This paper proposes a license plate recognition algorithm based on LSTM theory and algorithm in complex environment, aiming at the problem that the recognition accuracy and response time are reduced due to natural light and rainy weather. In order to enhance the contrast of license plate area and reduce the difficulty of location, the license plate image is preprocessed with corrosion algorithm, gray scale and binarization. Secondly, image recognition processing algorithm is used for license plate location, character segmentation, character recognition and other operations. The characters obtained after the license plate segmentation are normalized and the appropriate size is unified as the input of the long and short duration memory network (LSTM), and the corresponding Chinese characters, numbers and alphabetic characters are as the output. The model of license plate recognition system is realized through the TRAINED LSTM network. After a lot of data collection and training, compared with the existing license plate recognition system, the algorithm proposed in this paper has a recognition accuracy of 98.90% Chinese characters, 99.40% alphanumeric characters and a single image recognition speed of 2.65ms.