TY - GEN
T1 - SOLOv2-cable
T2 - 3rd International Conference on Computer Vision, Image and Deep Learning and International Conference on Computer Engineering and Applications, CVIDL and ICCEA 2022
AU - Luo, Xinhao
AU - Zeng, Wentao
AU - Chen, Yilin
AU - Liu, Shaohui
N1 - Publisher Copyright:
© 2022 IEEE.
PY - 2022
Y1 - 2022
N2 - In this paper, we propose SOLOv2-cable: a power cable segmentation algorithm based on neural network in complex scenarios, to solve the problems of miss predictions in the traditional algorithm in the process of power cable identification. The method is based on the SOLOv2 algorithm. First, the attention mechanism is introduced, and the ResNeSt network with the separation attention mechanism is used to design the backbone network to obtain multi-scale feature representation. Second, we optimize the loss function by adding the coordinate information of the instant center to the loss function, replace focal loss with GHM loss, and discuss the potential of GHM in imbalanced datasets; finally, data enhancement and cosine annealing step size are used to further improve the generalization ability of the model. Compared with the original SOLOv2 algorithm, the AP, AP50, and AP75 of SOLOv2-cable are improved by 6.4%, 7.1%, and 8.2%, and the reasoning time is only increased by 17%. This work can better identify power cables and provide technical support for subsequent laying of cables and measurement of cable curvature.
AB - In this paper, we propose SOLOv2-cable: a power cable segmentation algorithm based on neural network in complex scenarios, to solve the problems of miss predictions in the traditional algorithm in the process of power cable identification. The method is based on the SOLOv2 algorithm. First, the attention mechanism is introduced, and the ResNeSt network with the separation attention mechanism is used to design the backbone network to obtain multi-scale feature representation. Second, we optimize the loss function by adding the coordinate information of the instant center to the loss function, replace focal loss with GHM loss, and discuss the potential of GHM in imbalanced datasets; finally, data enhancement and cosine annealing step size are used to further improve the generalization ability of the model. Compared with the original SOLOv2 algorithm, the AP, AP50, and AP75 of SOLOv2-cable are improved by 6.4%, 7.1%, and 8.2%, and the reasoning time is only increased by 17%. This work can better identify power cables and provide technical support for subsequent laying of cables and measurement of cable curvature.
KW - GHM loss
KW - SOLOv2
KW - attention mechanism
KW - instance center distance
KW - power cable
UR - https://www.scopus.com/pages/publications/85135392066
U2 - 10.1109/CVIDLICCEA56201.2022.9825426
DO - 10.1109/CVIDLICCEA56201.2022.9825426
M3 - 会议稿件
AN - SCOPUS:85135392066
T3 - 2022 3rd International Conference on Computer Vision, Image and Deep Learning and International Conference on Computer Engineering and Applications, CVIDL and ICCEA 2022
SP - 1156
EP - 1161
BT - 2022 3rd International Conference on Computer Vision, Image and Deep Learning and International Conference on Computer Engineering and Applications, CVIDL and ICCEA 2022
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 20 May 2022 through 22 May 2022
ER -