Research On Deep Homography Network Based On Super-resolution And Reparameterization

Image stitching is a crucial research area in computer vision,aimed at merging multiple narrow field-of-view images into a single wide field-of-view image.Traditional and deep learning-based methods are the two main solutions.Both involve feature extraction,image registration,homography matrix calculation,deformation,and fusion,but differ in feature point detection.Traditional methods use SIFT or Harris for feature detection and one or more homography matrices for stitching.Deep learning-based methods are still developing,with two popular approaches: Use deep learning methods to extract keypoints from images and calculate a homography matrix based on keypoints or using a deep learning network to predict the offsets of image corners.Homography estimation is one of the most important parts of image stitching technology.Improving homography estimation ability is equivalent to improving image stitching ability.Nowadays,deep learning-based stitching solutions have certain advantages over traditional manual feature extraction methods.Therefore,this paper intends to research deep learning-based homography prediction networks.The main work includes:(1)Proposing a super-resolution-based deep homography estimation network,adding a feature fusion module based on dilated convolution to the pyramid network structure of the deep homography network,which can achieve better intermediate layer information extraction and better utilization of the feature maps output by the simple backbone network.By predicting image offsets,the network can perform homography estimation from coarse to fine.The homography prediction effect using the feature fusion module is better than that without the module.(2)Deep convolutional neural networks are becoming more complex and computationally expensive due to their increasing number of parameters.Therefore,there is an urgent need to improve model performance while reducing computation time.Reparameterization is a method that can effectively improve performance while reducing computation time and inference complexity during the inference stage.This method is very compatible with the goal of deep homography estimation,which is to achieve high performance during training and quickly find the corresponding homography for subsequent work during forward inference.Therefore,this paper proposes a reparameterizationbased backbone network to improve the performance of deep homography prediction.The experimental results show that this network outperforms deep homography networks that use traditional backbone networks.