Abstract: Existing illegal dress recognition algorithms mainly locate and extract the image features of helmets and work clothes based on the proportion of human body structure, and cannot accurately recognize large postures such as bending, squatting, and climbing. Aiming at the above problems, an illegal dress recognition algorithm based on key-point detection and attention mechanism was proposed, which could be applied to a variety of human postures and different shooting angles, with good performance. Based on pedestrian detection, a single-person key-point detection model was designed, and a region localization strategy was proposed to accurately locate the pedestrian's head, upper body and bottom body areas, which reduced background interference and made dress feature extraction easier. In addition, the attention mechanism was introduced into the basic image classification model to further improve the accuracy of the identification of illegal dress. Experimental results based on human body posture images in substation work scenes verify the effectiveness of the algorithm.