Detection of korea license plate by mask r-cnn using composite image

License plate detection on an image is the first necessary step for the automatic recognition of license plate. In this paper, we adopt Mask R-CNN [11] to detect a license plate on an image. It requires many training images to cope with over-fitting that occurs when training samples are smaller than numbers of parameters. In general, more than 200K images are required for the stable training, but it requires large amount of time and cost. In this paper, we present a method that uses already available open dataset. We use two open dataset of CCPD [8] and BDD [18]. CCDP dataset provides locations of four corner points on an image. But, they contain license plate of China. First, Korean license plate image is made by referencing the design rule. Then, Korea license plate image is projected onto the corresponding positions of CCPD image. Four corresponding points between Korea license plate and CCPD images is used in the computation of perspective transform. Two different types of training images from CCPD and BDD dataset are used in the training of Mask R-CNN, and they are applied to image that contains real Korea license plate. Training using composite images from CCPD shows better performance than that of BDD on the real Korea license plate image. Experimental results show the feasibility of presented approach.