Unsupervised Stitching Method for Blisk DR Images with Dual-Energy X-Ray Radiography

Traditional image stitching methods rely heavily on the quality of feature matching. However, the blisk Digital Radiography (DR) images tend to exhibit low contrast and repetitive textures, which can frequently cause incorrect alignment during the stitching process. Consequently, this often leads to the appearance of artifacts and distortions in the stitched image. In this paper, we propose an unsupervised image stitching method specifically designed for blisk using dual-energy radiography. Firstly, we adopt a multi-scale enhancement process to enhance image contrast and improve detail clarity in the original images. Secondly, we introduce an unsupervised image stitching method consisting of a coarse alignment module, an image reconstruction module, and a multi-energy image fusion module. The unsupervised coarse registration module automatically learns image features and achieves initial coarse alignment through homography transformation. The unsupervised image reconstruction module learns spatial transformations and deformation patterns of the images, resulting in the reconstruction of pixel-level features in the stitched image and minimizing stitching artifacts. Finally, a dual-energy image fusion module based on Nonsubsampled Contourlet Transform (NSCT) is employed to fuse the stitched images, resulting in globally, high-resolution blisk DR images. Through subjective visual evaluations and quantitative metric analysis, our method demonstrates optimal stitching results on both simulated and real datasets. Our approach effectively preserves image details and structures while minimizing stitching seams, thereby achieving a high-resolution, artifact-free stitching result.