Multimodal image registration seeks to establish spatial correspondences between images from different modalities. However, the generalization of existing methods to diverse multimodal scenarios remains limited due to significant modality gaps. To address this challenge, we propose CrossNet, a feature-based multimodal registration framework that integrates automated keypoint detection, cross-modal descriptor extraction, and correspondence consistency enforcement through constraints on intermediate feature maps. To further enhance robustness and minimize annotation requirements, we introduce a cross-supervision strategy that promotes consistent predictions, generates pseudo-aligned images, and leverages unlabeled image pairs during training. Extensive experiments across a variety of multimodal settings demonstrate that CrossNet significantly reduces the reliance on aligned training data while achieving state-of-the-art performance in both feature matching and image registration tasks.