Small Target Detection Method for Safety Helmet Based on Lightweight Transformer

This paper addresses the critical challenge of detecting small-scale safety helmets in complex industrial surveillance scenarios. Traditional methods face difficulties due to low pixel occupancy (often <0.05%), frequent occlusion, and computational constraints. To tackle these issues, we propose LW-DETR, a lightweight Transformer-based architecture. It integrates three innovative modules: the Large Kernel Selection Network (LSKNet) dynamically selects large convolutional kernels to capture local contextual details; the Frequency-Spatial Fusion Transformer (FSF Transformer) models long-range dependencies by fusing spectral and spatial features; and the Boundary-aware Selection Aggregation (SBA) module enhances edge-sensitive feature representation for precise localization. Extensive experiments demonstrate that LW-DETR achieves a 1.4% precision gain compared to the baseline, while reducing parameters by 6.4 million. These results establish LW-DETR as an effective solution for practical safety monitoring systems, balancing detection precision with computational efficiency.