Safety helmet is one of the most widely used and important head safety protection equipment for laborers, widely used in machinery, construction, mines, transportation, metallurgy, electric power and other industries. However, due to the lack of safety awareness, some laborers often don't wear safety helmets correctly. Therefore, how to correctly identify operators who are not wearing safety helmets becomes very necessary. The existing helmet wearing recognition method has low recognition accuracy in complex posture and complex background, and it is easy to misjudge wearing other hats as wearing helmets. In this paper, an improved helmet recognition method is proposed to try to solve the above problems. First, for each image of human body, the joint information of the human body is obtained by human pose estimation, and then the subimages of his head-neck are cropped from the aforementioned human body image using an improved three-point localization method, and finally, a classification network is designed to classify all the obtained head-neck subimages to determine whether the laborer wears a helmet or not. The experimental results show that the accuracy of the helmet recognition method proposed in this paper can reach 99.0%, which is significantly higher than the comparison methods, and has strong robustness in complex environments.