TY  - EJOU
AU  - Chen, Xuefei 
AU  - Tan, Wenhui 
AU  - Wu, Qiulan 
AU  - Zhang, Feng 
AU  - Guo, Xiumei 
AU  - Zhu, Zixin 

TI  - Contamination Identification of Lentinula Edodes Logs Based on Improved YOLOv5s
T2  - Intelligent Automation \& Soft Computing

PY  - 2023
VL  - 37
IS  - 3
SN  - 2326-005X

AB  - In order to improve the accuracy and efficiency of Lentinula edodes logs contamination identification, an improved YOLOv5s contamination identification model for Lentinula edodes logs (YOLOv5s-CGGS) is proposed in this paper. Firstly, a CA (coordinate attention) mechanism is introduced in the feature extraction network of YOLOv5s to improve the identifiability of Lentinula edodes logs contamination and the accuracy of target localization. Then, the CIoU (Complete-IOU) loss function is replaced by an SIoU (SCYLLA-IoU) loss function to improve the model’s convergence speed and inference accuracy. Finally, the GSConv and GhostConv modules are used to improve and optimize the feature fusion network to improve identification efficiency. The method in this paper achieves values of 97.83%, 97.20%, and 98.20% in precision, recall, and mAP@0.5, which are 2.33%, 3.0%, and 1.5% better than YOLOv5s, respectively. mAP@0.5 is better than YOLOv4, Ghost-YOLOv4, and Mobilenetv3-YOLOv4 (improved by 4.61%, 5.16%, and 6.04%, respectively), and the FPS increased by two to three times.
KW  - Lentinula edodes logs; contamination identification; deep learning; attention mechanism; loss function

DO  - 10.32604/iasc.2023.040903