Super-Resolution GANs for Enhancing License Plate Detection from Distorted Inputs

Capturing an image of a vehicle's license plate is the most utilized method to identify vehicles. However, during a real-world investigation, the vehicle's license plate is not always visible in the image due to poor resolution, motion blur, and non-normal angle. This paper proposes a novel approach to address this limitation by designing and training multiple neural networks in the Super-Resolution Generative Adversarial Network structure. To generalize the trained networks to real-world images affected by rotation, motion blur, and low resolution, we introduce homography transformations during training data generation. Then, the generated data were used to train multiple networks. After the networks has been trained they were tested with the validation dataset. The trained networks were evaluated using three key metrics: visual quality, peak signal-to-noise ratio (PSNR), and optical character recognition (OCR). Results demonstrate significant improvements in visual clarity, with a notable increase in PSNR and OCR accuracy compared to traditional interpolation methods.