Deep learning based detection of goggle wearing in industrial plants

In industrial environments, the monitoring of safety goggle usage among onsite personnel primarily relies on manual supervision. However, this manual approach is inefficient, lacks precision, and cannot meet real-time requirements. To address this issue, this study proposes an improved SSD target detection model to automatically recognize the wearing conditions of safety goggles. The research focuses on two main aspects: 1) Constructing a dataset named GWD, containing images of four common types of industrial safety goggles. The dataset encompasses five categories: improperly worn goggles, correctly worn anti-radiation goggles, correctly worn dust-proof goggles, correctly worn impact-resistant goggles, and correctly worn chemical-resistant goggles; 2) Designing a comprehensive lightweight SSD model named BiFPN-FE-SSD. This model utilizes an improved MobileNetV2 as its backbone network and integrates weighted bidirectional feature pyramids (BiFPN) for multiscale fusion of high and low-level features. Additionally, it incorporates Feature Enhancement Modules (FEM) and Efficient Channel Attention (ECANet) mechanisms at strategic positions to enhance the sensitivity in detecting small and medium-sized objects.Experimental results demonstrate that the proposed model achieves higher accuracy and speed compared to common SSD optimized models such as DSOD, RSSD, DSSD, and FSSD, both on the GWD dataset and in real-world scenarios.