Efficient helmet detection based on deep learning and pruning methods

Ensuring correct helmet usage is vital for mitigating head injuries in construction environments. Although YOLO-based detectors are proficient in dress code recognition, their performance is often constrained when detecting small objects. We introduce ADSP-YOLO, an advanced algorithm derived from YOLOv8n and specifically optimized for helmet detection. Central to this design is the adaptive boundary-semantic aggregation module, which replaces the traditional feature pyramid network and incorporates a specialized detection head tailored for small targets. To further enhance detection efficacy, we propose the scale sequence feature fusion, leveraging edge information from the P2 feature layer, and integrate the dynamic head with attention modules to refine accuracy. Model optimization is achieved through the layer-adaptive magnitude-based pruning method, enabling a balance between compression and performance. Evaluations on the safety helmet wearing dataset demonstrate that ADSP-YOLO achieves a [email protected] 92.7% on the test set, surpassing YOLOv8n by 2.2% and 4.2% in [email protected] and APs, respectively, while reducing model parameters to 39.3% and model size to 45.8% of the original. Moreover, ADSP-YOLO achieves a [email protected] of 82.3% on the GDUT-HWD dataset and demonstrates robust adaptability on the remote sensing object detection dataset, highlighting its potential for broader applications in small object detection.