Research On Sonar Image Registration Based On Convolutional Neural Network

Sonar image registration aims to find the best global transformation model for mapping source images to moving images.It is the basic link of sonar image processing and an important technical means for marine resource detection.From the perspective of spatial transformation parameters and similarity measurement of image registration,its application is studied,and sonar image registration algorithm based on regression-correction network and similarity calculation method based on similarity ranking model are successively studied.The main work of this thesis can be summarized as:(1)Introduce hierarchical processing strategy in sonar image registration,and design a sonar image registration algorithm based on regression-correction network.The regression network is used to predict the 3D geometric transformation parameters of the registration pair.The correction network includes the geometric transformation network and the comparison network,and the transformation parameters are corrected according to the similarity of the image pair.Through the combination optimization of the two networks,the registration from coarse to fine is realized.Comparative experiments are performed on two sonar datasets and three registration methods,and the experimental results verify the effectiveness of the algorithm.(2)Similarity calculation method based on similarity ranking model: A model that can characterize the similarity of images is studied.By introducing a ranking loss function based on triples into a multi-scale deep neural network,supervised learning of similarity information achieves similarity measurement.The two sonar datasets are compared with three classical similarity metric functions and three classical similarity calculation methods to verify the effectiveness of the algorithm.