딥 러닝 피쳐를 통한 영상 정합 기술

This paper proposes an image registration algorithm that uses features obtained from a deep convolutional neural network. The proposed algorithm first extracts feature maps from the image using pre-trained convolutional neural network. Then we construct cost function based on the feature similarity and motion smoothness. The cost function is optimized by belief-propagation algorithm and coarse-to-fine scheme. Experimental results show that the features trained by deep learning leads better registration results than conventional descriptors such as SIFT.