In the early days, visual sensors were highly applied to supervise the wearing of fire-resistant clothing to ensure safe production in calcium carbide factories. However, artificial intelligence is difficult to apply to these visual sensing systems with the rapid development of information technology. Considering this, an improved approach is established that allows for the operation of old visual sensing system without adding too much financial expenditure in this paper. Our primary contribution involves dual architectural enhancements targeting both the Efficient Channel Attention (ECA) mechanism and Pyramid Split Attention (PSA) module. Two attention modules are also integrated into bottleneck module of the YOLOv5 network, which is named as Dual Attention based on YOLOv5 (DA-YOLOv5). Public dataset and self-made dataset are utilized to validate the effectiveness of the proposed approach. At the same time, the impact of data augmentation and annotation quality on performance was studied to guide the application of the algorithm in practical engineering. The results show that the improvement of old visual sensing system is successful in calcium carbide factories.