Multi-source image registration is a complicated but essential processing task in various vision problems, such as image fusion and object detection. Conventional methods are only capable of handling images with negligible parallax and near-infinite sight distance, such as remote sensing images. However, when the parallax between multiple images is significant, the lighting conditions are poor, or there is significant interference between the target and foreground, the registration performance can dramatically degrade. To address these challenges associated with image acquisition, in this paper, we propose a novel and robust registration method for multimodal images by utilizing an adaptive training scheme. The proposed method begins by detecting basic feature points and generating an initial coarse registration result using the SuperPoint network and the SuperGlue network. Optimal registration points are suitably determined using the DEGENSAC algorithm with a reasonable threshold. Abundant experimental results and quantitative comparisons demonstrate that our proposed scheme achieves robust and state-of-the-art registration performance for multimodal images, even complicated imaging scenarios. Additionally, for the first time to the best of our knowledge, we experimentally determine an optimal selection scheme of the target image for registration, providing valuable insights for registration tasks involving more than two images in a practical sense.