In view of such flaws as the high miss detection rate of small targets including constructors and the requirement of real-time monitoring in the process of visual safety helmet wearing detection by deep learning method,an improved target detection model has thus been proposed in this paper. First of all,a residual network module is added to the original network of the algorithm,so that the characteristics of small targets will not cause the gradient to disappear as the network deepens,thus better improving the problem of high missed detection rates for small targets. Moreover,an optimization can be achieved of the loss function and the screening prediction frame. Theoretical analysis and results show that,compared with the original algorithm,the recognition accuracy rate of the proposed method can be improved by 4.6%,meanwhile the recall rate is increased by 3.9%,and the average accuracy rate is increased by 4.1%,with the frame rate being 63 frames/s. The results show that the improved algorithm proposed in this paper is characterized with a better performance in extracting small target features,with the position error of the bounding box reduced accordingly as well.