<p>Image stitching is a fundamental task in computer vision that plays an important role in applications such as panoramic virtual reality generation and environmental perception for autonomous driving. The primary technical challenges lie in improving geometric alignment accuracy and visual fusion quality in complex scenes. Traditional feature point-based methods often fail in low-texture or large-parallax scenarios, while existing deep learning approaches have limited adaptability to local non-rigid deformations and commonly suffer from color inconsistencies and visible seams during fusion. To address these issues, we propose a dual-stage unsupervised image stitching framework. In the alignment stage, a three-level cascaded network progressively extracts and couples high-frequency edge details with low-frequency structural information through a frequency-aware allocation mechanism. By integrating a global homography matrix with local thin-plate spline transformations, the framework achieves coarse-to-fine multi-scale registration. The fusion stage employs a hybrid Mamba-CNN architecture incorporating SR2D-based pixel rearrangement upsampling and a difference-aware module, enhanced by a multi-scale attention mechanism to improve seam quality and reduce texture and color mismatches. Extensive experiments on the standard UDIS-D dataset demonstrate that our method achieves a peak signal-to-noise ratio of 26.58 and a structural similarity index of 0.874, surpassing existing methods in visual quality. Additionally, evaluations on the cross-domain CrossBench dataset validate the strong generalization capability of our approach. Moreover, our method maintains a real-time processing speed of approximately 0.163 seconds per image pair under large-parallax conditions, showing strong robustness and promising potential for practical deployment.</p>