An Efficient License Plate Detection Approach With Deep Convolutional Neural Networks in Unconstrained Scenarios
Jia Wei, Mingshan Xie
Abstract
License plate (LP) detection is a crucial task for Automatic License Plate Recognition (ALPR) systems. Most existing LP detection networks can detect License plates, but their accuracy suffers when license plates (LPs) are tilted or deformed due to perspective distortion. This is because these detectors can only detect the region where the LP is located, and even the most advanced object detectors struggle in unconstrained scenarios. To address this problem, we propose a lightweight Deformation Planar Object Detection Network (DPOD-NET), which can correct the deformed LPs of various vehicles (e.g., car, truck, electric motorcycle, bus) by detecting the LP corner points. Accordingly, the distortion associated with perspective is mitigated when we adjust the LP to a frontal parallel view through the LP corners. To optimize small errors between the predicted and true values of the LP corner points, we propose an LPWing loss function. Compared with the commonly used L1 function, the LPWing loss is derivable at the zero position, and the gradient will be bigger when errors are smaller. This enables the model to converge faster at the position where the error is close to zero, resulting in better convergence when the error between the true values and predicted values is small. In addition, the paper presents a stochastic multi-scale image detail boosting strategy, which effectively augments the dataset. Finally, to objectively evaluate the effectiveness of LP corner detection approaches, we present a dataset (LPDE-4K) including various LP types (e.g., color, country, illumination, distortion). We test the performance on various datasets, and our approach outperforms other existing state-of-the-art approaches in terms of higher accuracy and lower computational cost.