Research On Convolutional Neural Network Based Image Matching Algorithm

In recent years,with the rapid development of intelligent facilities,computer vision technology has become one of the important technologies in many fields.Image feature extraction and matching is an important part of image processing.Feature detection of images is to extract images from high-dimensional features with certain representative features,which is a kind of compression of images,and these features can satisfy the application of images in their target tasks.Feature extraction in image matching is generally based on mathematical methods to extract feature points and calculate feature descriptions.But the ability of these feature descriptions express weakly,and there are limitations in the use of some scenarios such as target positioning and target tracking with the large deformation in the image.In view of the problem of insufficient characterization ability,this paper used the deep learning method to extract the features of the image.Due to the depth of the model structure of deep learning,the characteristics of the bottom layer of the image are continuously merged and refined to obtain the deep layer of the image on the semantic level.The main research content completed in this paper includes:(1)Investigating the research status and characteristics of image matching,Combining the advantages of deep learning in image feature extraction,the image feature extracted by convolutional neural network are used to optimize the image matching accuracy.(2)A feature descriptor based on convolutional neural network is proposed.In the process of training,the network not only considers the similarity of feature descriptions of the same feature points of two images,but also considers the dissimilarity between different feature points,which makes the network have stronger expression.When optimizing the objective function,the stochastic gradient descent method is used.For the data selection,an online method is used to construct a triplet data set.And in this way,it can reduce the problem of data redundancy of the training data set which used randomly selecting heterogeneous samples in offline mode.(3)An image matching algorithm based on convolutional neural network is proposed.Feature points with certain invariance are used in the algorithm.For the feature description,the convolutional neural network is employed.Based on the feature points and their descriptions,the K-nearest neighbor algorithm of two-way search strategy is used to optimize the search process and matching accuracy.For eliminate the mismatch matching,the random sample consensus algorithm is performed.The experimental results show that the performance of this algorithm has a better image matching effect than traditional algorithms in the difference of illumination intensity,geometric and appearance.The algorithm in this thesis is based on a large amount of data for training,and the acquired knowledge is stored in its model structure.Through the model extracted high-level features,it can express images more abstractly and can face more complex target problems.Therefore,the algorithm of this thesis is applied to the feature description after the feature extraction of the image,which gives the feature better expression ability and enhances the matching effect.Also,it can be applied to the image-based target localization or target tracking.