License Plate Image Resolution Enhancement Using Super-Resolution Generative Adversarial Network

Abstract

Car license plates play a crucial role in vehicle identification. However, obtaining high-quality images of license plates, particularly in real-world scenarios, poses a significant challenge due to limitations in camera resolution. To address this issue and enhance image quality for license plate identification, this research proposes a novel approach employing two Convolutional Neural Networks: a Generator and a Discriminator. These networks are trained simultaneously within a Generative Adversarial Network framework. The primary objective is to improve the visual quality of low-resolution license plate images to facilitate accurate tag recognition. The training process involves generating high-resolution (HR) license plate images with random tag numbers, followed by downsampling and inputting them into the Generator to produce an upscaled super-resolution (SR) image. The Discriminator then discerns whether the image is SR or HR, and this feedback refines the Generator's performance. Low-resolution images spanning from 16×16 to 128×128 are up-scaled to super-resolution, ranging from 32×32 to 256×256. Evaluation metrics, including Peak Signal-to-Noise Ratio (PSNR) and Optical Character Recognition (OCR), are employed to compare the resultant SR images with interpolated images of the same size combinations. Notably, the PSNR of the SR images is nearly double that of the interpolated low-resolution (LR) images. Additionally, the OCR accuracy of LR images, such as 16×16 and 32×32, improved to 66.59% and 71.28% from 0.02% and 5.74%, respectively. This proposed image enhancement method was utilized for license plate image enhancement; which will offer law enforcement agencies a powerful tool to extract actionable intelligence from surveillance footage.