Image Feature Matching Based On Content-Aware And Its Application In Model Fitting

Computer vision technology has become a research hotspot in the field of artificial intelligence in recent years.Image fusion,threedimensional reconstruction,image stitching,and image registration are widely used in fields such as medical assisted diagnosis,intelligent transportation,and so on.Robust model fitting algorithms calculate fundamental matrices or essential matrices through feature matching to achieve computer vision analysis tasks such as image stitching and image registration.Image feature matching is a key step of robust model fitting,which aims to find matches between images,i.e.,matching feature points in two or more images one by one.Existing image feature matching methods have poor performance in situations such as multi-planar,low texture,and large variations in light intensity.The matching pairs they obtain may contain more mismatching and noise,resulting in reduced model fitting accuracy.This article focuses on feature matching and robust model fitting,and mainly conducts the following research:(1)A feature matching method based on content-aware is proposed.Firstly,this method uses the existing marginalizing sample consistency model fitting strategy to pre-align the input images to reduce the impact of image rotation or distortion on feature matching accuracy;Then,in response to the problem of low accuracy of correct matching due to the small number of valid features extracted by existing methods,this method designs a content-aware block composed of a pre feature extractor and a mask predictor,wherein the pre feature extractor is used to calculate the initial feature map,and the mask predictor is used to generate a probability map.The probability map is used to weight the initial feature map as input to subsequent feature extractors.The weighted feature map guides the feature extractor to obtain more features of the interested region;Finally,in view of the problem that there are still mismatches in the obtained matching pairs,this method introduces a progressive pruning strategy to remove the mismatches and improve the matching accuracy.This method performs feature matching and homography estimation experiments on the publicly available HPatches dataset.The experimental results of feature matching show that this method achieves good results,with an average matching accuracy of 88%,surpassing the suboptimal method(87%);The experimental results of homography estimation show that the average estimation accuracy is 80%,which is 1% higher than the suboptimal method.(2)An efficient robust model fitting method based on preference analysis and information theory is proposed.Aiming at the negative impact of data point sampling from different structures(different planes,objects,etc.)on model fitting performance when the proportion of outliers contained in input data is high,this method first uses a data sampling strategy based on preference analysis to generate more reliable model assumptions to calculate the discriminant matrix;Then,using information theory principles to calculate a more discriminative sparse matrix,spectral clustering of data points is performed to accurately segment data points,thereby improving the performance of model fitting.This method uses the published Adelaide RMF and Hopkins 155 datasets for fundamental matrix estimation and three-dimensional motion segmentation experiments.The fundamental matrix estimation experiment shows that compared with the other seven methods,the minimum average estimation error of this method on all test image pairs is reduced by 35% compared to the suboptimal method,and the standard deviation of the average estimation error is reduced by 32% compared to the suboptimal method;Experiment on 3D motion segmentation shows that the average estimation error of this method on all video sequences containing moving objects is reduced by5.2% compared to the suboptimal method.