Infrared Image Stitching Method Based on Channel Attention and Transformer

Infrared panoramic observation systems typically rely on image stitching techniques to achieve wide field-of-view and high spatial resolution imaging. However, traditional stitching methods are heavily dependent on the quality and quantity of feature points, making it difficult to obtain high-quality stitched images when there is significant disparity or poor feature extraction. To address this challenge, we propose an unsupervised homography-based image stitching framework. First, we use a transformation model that transitions from global homography to local Thin Plate Spline to precisely align overlapping regions while preserving the shape of non-overlapping areas. Next, we design a feature extraction network that incorporates channel attention mechanisms to help the network more effectively extract key features. Finally, we apply seam-driven unsupervised image fusion techniques to further eliminate disparity artifacts. Experimental results show that this method does not require complex geometric feature designs and, compared to existing mainstream algorithms, generates clearer and more natural panoramic images, demonstrating broad practical application potential.