Panoramic Image Stitching Using Double Encoder–Decoders

Panoramic image stitching is to synthesize similar images into a complete image. It is one of the critical scientific issues in computer vision. It has important applications in medical imaging and virtual vision. The traditional algorithms of the panoramic image stitching mainly include three steps: (1) extract feature points; (2) match the similar feature points; and (3) blend the combined images. However, the traditional method cannot adaptively learn from the external environment, leading to the restricted applications. Inspired by deep learning, this paper proposes a novel algorithm based on the neural networks. This network can stitch two related patches into a panoramic image accurately. Instead of extracting the feature points in the images, it directly learns the transformation relationships between images. Simultaneously, no complicated post-processing is needed to reprocess the combined images, but a neural network is used to directly generate the panoramic image with better color and texture information. © 2021, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. part of Springer Nature.