Application of Improved YOLOv5 in Identification of Engineering Laboratory Devices

• Published in: International Conference on Automation, Control and Robotics Engineering (Online) pp. 1 - 8
• Main Authors: Liang, Devu, Liu, Zhibo, Cheng, Jian, Zhao, Tingqi
• Format: Conference Proceeding
• Language: English
• Published: IEEE 16.07.2025
• Subjects: Accuracy; Computational modeling; Computer Vision; Deep learning; Neck; Object recognition; Optimization; Performance evaluation; Real-time systems; Smart Laboratory; Training; YOLO; YOLOv5
• ISSN: 2997-6278
• DOI: 10.1109/CACRE66141.2025.11119602

In smart laboratory scenarios, autonomous mo-bile operating robots are often used to monitor and operate laboratory equipment. One of the important issues is to use deep learning based object detection algorithms to quickly and accurately identify and locate equipment and its operating components in laboratory scenes. In response to the universality of YOLOv5 model in object detection, we choose YOLOv5s as the basis for algorithm improvement and carried out the following work: to reduce network complexity and meet the real-time requirements of detection, we replace the Conv module of the network neck layer with the ethereal convolution module GSConv to construct the Slim-Neck module; Secondly, we introduce a new Decouple_Head that separates classification tasks from regression tasks to better parse network feature information; In terms of network loss function, the improved SIOU loss function takes the place of the original GIOU loss function. Based on the application scenario, this paper constructs a dedicated database(BED25) for model training and performance evaluation. Through data analysis of the iteration interval of the improved network, ours has a fast convergence speed and an average detection accuracy of 94.0%, which is 3.0% higher than the original model. The ablation experiment demonstrated that each introduced optimization scheme has an effect on improving the comportment of the model. Under the same training conditions on the dataset, our network in this paper has an advantage in detection behavior compared to other models. In summary, the upgraded YOLOv5 model, discussed in this paper, can accurately identify and locate devices in smart laboratory scenarios with high precision, while sacrificing a small amount of computing speed, meeting application requirements and real-time requirements.