Traffic Surveillance License Plate Recognition System in Heavy Rainfall

In adverse weather conditions such as heavy rainfall, visibility obstruction significantly reduces the image quality captured by surveillance cameras, impacting the recognition of vehicles and their license plates. This leads to ineffective monitoring of traffic violations and other situations. To address this issue, we have developed a vehicle and license plate detection and recognition system specifically designed for heavy rain conditions, which consists of three main components: (a) image de- raining, (b) license plate localization, and (c) license plate information extraction. For image de-raining, we have introduced the NeRD-Rain algorithm and created a de-raining dataset containing vehicles by adding random noise. We then utilized pre- trained models for transfer learning. In the license plate localization stage, we employed the YOLOv5 model for accurate localization of license plates. Lastly, for license plate information extraction, we constructed a character extraction network. Through experiments conducted on our dataset, the complete version and lightweight version of the de-raining module achieved PSNR scores of 38.42 and 36.82, respectively. Moreover, the accuracy of license plate information extraction reached 96.33%.