License plate recognition system for complex scenarios based on improved YOLOv5s and LPRNet

Traditional license plate recognition (LPR) algorithms perform well in controlled environments but often suffer from accuracy degradation in complex scenarios (such as adverse weather, plate tilt, and varying capture distances) as well as deployment difficulties under hardware constraints. This study proposes a lightweight, end-to-end method for license plate detection and recognition that integrates an improved YOLOv5s with LPRNet. First, we incorporate a Triplet Attention mechanism into the YOLOv5s backbone to enhance feature extraction, more precisely focus on license plate regions and suppress background interference from adverse weather. In the detection post-processing stage, we introduce a Soft-NMS strategy that applies Gaussian-weighted smoothing suppression to overlapping candidate boxes, thereby alleviating the over-suppression of overlapping license plates by traditional NMS and enhancing detection robustness. To address the issue of decreased recognition accuracy caused by license plate tilting, we introduce a Spatial Transformer Network (STN) before the recognition stage to geometrically correct tilted or distorted license plate images, thereby improving recognition accuracy. Experiments conducted on the CCPD2019 and CRPD datasets demonstrate that the proposed method achieves a detection precision of 98.9% and a recognition accuracy of 91.5% on CCPD2019, representing improvements of 3.7% and 8.38% over the baseline YOLOv5s + LPRNet, respectively. The model contains only 7.5 M parameters and 18.1 GFLOPs; it achieves 147 FPS for detection and 0.1138 ms inference time for recognition, indicating potential feasibility for deployment on resource-constrained platforms such as mobile devices and embedded systems.