A Research Of Unsupervised Learning Of Interest Point Detection And Description

Representation learning methods such as Deep Learning have provided representative features for a number of computer vision applications which improves their performance significantly.But the local features outputted by this kind of methods are dense and contain a great deal of redundant information,which increase computation complex and lower overall accuracy.Interest point methods can extract sparse representative features so that maintain stable accuracy and significantly improve computation efficiency.That's why interest point methods are widely used in a number of field such as image matching.With analyses of weakness of existing interest point methods,an unsupervised learning method of interest point is studied in order to achieve high performance.Firstly,existing methods of interest point are introduced,which can be divided into hand-craft interest point and learning based interest point.The superiorities and weaknesses of existing interest point models are discussed.Then learning based interest point are summarized from four aspects: training datasets,optimization objectives,model architectures and optimization algorithms.Secondly,an unsupervised learning method of interet point detection is presented.The objective of this method is constructed with sparsity constraint and repeatability measure so that the sparsity and repeatability of interest point can be maximized.An alternative optimization algorithm is proposed to maximize this objective iteratively,which has good convergence and efficiency.In experiments the detection model is named as Sparse Repeatable Network(SRN)which is implemented with fully convolutional network.Experimental results on several image matching benchmarks demonstrate the combinations of SRN detector and a number of existing descriptors give better or similar matching performance compared to the original detectors designed to applicable for these descriptors,which exhibits powerful generalization ability of SRN.Further experiments reveal this generalization ability benefits from the objective function constructed with sparsity and repeatability.Finally,an unsupervised learning framework of interest point detection and description are presented by extending above method of interest point detection.The objective of the framework is formulated as joint probability distribution of the properties of the extracted points,so any properties with probability formulations can be integrated into this objective.Sparsity,repeatability and discriminability are selected as the essential properties to instantiate the joint probability distribution.The optimization algorithm of this framework is Expectation Maximization algorithm(EM).An approximation of EM is proposed to further improve efficiency of training.In the experiments the detection and description model is named as Property Network(PN)which is implemented with fully convolutional network.The experiments demonstrate that PN outperforms state-of-the-art methods on several image matching benchmarks.