CN-121982081-A - Offset image registration method based on side window Gaussian edge condition diffusion model

CN121982081ACN 121982081 ACN121982081 ACN 121982081ACN-121982081-A

Abstract

The invention discloses an offset image registration method based on a side window Gaussian edge condition diffusion model, and belongs to the technical field of remote sensing image processing. The method comprises the steps of firstly obtaining a reference image and an offset image to be registered, performing rough registration to obtain initial transformation parameters and a rough registration image, then extracting a side window Gaussian edge image of the reference image as a condition constraint, performing iterative refinement on the rough registration image by using a diffusion model to obtain a refined image, determining residual transformation parameters according to the refined image and the reference image, combining the initial transformation parameters to obtain final transformation parameters, and finally transforming the offset image to obtain a registered image. In the aspect of registration accuracy, stable sub-pixel level registration is realized through strong constraint of the side window Gaussian edge, large-scale training data and additional geographic reference information are not needed, and nonlinear radiation difference and large-scale geometric offset interference are effectively resisted.

Inventors

SHAN XIN
QIN HONG
Zhao Lingyuan
KANG RUOFAN
Luo Zifei
ZHANG YAN

Assignees

环天智慧科技股份有限公司
环天智慧(成都)科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260407

Claims (10)

1. A method for registering offset images based on a side window Gaussian edge condition diffusion model is characterized by comprising the following steps: Acquiring a reference image and an offset image to be registered; performing coarse registration on the reference image and the offset image to obtain initial transformation parameters and a coarse registration image; Extracting a side window Gaussian edge graph of the reference image; Performing iterative refining on the coarse registration image by using a diffusion model under the condition of a side window Gaussian edge map to obtain a refined image; determining residual transformation parameters according to the refined image and the reference image; determining final transformation parameters according to the initial transformation parameters and the residual transformation parameters; And transforming the offset image according to the final transformation parameters to obtain the registered image.
2. The offset image registration method based on the side window Gaussian edge condition diffusion model of claim 1, wherein the coarse registration comprises the following steps: respectively constructing a multi-scale side window Gaussian filtering scale space for the reference image and the offset image to obtain a multi-scale image sequence, wherein each layer corresponds to different scale parameters; Extracting a binary edge graph on each layer of image in the scale space, and detecting key points on the binary edge graph to obtain a key point set; constructing a descriptor for each key point and estimating a main direction; and matching the key point set of the reference image and the key point set of the offset image, estimating initial transformation parameters and generating a rough registration image.
3. The offset image registration method based on the side window Gaussian edge condition diffusion model of claim 2, wherein when a multi-scale side window Gaussian filter scale space is constructed, a side window Gaussian filter is applied to an image, gaussian kernels are respectively applied to a plurality of directional windows around a pixel by the side window Gaussian filter, and window output with the smallest difference with a target pixel is selected as a filtering result, and the mathematical expression is as follows: Wherein, the For the intensity of the target pixel, For a local window bounded by pixel i, The direction is indicated as such, Representing a set of discrete directions, Normalized Gaussian weights; Is a node Is a feature of the original feature of (a).
4. The offset image registration method based on the side window Gaussian edge condition diffusion model of claim 3, wherein the steps of extracting the binary edge map and detecting the key points comprise: A gradient operator is applied to each layer of image in the scale space to calculate a gradient amplitude, and fixed thresholding is carried out on the gradient amplitude to obtain a binary edge map; Performing corner detection on the binary edge graph to obtain a corner response graph; and obtaining a uniformly distributed key point set by adopting square coverage self-adaptive non-maximum suppression.
5. The method for registration of offset images based on side window Gaussian edge condition diffusion model of claim 4, wherein constructing descriptors and estimating principal directions comprises: Applying first-order steerable filtering to the key point neighborhood to obtain a first-order gradient map, and then calculating second-order gradient amplitude and phase; determining a main direction of the key point according to the peak value of the second-order phase histogram; improved gradient position and direction histogram descriptors are constructed using log polar partitioning.
6. The offset image registration method based on the side window Gaussian edge condition diffusion model of claim 5, wherein the matching of the key point set comprises the following steps: screening high-confidence matching point pairs by adopting a nearest neighbor distance ratio, and estimating initial transformation by utilizing a rapid sampling consistency algorithm; And in the second stage, matching is carried out in a local search window of the initial transformation constraint by using the combined distance, and the matching is refined again by using a rapid sampling consistency algorithm, so that updated initial transformation parameters and a rough registration image are obtained.
7. The method for registering offset images based on a side window Gaussian edge condition diffusion model according to claim 1, wherein the forward noise adding process of the diffusion model is defined as gradually adding Gaussian noise to a coarse registration image until the coarse registration image approaches a standard Gaussian distribution, and the noise adding amount of each step is controlled by a predefined noise scheduling coefficient.
8. The offset image registration method based on the side window Gaussian edge condition diffusion model of claim 1 is characterized in that a lightweight U-shaped network architecture is adopted in the inverse denoising process of the diffusion model, a side window Gaussian edge graph is injected as a condition through a cross attention mechanism, and the calculation mode of the cross attention mechanism is as follows: Wherein, the Representing the target currently to be queried, query Features, key matrix and value matrix from noise image Splicing features from the side window Gaussian edge map and the reference image; representing dot product similarity calculation to obtain each And all of The larger the value the stronger the correlation is represented by the original correlation score of (a); And representing a scaling factor, predicting the noise added in the current step by the lightweight U-shaped network, and gradually recovering the image through iterative denoising.
9. The offset image registration method based on the side window Gaussian edge condition diffusion model of claim 8, wherein the parameter quantity of the lightweight U-shaped network is smaller than 10M, the backbone network is initialized by adopting the weight of the pre-training image denoising model, and fewer iterations smaller than the preset maximum steps are adopted in reasoning.
10. The method for registering offset images based on the side window Gaussian edge condition diffusion model of claim 1, wherein determining residual transformation parameters comprises performing dense optical flow estimation or subpixel matching on a refined image and a reference image to obtain a residual transformation matrix representing a tiny geometric offset.

Description

Offset image registration method based on side window Gaussian edge condition diffusion model Technical Field The invention belongs to the technical field of remote sensing image processing, and particularly relates to an offset image registration method based on a side window Gaussian edge condition diffusion model. Background The remote sensing image registration is a front core technology for realizing multi-source, multi-time-phase and multi-view remote sensing data collaborative analysis, and has the core function of mapping an image to be registered with geometric offset into a standard coordinate system of a reference image through geometric transformation, so as to provide geometric consistency guarantee for subsequent image fusion, ground feature identification, dynamic monitoring and other applications. The current remote sensing image registration technology is mainly divided into three technical routes based on regional gray scale, local feature and deep learning, and various schemes have technical short plates which are difficult to overcome in practical engineering application. The registration method based on the combination of frequency domain migration and traditional features relies on global Fourier transformation to realize image domain self-adaption processing, is high in computation complexity and long in time consumption when processing large-size single-pair remote sensing images, is prone to low-frequency component distortion problems when facing nonlinear radiation differences and weak texture areas, directly causes feature matching failure and registration deviation, is limited in modeling capacity by relying on a large-scale labeling data set to complete generation of countermeasure network training, is extremely poor in adaptation to extremely simple application scenes of single reference images matched with single offset images, can introduce irreversible geometric blurring in the pseudo-optical image reconstruction process, greatly reduces registration accuracy, is limited to an inherent framework of traditional feature matching although the registration method based on directional filtering and traditional feature matching has certain robustness in structural feature extraction, and is required to rely on geographic reference data such as DEM (digital elevation model) as auxiliary constraints, and the hardware and data cost of technical application are remarkably improved. In summary, the existing registration technology cannot simultaneously consider the registration efficiency, robustness and sub-pixel level precision under the extremely simple condition of only relying on a single group of reference-offset images, and cannot effectively overcome the registration problems caused by nonlinear radiation difference, large-scale geometric offset and weak texture area interference, and cannot meet the actual requirements of high-precision remote sensing data processing. Aiming at the defects of the prior art, the invention provides an offset image registration method based on a side window Gaussian edge condition diffusion model, stable coarse registration is realized by constructing a multi-scale side window Gaussian filter scale space, then a light diffusion model is driven by taking a reference image edge structure as a strong priori condition to finish residual refining, and on the premise of no additional geographic reference data and no large-scale training data, the high-efficiency, robust and high-precision registration of remote sensing images is realized, so that the technical blank of the prior art in a single-pair image extremely-simple registration scene is filled. Disclosure of Invention The invention aims to provide an offset image registration method based on a side window Gaussian edge condition diffusion model, which aims to solve the problems of large calculated amount, high data dependence, poor weak texture suitability, insufficient registration precision and high application cost in the existing remote sensing image registration technology in the background technology. In order to solve the technical problems, the invention adopts the following technical scheme: an offset image registration method based on a side window Gaussian edge condition diffusion model comprises the following steps: Acquiring a reference image and an offset image to be registered; performing coarse registration on the reference image and the offset image to obtain initial transformation parameters and a coarse registration image; Extracting a side window Gaussian edge graph of the reference image; Performing iterative refining on the coarse registration image by using a diffusion model under the condition of a side window Gaussian edge map to obtain a refined image; determining residual transformation parameters according to the refined image and the reference image; determining final transformation parameters according to the initial transformation parameters and the residual transf