Research On Image Matching Based On Multi-layer Semantic Features

Image matching aims to generate a pixel-level correspondence for semantical similar images,i.e.a flow field which describe the correspondence between image pair.Due to the basicity of image matching tasks,the increase of accuracy and the improvement of computational efficiency is the key for many application tasks of computer vision.Image matching is an important task to understand the correspondence between image semantics and pixel points.There are still many problems to be solved,especially for matching the objects with same class label.The angle change of the image capturing,the scale change of the object,and the complexity of the background limit the useable range of matching algorithm.Therefore,this paper will proceed from the huge difference between low-level pixels point and high-level semantics,use the convolutional neural network effectively to solve the huge intra-class difference between image pairs to be matched.The main work is as follows:Firstly,we propose an image matching method based on convolutional neural network.By analyzing the correspondence between convolutional neural network features and image visual information,we confirm the validity of its application to image-intensive semantic matching.Introduced convolutional neural network features with stronger visual information expression ability into image matching task which makes the accuracy improved from the feature level.Secondly,we propose an image matching method based on multi-layer semantic features.We match the image pair hierarchically using feature complementary advantages between different layers of convolutional neural networks.The structure of the feature pyramid is used to fuse different levels of information,and the matching guide constraint is added to construct the connection between different levels.Finally,we propose an optimization method of image matching based on cycleconsistency,which applies the cycle-consistency masterly.We modify the existing guidelines to improve the scope and strength of reliable matching.Combining the method of hierarchical matching,we use the guidance graph generated by the cycle-consistency and optimized objective function to achieve better purpose of processing large intra-class changes between images.