TY - JOUR
T1 - Faster and accurate green pepper detection using NSGA-II-based pruned YOLOv5l in the field environment
AU - Nan, Yulong
AU - Zhang, Huichun
AU - Zeng, Yong
AU - Zheng, Jiaqiang
AU - Ge, Yufeng
N1 - Funding Information:
This work is supported by Funding for school-level research projects of Yancheng Institute of Technology (Grant No. xjr2021012), National Natural Science Foundation of China (Grant No. 32171790), Key Research and Development Program of Jiangsu Province (BE2021307), Jiangsu Province Modern Agricultural Machinery Equipment and Technology Demonstration Promotion Project (NJ2020-18). The authors would like to thank the Alstonia pepper planting base was located in Yanyang Village, Tinghu District, Yancheng City for providing permission and a venue for taking photos in the Alstonia pepper field.
Publisher Copyright:
© 2022 Elsevier B.V.
PY - 2023/2
Y1 - 2023/2
N2 - Rapid and accurate detection of green peppers are essential for their growth monitoring, yield estimation, phenotypic monitoring, and robotic harvesting. In order to simplify the detection model and improve the detection efficiency, the NSGA-II-based (Non-dominated Sorting Genetic Algorithm-II-based) pruning algorithm was proposed to obtain an optimal pruning that balanced the detection accuracy and speed of the pruned model. A green pepper detection model was trained using YOLOv5l, and the NSGA-II-based pruning algorithm was implemented to obtain a YOLOv5l pepper detection model. The number of model parameters, model size, and GFlops of the pruned model were reduced by 73.9 %, 73.5 %, and 62.7 %, respectively. The mAP0.5 of the pruned model was 81.4 %, only slightly lower (by 0.973 %) than that of the original model.The detection speed of the pruned model was 70.9f/s, which was 59.0 % higher than that of the model before pruning. The NSGA-II-based pruning also significantly outperformed other two algorithms, namely, Slim pruning and EagleEye pruning, in terms of number of parameters, model size, GFlops, and detection speed, with a slight reduction in mAP0.5 0.973 % compared to EagleEye pruning. Finally, the NSGA-II-based pruned YOLOv5l pepper detection model was compared with other 11 deep learning models. Except that the mAP0.5 was only 0.367 % lower than that of YOLOv4, our method again showed had obvious advantages in terms of parameter quantity, model size, GFlops, mAP0.5, and detection speed. This research provided a new method and insights for the pruning of deep learning models, which is a necessary step to deploy them in compact mobile devices for real-time applications.
AB - Rapid and accurate detection of green peppers are essential for their growth monitoring, yield estimation, phenotypic monitoring, and robotic harvesting. In order to simplify the detection model and improve the detection efficiency, the NSGA-II-based (Non-dominated Sorting Genetic Algorithm-II-based) pruning algorithm was proposed to obtain an optimal pruning that balanced the detection accuracy and speed of the pruned model. A green pepper detection model was trained using YOLOv5l, and the NSGA-II-based pruning algorithm was implemented to obtain a YOLOv5l pepper detection model. The number of model parameters, model size, and GFlops of the pruned model were reduced by 73.9 %, 73.5 %, and 62.7 %, respectively. The mAP0.5 of the pruned model was 81.4 %, only slightly lower (by 0.973 %) than that of the original model.The detection speed of the pruned model was 70.9f/s, which was 59.0 % higher than that of the model before pruning. The NSGA-II-based pruning also significantly outperformed other two algorithms, namely, Slim pruning and EagleEye pruning, in terms of number of parameters, model size, GFlops, and detection speed, with a slight reduction in mAP0.5 0.973 % compared to EagleEye pruning. Finally, the NSGA-II-based pruned YOLOv5l pepper detection model was compared with other 11 deep learning models. Except that the mAP0.5 was only 0.367 % lower than that of YOLOv4, our method again showed had obvious advantages in terms of parameter quantity, model size, GFlops, mAP0.5, and detection speed. This research provided a new method and insights for the pruning of deep learning models, which is a necessary step to deploy them in compact mobile devices for real-time applications.
KW - Detection
KW - Field environment
KW - Green pepper
KW - NSGA-II-based pruning algorithm
KW - YOLOv5l
UR - http://www.scopus.com/inward/record.url?scp=85145250345&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85145250345&partnerID=8YFLogxK
U2 - 10.1016/j.compag.2022.107563
DO - 10.1016/j.compag.2022.107563
M3 - Article
AN - SCOPUS:85145250345
SN - 0168-1699
VL - 205
JO - Computers and Electronics in Agriculture
JF - Computers and Electronics in Agriculture
M1 - 107563
ER -