Vehicle Model Recognition using SRGAN for Low-resolution Vehicle Images

초록

An enhanced vehicle model recognizer for low-resolution images is proposed in where SRGAN (Super Resolution Generative Adversarial Network) is used to enhance the image quality and CNN (Convolutional Neural Network) is used to classify the vehicle model from the enhanced images. Many previous vehicle model classifiers trained with only the high-resolution front-images of vehicles have low accuracy against the low-quality images captured by CCTV cameras in real environments. To correctly classify the vehicle model from the low-quality images of arbitrary directions, SRGAN is first used to transform the low-resolution image into the corresponding high-resolution image. Then the direction of the vehicle in the image is determined and the vehicle model is recognized based on the pre-determined direction. The accuracy of the proposed vehicle model classifier is evaluated as 78%, higher than that of the classification without SRGAN.

키워드