Research On Image Matching Algorithm Based On Siamese Network

Image matching technology is used to match two or more images transformed by translation,rotation or folding.At present,face recognition technology,image dryness,human pose recognition and other technologies have been widely used in our life,and the development of these technologies is based on image matching technology.Visible image matching technology is the basis of many image editing or processing,which has important research significance and practical value.An image matching algorithm based on siamese network is studied in this paper.The main contents of this paper are as follows:Firstly,the principle of image matching algorithm based on convolutional neural network is analyzed and introduced.Compared with the traditional image matching algorithm,the image matching algorithm based on convolution neural network shows better performance.Secondly,the data set was preprocessed,and the images in the original data set were rotated horizontally,vertically,90 degrees,180 degrees,270 degrees,and the number of images was amplified to 6 times the original.in addition,according to the image features extracted from the pixel domain and the image features extracted from the transform domain,there is a great degree of complementarity.therefore,the discrete cosine transform is used to extract the image features from the transform domain,thus realizing the feature fusion between the pixel domain and the transform domain.Then,the pyramid residual module is used to achieve invariance on the scale of convolutional neural networks.Pyramid residual module enriches the extraction of image detail features and improves the matching accuracy.for the above improved methods,the model is built and trained.the experiment proves that the proposed algorithm is the best in the dual-channel network structure.the experimental results are tested on the Brown dataset with a score of 3.65,which is superior to the original model.Finally,the improved siamese network structure is optimized by adding batch normalization layer and depth separable convolution,respectively.Compared with the two optimization methods,the experiment shows that the addition of batch normalization layer accelerates the convergence of the network and the training speed of the network increases to 2 times.At the same time,the matching accuracy of the algorithm is improved,and the score is 3.36.in addition,the deep separable convolution is used instead of the traditional convolution mode,which reduces the number of parameters of the network and optimizes the overall structure of the network.Experiments show that the overall parameters of the network are reduced under the premise that the accuracy of the algorithm is slightly reduced.