SvRetina-LPD: A Sliding Vertex-Based RetinaNet for Robust Multi-Oriented License Plate Detection

The performance of license plate detection has greatly improved with the development of deep learning. However, two challenges remain. First, in unconstrained scenarios, such as rotation and uneven lighting, license plate detection still faces significant challenges. Secondly, traditional horizontal bounding boxes are not suitable for representing multi-oriented license plates. And quadrilateral boxes can effectively represent them though, they cannot avoid the confusion caused by sequential label points. To address these challenges, we propose a license plate detection method called SvRetina-LPD. Specifically, we design an inception residual fusion pyramid network, which fuses multiple layers of features thoroughly. Then, by integrating spatial attention and channel attention, the inception residual multi-dimensional attention network is developed to weaken the noise and highlight the features of the license plate. Finally, we introduce a sliding vertex head network that incorporates a sliding vertex branch and regresses four length ratios to represent the relative sliding offsets of the corner points of the quadrilateral to the corresponding sides of the bounding rectangle. This method can accurately detect multi-directional license plate regions and effectively avoids the sequential label points. Extensive experimental tests were conducted on datasets such as CCPD, AOLP, and CLPD. On the CCPD test set, SvRetina-LPD demonstrated a high accuracy of 97.9%, surpassing existing methods. In addition, SvRetina-LPD also demonstrated excellent detection accuracy in other subsets of CCPD, including various challenging scenarios, demonstrating its ability to accurately detect multi-directional license plates in unconstrained scenarios.