Convolution Neural Network for Image Registration

Image registration plays a fundamental role in many computer vision applications, such as medical image processing, camera pose estimation, etc. It includes the estimation of geometric transformation and image warping. The goal of this study is to design, develop and evaluate an end to end trainable convolution neural network (CNN) which can learn homography and affine transformation’s controlling parameters. The trained model can be used for registering unseen image pairs to provide a better-quality image registration. The training data is based on publicly available image dataset and truth tagging is accomplished by generating synthetic image pairs instead of depending on manual annotations.