Micro Laboratory Safety Hazard Detection Based on YOLOv4: A Lightweight Image Analysis Approach

In hazardous chemical laboratories, identifying and managing safety hazards is critical for effective safety management. This study, grounded in safety engineering principles, focuses on laboratory environments to develop an efficient hazard detection model using deep learning and object detection techniques. The lightweight YOLOv4-Tiny algorithm, with fewer parameters, was selected and optimized for detecting unsafe factors in laboratories. The CIOU loss function was employed to enhance the stability of candidate box regression, while three attention mechanism modules were embedded into the backbone feature extraction network and the feature pyramid's upsampling layer, forming an improved YOLOv4-Tiny object detection algorithm. To support the detection tasks, a specialized dataset for laboratory hazards was created. The improved YOLOv4-Tiny model was then used to construct two detection models: one for identifying the status of chemical bottles and another for detecting general laboratory safety hazards. The chemical bottle status detection model achieved AP values of 93.06% (normal), 95.31% (disorderly stacking), and 90.72% (label detachment), with an mAP of 93.03% and an FPS of 272, demonstrating both high accuracy and speed. The laboratory hazard detection model achieved AP values of 97.40%, 90.14%, 96.80%, and 68.95% for normal experimenters, individuals not wearing protective equipment, individuals smoking, and open flames, respectively, with a mAP of 88.32% and an FPS of 116. These results confirm the effectiveness of the proposed models in accurately and efficiently identifying laboratory safety hazards.