DMS-DETR: Dynamic Multi-Scale Detection Transformer With Progressive Path Attention for PPE Detection

The widespread deployment of automated safety monitoring systems in industrial environments has highlighted the critical need for reliable personal protective equipment (PPE) detection. Existing computer vision approaches face significant challenges in real-world industrial settings due to the inherent difficulty in detecting small, often partially occluded PPE items under varying environmental conditions. To address these challenges, we present DMS-DETR (Dynamic Multi-Scale Detection Transformer), a framework for efficient PPE detection. The core of DMS-DETR is a backbone network enhanced with Dynamic Multi-Scale Deformable Convolution (DMDConv) for adaptive feature learning. We enhance the framework with a Dynamic Path Attention mechanism for attentive feature selection through spatial-channel attention, alongside a Kolmogorov-Arnold Network for nonlinear feature fusion. DMS-DETR achieves 66.6% F1 score and 64.0% [email protected] with only 95.9 GFLOPs, demonstrating its suitability for real-time industrial applications.