Research On Some Key Technologies Of Image Correspondence And Its Applications

Images,as the media through which humans learn about the world,carry rich knowledge.When identifying and inferring an unknown object in an image,people not only identify its category(e.g.,“What is it?”),but also try to find its connection with the known objects in the memory(e.g.,“What is it like?”).By analyzing the similarity between them,people can get meaningful information.Image correspondence,as a key technology in computer vision,can quickly build this visual similarity and enable computers to identify and understand an object like humans do.However,in real-world scenarios,the difference in illumination,camera location and angles will lead to huge variations across images,even when these images are taken from the same scene.How to overcome the illumination,scale and angle variations and accurately build visual similarity between them is an important technical problem for image correspondence.In addition,when handling images which only share semantic similarity,how to overcome the variations of object appearances,colors and textures to build accurately correspondences between them have become another important problem.These problems have posed a great challenge to image correspondence estimation,and are worth studying.By centering on several key technologies on image correspondence,this paper mainly focuses on dense image correspondence related problems.This paper studies scale-and rotation-invariant dense image correspondence algorithm,dense image correspondence algorithm for semantically similar objects,object skeleton detection algorithm and its application in image correspondence,and the application of dense image matching in computer vision tasks.The contributions of this paper are as follows:1.The existing dense image correspondence algorithms cannot well handle scale,rotation and other geometric transformation issues.Hence,this paper proposes a propagation-induced dense image correspondence algorithm.By taking into account both low-level representation and geometric transformation,this paper proposes a scale-and rotation-invariant dense descriptor and a propagation-induced matching framework.This method adopts the estimated geometric transformation to guide the matching of low-level descriptors.Therefore,it successfully solves the problem during image matching caused by the scale,rotation and viewpoint variations.Experiments show that this algorithm can produce accurate results.2.Dense descriptors represent the local structure of input image and contain little semantic information,and thus they often become invalid when matching images with similar semantic content but different appearances.To solve this problem,this paper proposes an object-aware dense semantic correspondence method.Based on an object-aware hierarchical graph model,this method performs dense semantic correspondence first from the whole object,then to regional structure and finally to every pixel.In this way,the interference from background noise can be avoided,and matching accuracy can be improved.Furthermore,this method learns feature representations of images in an object-driven manner and better defines the visual similarity between images.3.Object skeleton contains the structure information of an object that can be used to improve the accuracy of dense semantic correspondence.Therefore,this paper first proposes a multi-scale bi-directional fully-convolutional network model for object skeleton detection.The proposed model builds a multi-scale feature pyramid and uses bi-directional propagation for feature fusion so as to improve object skeleton detection.According to the experimental results in public skeleton detection databases,the proposed model produces the acurrate results.Then this paper also integrates object skeleton information into the dense semantic correspondence framework,and proposes a skeleton-aware dense semantic correspondence method.The experiments show that the accuracy of dense semantic correspondence is boosted due to the introduction of object structure information into dense semantic correspondence framework.4.To illustrate how image correspondence technology aids computer vision tasks,this paper integrates the dense image correspondence technique into two different computer vision tasks.Specifically,we take co-segmentation and semantic segmentation as examples,and use dense image correspondence technique as the key components of the two tasks.Because dense image correspondence technique is able to provide rich information,these problems can be solved effectively.