Unsupervised Homography Estimation Method Based on Edge Detection and Feature Fusion

Traditional homography estimation uses methods such as SIFT and ORB in feature extraction but suffers from problems such as feature sparsity in large baseline scenes with large viewing angles and low overlap, causing inaccurate homography estimation. This paper proposes an unsupervised method that combines edge detection and feature fusion to solve the above problems. Firstly, three layers of image features are extracted by the feature pyramid; secondly, the edge detection module extracts the edge features of the input image and feature map; then, the features are fused by adaptive feature fusion and attentional feature fusion; finally, correlation is computed layer-by-layer using the fused features to estimate the homography. In this paper, experiments have been conducted to verify the method's effectiveness by calculating the PSNR and SSIM metrics. Experimental results show that this method can estimate homography more accurately than other related methods when stitching large baseline images.