Image stitching refers to the process of combining multiple local images from the same scene with overlapping areas into a panoramic image through algorithms. In low-light scenes, image stitching is of practical significance in many fields. However, images captured in low-light conditions often suffer from issues such as noise, loss of detail, and uneven brightness, which can lead to stitching failure. To address these challenges, this paper proposes an image stitching framework that combines low-light image enhancement and an improved U-Net structure. The framework is divided into two stages: in the image enhancement stage, the details are restored and the colour is enhanced by a dual-branch enhancement network. The two branches exchange information through a fusion module to generate an enhanced image. In the image stitching stage, the warped image is obtained using an image alignment method, and then the stitching is optimised using an improved U-Net structure. A residual block is introduced to improve the stitching quality and generate the final stitched image. In addition, in order to evaluate the performance of the proposed framework in real low-light scenes, a special low-light scene dataset was constructed for network training. The results show that the algorithm described in this paper performs well in low-light image enhancement, significantly improving the image’s detail and colour reproduction capabilities, and the overall image quality is significantly improved. The enhanced image significantly improves the stitching accuracy when stitching, effectively eliminating stitching artifacts.