Dual-Fisheye Image Stitching via Unsupervised Deep Learning

Constructing panoramic images from a dual-fisheye lens has been increasingly used along with the recent booming of new computer vision applications, such as virtual reality (VR) and augmented reality(AR). The recent development of deep learning (DL) techniques has shed new light on the field of image stitching, but little research has been conducted on DL-based dual-fisheye image stitching. In this work, we propose an unsupervised deep learning method for dual-fisheye image stitching. Specifically, we construct a stitching system consisting of fisheye distortion correction, unsupervised image reconstruction, and image edge rectangularization blocks. Experiment results show that the proposed scheme can perform accurate and natural stitching of two images, and exceed the traditional method in PSNR, SSIM, RMSE, MSE, and other performance indicators.