Detection of QR code based on mask R-CNN using synthetic image under perspective transform

Various types of 2D barcodes including QR code, Data Matrix are widely used in diverse industries. Recently, QR code is adopted in mobile phone. They require QR code occupy some amount of area on the image to correctly operate. In other application of mobile robot navigation where QR code is used as landmarks, we need to detect QR code in various scale and lighting condition. Traditional approaches operate well under restricted conditions also they require the setting of many parameters. In this paper, we deal with the detection of QR code in wild including various lighting condition and scale. We adopt the Mask R-CNN [11] for the detection of QR code. It requires many training images to have good performance. We present a method that uses composite images which is made under general perspective transform. QR code under reference poses and real images are composited by the presented method. We present various experimental results under diverse configuration of hyperparameters. Sequential step of first training using many composite images then finally training using real images show the best performance. Experimental results show the feasibility of presented approach.