Scale-Invariant Multidirectional License Plate Detection with the Network Combining Indirect and Direct Branches

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• License plate detection in natural scenes is difficult due to scale and orientation variations.
• Existing indirect and direct methods have limitations in precision and detecting missed vehicles.

Purpose of the Study:

• To propose a novel end-to-end network for robust license plate detection in the wild.
• To improve the detection of small-sized and multidirectional license plates.
• To combine indirect and direct detection strategies for enhanced performance.

Main Methods:

• A novel network integrating indirect and direct detection branches is developed.
• The indirect branch uses a coarse-to-fine scheme with vehicle-plate relationships for high-precision small plate detection.
• The direct branch mitigates false negatives by detecting plates directly, complemented by a four-corner localization refinement for multidirectional plates.

Main Results:

• The proposed end-to-end network effectively detects small-sized and multidirectional license plates.
• Experimental results demonstrate significant performance improvements over standalone indirect and direct methods.
• The method achieves superior accuracy in real-world license plate detection scenarios.

Conclusions:

• The combined indirect and direct approach offers a significant advancement in license plate detection.
• This novel architecture provides a robust solution for detecting challenging license plates in diverse natural scenes.
• The method represents a pioneering integration of indirect and direct techniques into a unified, end-to-end trainable network.