Stitching of drone images using unsupervised deep learning

This research presents a method to produce extensive images using images taken by UAVs. Which faces various challenges such as the existence of parallax, low overlap between images and change of depth of field in images. Various methods have been tried to solve these challenges. Each of these methods has problems. To solve these challenges, in this research, an extensive image generation framework using unsupervised deep learning was proposed. This method has investigated the problem in the stationary UAV mode with low overlap between images and the moving UAV mode with medium overlap between images. The proposed method includes two separate neural networks that are used for the two stages of image alignment and the stage of combining and reconstructing the image. In the first stage, the unsupervised deep homography estimation method, in which multi-scale feature extraction using the feature pyramid method, along with the use of a content-based correlation method to match the features and to improve the alignment of images from a spline transformation Thin plate and an accurate alignment matching method were used and a method was proposed to modify the geometric structure in the non-overlapping area of the target image. In the second step, a UNet network was used to combine and reconstruct the images and solve challenges such as distortion and brightness differences in the stitched images. Compared to the existing methods, the proposed method provides a better output compared to similar works in both scenes with moderate parallax and scenes with low overlap, and the stitched image is free of distortion or differences in brightness.

کلیدواژه ها:

Image Stitching ، Panorama ، deep homography ، thin plane splines