Research On Multi-source Image Matching Algorithm Based On Twin Neural Network

The problem of large-scale multi-source image matching mainly focuses on the computational work of paired image matching and geometric verification.This problem has been a hot topic in the field of image processing.For most image processing,the images that can be matched are meaningful,which requires finding the relationship between the images.The more images with the same features,the higher the matching rate.However,with the development of photography technology,more and more images are acquired,and the images actually needed are getting lower and lower in total acquisition images,and it is more and more difficult to obtain useful image information.At present,for large-scale multi-source image matching,traditional algorithms based on manual design features have low precision and poor robustness.The use of supervised depth features requires high training,and the pairing method based on twin neural networks cannot extract local similarities.Therefore,in view of these problems,this paper proposes a large-scale multi-source image matching method based on previous research.The main research contents are as follows:1.Unsupervised depth feature extraction method based on stack self-encoder.Image feature quality has a direct impact on the final effect of the image algorithm.Compared with the limitations of manual design features and lack of generalization,depth features have good generalization and multi-level semantics,and show good performance in image matching.At the same time,in order to reduce the computational cost of deep feature training and large-scale sample dependence,this paper designs a stack convolutional self-encoder model suitable for image pairing for unsupervised depth feature extraction.A multi-layer network is used to implement a stack-type self-encoder,and unsupervised feature learning is performed to obtain high-order sparse expression.2.Research on image matching algorithm based on improved siamese network structure.Aiming at the problem that the traditional siamese network structure can not measure the local similarity well,this paper proposes an improved siamese network structure,which improves the objective function and corrects the relative difference function into a weighted dot product.The global similarity measure becomes a measure of layout similarity.Experiments were carried out on the image set data provided by the Visual Geometry Group of Oxford University,and the method was verified and compared with the existing mainstream methods.Experiments show that the method of this paper can effectively improve the accuracy and efficiency of image matching.