Super Resolution for number plate detection using Generative Adversarial Network.
Identification and recognition of number plate is very difficult from low resolution images due to poor boundary and contrast. Our goal is to identify the digits from a low-quality number plate image correctly, but correct detection was exceedingly difficult in some cases due to the low-resolution image. Another goal of this paper was to upscale the image from a very low resolution to high resolution to recover helpful information to improve the accuracy of number plate detection and recognition. We have used Enhanced- Super-Resolution with Generative Adversarial Network (SRGAN). We modified native Dense Blocks of the Generative Adversarial Network with a Residual in Residual Dense Block model. In addition to Convolutional Neural Networks for thresholding. We also used a Rectified Linear Unit (ReLU) activation layer. The plate image is then used for segmentation using the OCR model for detection and recognizing the characters in the number plates. The Optical character recognition (OCR) model reaches an average accuracy of 84% for high resolution, whereas the accuracy is 4% - 7% for low resolution. The model’s accuracy increases with the resolution enhancement of the plate images. ESRGAN provides better enhancement of low-resolution images than SRGAN and Pro-SRGAN, which the OCR model validates. The accuracy significantly increased digit/alphabet detection in the number plate than the original low-resolution image when converted to a high-resolution image using ESRGAN.
If you find this work useful please cite our original paper here - https://link.springer.com/article/10.1007/s11042-022-14018-0