A Fast Unsupervised Image Stitching Model Based on Homography Estimation

Image stitching is the synthesis of multiple partial image segments into a complete and continuous panoramic image through effective image alignment and seamless fusion techniques. It can achieve a wider field of view and richer information for display and analysis. Most deep learning-based image stitching methods have significant advantages in improving accuracy, but they are not suitable for real-time applications due to multiple iterations of computation or deeper network depth. To deal with this problem, a fast unsupervised image stitching model is proposed in this article. In the proposed model, an adaptive feature extraction module (FEM) for deformation is designed, and then a fast unsupervised learning-based image alignment network is proposed. In addition, a stitching restoration network with a smaller number of parameters is presented to remove the redundant and unnecessary sampling and convolution operations in general deep learning-based models. Finally, some experiments are conducted on both the synthetic and real-scene datasets. The total stitching accuracy of the proposed model is higher, and the details of the output images are clearer. The proposed can achieve 1.79, 26.54, and 0.86 in RMSE, peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) on the alignment results, respectively, which are better than those of the state-of-the-art methods. Furthermore, the comparison results prove that the proposed model can effectively reduce memory loss, and achieve a fast unsupervised image stitching, with a very small model size.