In the construction industry, safety accidents caused by not wearing helmets account for a relatively large proportion. Aiming at the problems of strong interference and low accuracy of small targets in helmet detection, an improved algorithm based on YOLOX is proposed. Firstly, an ECA-Net attention mechanism is added to the enhanced feature extraction network to carry out cross-channel interaction, suppress the interference information according to the corresponding channel weight value generated, strengthen the model's attention to the target feature, and then fuse the recalibrated feature map more deeply to improve the expression ability of the target feature. Secondly, the CIoU is used to calculate the loss, the distance between the two boxes of center points and the aspect ratio are considered into the penalty term, and the loss function is constantly adjusted and updated to accelerate the model convergence speed. Finally, a small target helmet dataset in a real construction scenario is constructed. Experimental results show that the improved algorithm mAP reaches 91.7%, which is 1.2% higher than the original YOLOX calculation, the average accuracy of the detection of workers who have worn helmets reaches 93.9%, the average accuracy of detection of those who have not worn helmets reaches 89.5%, and the detection speed reaches 71.9 frames/s, which ensures that the real-time detection of helmet wearing has a high accuracy rate.