Transformer-Based Unsupervised Image Registration Using SSIM and Homography Loss for Steady Camera and Aerial Videos
Golnoosh Abdollahinejad, Matin Hashemi · 2023
Image registration is an essential and initial block in the pipeline of computer vision tasks and systems. It is defined as transforming a moving image into a target image with the minimum difference when aligned. Unlike previous work for general-purpose datasets, aerial images suffer from mechanical shakes, leading to deformed distortion similar to medical volumetric image registration task. Inspired by medical approaches, we use a Transformer-based network with a semi-unlimited receptive field Swin block to produce a general output for each pixel named flow matrix. Flow matrix is utilized instead of regressing parameters of transformation matrix with a fixed degree of freedom that cannot handle the structural difference between images. This leads to introducing a new loss function based on the Structural Similarity Index Measure (SSIM) and embedding Homography transformation as a regularization term. Combining a generalized-designed network and loss function based on problem definition significantly enhanced results.