolution and feature fusion module was added to the neck network to enhance the multi-scale feature extraction. Secondly, considering the boundary problem of angle range, a circular smooth label was introduced after the prediction network, and the logistic regression cross-entropy was used to realise rotation angle classification. Finally, a SIoU intersection ratio model was used to introduce an angle vector penalty index. The experimental results show that the parameters of rotated object detection model are reasonably optimised on the cylindrical battery dataset. The model accuracy reaches 98.6%, the recall rate reaches 97.2%, and the inference speed of single frame image reaches 10.5 ms, which meets the performance requirements of practical applications.
computer science, information systems,telecommunications