Geometric Information Modeling Of Local Features In Image Retrieval

In the past few decades,the rapid development of digital technology has led to an explosive growth in the number of digital images.Especially,with the advent of the era of mobile Internet,hundreds of millions of users begin to take photos and upload them through mobile terminals everyday.These ever-increasing images pose an urgent need for highly-efficient image retrieval techniques.Bag-of-Words is a classical image retrieval model due to its effectiveness of efficiency.However,Bag-of-Words model ignores the geometric/spatial information of local features,such as: angle,scale,and space coordinates,thus limits its retrieval performance.In this thesis,a density-based geometric verification method is proposed to improve image retrieval performance.In the reranking stage of image retrieval,the score of each local-feature match is weighted by its probability density in the parameter space of similarity transformation before adding to the overall similarity.Experiments show that our geometric verification method can effectively improve the MAP of local-feature-based image retrieval.Nowadays,good image retrieval performances have been achieved by using aggregated features extracted from convolutional neural networks,and there still exists room to make further improvements.The second work of this thesis is to study spatial weighting scheme for aggregated features to improve image retrieval.We first improve the spatial weighting method based on class activation mapping(CAM).The CAM operation is firstly preposed to improve the aggregation efficiency,and then a channel weighting method is proposed based on the response mean to further improve the MAP.Finally,the improved CAM-based weighting scheme is combined with a response aggregation weighting scheme to achieve better MAP.