Unsupervised registration method for multi-modal image

Abstract

The invention belongs to the field of image processing, discloses an unsupervised registration method for a multi-modal image, and is used for solving the problem of large error caused by direct training through similarity loss between multi-modal images due to large appearance difference of the multi-modal images in the existing unsupervised image registration. The image registration method comprises the following steps: firstly, putting a floating image B into a convolutional neural network in which features such as color and texture are similar to those of a fixed image A to obtain an image C converted from the image B; and secondly, putting the image C and the fixed image A into an unsupervised registration network, outputting transformation parameters, re-sampling the image C, performing similarity measurement on the image C and the image A, updating the network by taking the image C as a loss function, iterating for multiple times to obtain optimal transformation parameters, and re-sampling the image B by using the transformation parameters to obtain a registered image. According to the method provided by the invention, unsupervised registration can be carried out on the multi-modal image, and the precision of unsupervised registration is greatly improved.

Description

technical field [0001] The invention relates to an image registration method, in particular to an unsupervised registration method for multimodal images. Background technique [0002] Image registration is an important field of image processing. Image registration refers to the geometric alignment of two or more images, so that each point on the floating image has a unique point corresponding to it on the fixed image. The purpose is to The purpose of finding the spatial transformation relationship between different images is to remove or suppress the geometric inconsistency between the images to be registered. Image registration is an important part of the processing and a necessary prerequisite for image fusion, analysis and target recognition. [0003] The traditional registration method is an iterative optimization process: first extract the feature information in the image, match the features between the images, then select the spatial transformation method to obtain the...

AI Technical Summary

Problems solved by technology

[0004] Direct registration methods from unsupervised learning have achieved great success in unimodal images, but there are many difficulties in dealing with the registration of multimodal images

Because the appearance of different modalities is very different, it is inefficient to train the network directly through the similarity loss between multi-modal images, and the registration accuracy is poor, which leads to the problem of multi-modal registration than single-modal registration. It is much more difficult, so the current unsupervised registration method is still mostly used to deal with single-modal registration problems, and semi-supervised registration methods tend to be used when dealing with multi-modal image registration problems.

Images