License Plate Image Super Resolution Using Generative Adversarial Network(GAN)

Abstract

Super resolution of images in the field of Computer Vision is a widely used for the conversion of images into high resolution without the loss of pixel data into the images. Due to fast movement of vehicles and low quality of camera the image cannot be verified easily so, the techniques of Generative Adversarial network have been applied for the Super resolution of license plate Images which works to recover the loss data of license plate images without loss of pixel data. Earlier, mean square error (MS E) and peak signal to noise ratio (PSNR) was used as content loss to minimize the error but at optimal minimization the images get over smoothen and pixel data were lost. This paper has proposed and applied VGG-19 as pretrained neural network along with MSE and PSNR to minimize the content loss which overall optimizes the perpetual loss, and over smoothness of the images gets controlled which saves pixel data. Later, the pre-trained neural network is integrated with Generative Adversarial Network [GAN] of discriminator and generator to produce high resolution images. Taking the PSNR as an evaluation metrices for the images, it increases from 26.184 to 28.696 and accuracy from 58% to 84%.