A neural network-based image registration method between static orbit spectrum segments

Abstract

The invention discloses a neural network-based image registration method between static orbit spectrum segments, and the method comprises the steps: carrying out the marking of multi-spectrum remote sensing images, collected by an area-array camera, of the same scene, carrying out the dividing of a training set and a test set according to a proportion; carrying out noise reduction on the originalimage pair by adopting a Kalman filtering algorithm; extracting image features of the training set and the test set by adopting an SIFT algorithm to construct a feature database; designing a network structure, determining network initial parameters, taking image feature points among spectrum segments as input to train a network model complex, and determining an optimal network parameter; changingthe network structure and repeatedly training to obtain an optimal network model; and finally, verifying by adopting the test set, and if the registration precision does not meet the requirement, expanding the data set and repeatedly training until the requirement is met. According to the invention, the network model is adopted to accurately represent the non-linear relationship between the feature points of the images of different wave bands, and the problem that the same feature point changes greatly and is unstable in the images of different wave bands in a traditional registration method is solved.

Description

technical field [0001] The invention relates to a neural network-based image registration method between geostationary orbital spectrum segments, which belongs to the intersecting field of aerospace remote sensing image processing and artificial intelligence method application. Background technique [0002] Image registration refers to the process of geometrically calibrating images of overlapping areas captured by the same or different sensors from different perspectives of the same scene at different time periods. Image registration technology has a wide range of applications in many fields, such as medical image processing, image analysis, remote sensing fusion, computer vision, etc. Image registration technology is a hot spot in the current scientific research field. Remote sensing image registration is an indispensable step in remote sensing image fusion, multispectral classification, environmental monitoring and image mosaic. [0003] The current remote sensing image ...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to overcome the problem that the registration accuracy cannot meet the requirements due to the large variation and instability of the same feature point in different bands in the traditional remote sensing image registration method based on image features, and propose a neural network based method. The image registration method between spectral segments of geostationary orbits based on the network constructs a training data set based on visible, short-wave and medium-wave images collected by on-orbit cameras, trains and adjusts the network structure to obtain higher registration accuracy, and provides remote sensing images between spectral segments. Registration Offers a New and Efficient Way

Images