Aiming at the problems of difficult license plate location,slow detec tion speed and low accuracy in complex scenes such as illumination,dense vehicles and low reso lution,a fast license plate detection method based on attention mechanism is proposed.Fi rstly,based on the characteristics of license plate,a lightweight network unit LeanNet is d esigned and used to build a backbone network with low computational complexity and accuracy. Secondly,MLA(multi scale light attention) module is designed to guide the netw ork to pay attention to license plates of different scales,generate local saliency images based on license plates,and suppress background noise.Finally,a four scale prediction network is designed,in which the FSPF(four scale pyramid fusion) module can generate four scale feature pyramid, which is conducive to the detection of license plates of different scales.The e xperimental results show that the accuracy of the proposed method is 99.12% in CCPD(Chinese city parking dataset).Compared with the latest you only look once v4(YOLOv4) detection method,the accuracy is improved by 1.9%,and the running speed is increased by 6times.The proposed me thod can realize license plate detection in complex scenes in embedded devices.