License plate recognition in low quality image by using Latent Diffusion YOLOv7

The biggest challenge focus by today’s license plate recognition systems is whether they can still operate normally when weather conditions are severe. Many current systems rely on deep learning methods to train the system, however, collecting data under severe weather often requires a lot of costs. Because the quality of collected data is generally poor, which will greatly increase the difficulty of manual annotation. In addition, since regional license plates vary greatly, it will take a lot of effort to collect plate data in different regional styles in order to develop a universally-applicable recognition system.This paper mainly explores the problem of rapid performance degradation of existing license plate recognition systems under harsh weather conditions. We propose a method that can simultaneously generate fully synthetic license plate data as well as noises under severe weather conditions. Furthermore, we designed a new Latent Diffusion YOLOv7 (LD-YOLOv7) neural network based on the existing object detection method—YOLOv7, which can effectively solve various problems of plate recognition in bad weather. In order to verify the effectiveness of this system in practical application, we also established a regional license plate dataset containing real harsh climate conditions for testing purposes. The experimental results show that our proposed model performs comparable to other existing methods trained with real data on AOLP dataset, and performs better under harsh weather conditions.