Unsupervised deep image stitching based on cascaded warping and multi-scale seam prediction for USV wide field-of-view generation
Zhilin Yang, Yong Yin, Qianfeng Jing, Zeyuan Shao, Haitong Xu, Carlos Guedes Soares · Autonomous Transportation Research · 2025
This study proposes an unsupervised deep image stitching method based on cascaded warping and multi-scale seam prediction to address the challenges of low texture and large parallax image stitching for maritime scenes. In the image alignment stage, convolutional and transformer architectures are integrated to enhance multi-scale feature modelling in low-texture scenes, achieving high-precision alignment through a three-stage cascaded process of global homography and thin-plate spline (TPS) transformations. In the image composition stage, a decoupled multi-scale seam prediction approach is introduced, effectively mitigating artefacts by fusing multi-resolution soft-encoded seam masks. Experimental tests are conducted on a self-constructed maritime image stitching dataset (MISD) covering diverse textures, parallax, and illumination conditions. Results demonstrate that the proposed method outperforms traditional and deep learning approaches in terms of alignment accuracy and seam quality while maintaining low computational complexity, achieving an effective balance between precision and efficiency. This study presents a robust and efficient solution for wide field-of-view generation in unmanned surface vehicles (USVs), significantly enhancing the perception range of maritime intelligent transportation systems and laying a foundation for multi-view video stitching.