RoTIR: Rotation-Equivariant Network and Transformers for Zebrafish Scale Image Registration

Image registration is essential for aligning features of interest from multiple images. With the recent development of deep learning techniques, image registration approaches have advanced to a new level. In this work, we present Rotation-Equivariant network and Transformers for Image Registration (RoTIR), a deep-learning-based method for aligning zebrafish scale images captured by light microscopy. This approach overcomes the challenge of arbitrary rotation, translation detection, and the absence of ground truth data. We employ feature-matching approaches based on Transformers and general E(2)-equivariant steerable CNNs for model creation. Besides, an artificial training dataset is employed for semi-supervised learning. Results show that RoTIR successfully achieves the goal of zebrafish scale image registration.