AN EFFECTIVE METHOD FOR DETECTING LICENSE PLATES USING DEEP CONVOLUTIONAL NEURAL NETWORKS IN UNCONSTRAINED ENVIRONMENTS

One of the most important tasks for Automatic License Plate Recognition (ALPR) systems is license plate (LP)detection. Although the majority of LP detection networks currently in use are capable of detecting licenseplates, perspective distortion causes these networks to lose accuracy when LPs are tilted or twisted. This isbecause even the most sophisticated item detectors have trouble in unrestricted situations, and these detectorsare limited to detecting the area where the LP is located. In order to solve this issue, we suggest a lightweightDeformation Planar Object Detection Network (DPOD-NET), which uses LP corner point detection to rectifythe distorted LPs of a variety of vehicles, including cars, trucks, electric motorcycles, and buses. Therefore, byadjusting the LP to a frontal parallel view through the LP corners, the perspective distortion is lessened.We provide an LPWing loss function to minimize tiny discrepancies between the true and forecasted values ofthe LP corner points. The gradient will be larger when mistakes are less, thus the LPWing loss can be derived atthe zero position in contrast to the widely used L1 function. This makes it possible for the model to convergemore quickly at the point where the error is near zero, leading to improved convergence when there is little errorbetween the true and predicted values. The study also offers a stochastic multi-scale picture detail boostingtechnique that successfully enhances the dataset. Lastly, we provide a dataset (LPDE-4K) with a variety of LPkinds (e.g., color, country, lighting, distortion) in order to impartially assess the performance of LP cornerrecognition techniques. Our method performs better than other state-of-the-art methods in terms of increasedaccuracy and reduced computational cost when tested on a variety of datasets.