An image registration method based on image diffusion features

Abstract

This invention belongs to the field of image processing technology and relates to an image registration method based on image diffusion features, comprising the following steps: Step 1: Acquire a real-time two-dimensional ground image, denoted as I(x,y), where (x,y) are the coordinates of pixels in the image; Step 2: Obtain the two-dimensional Fourier spectrum I based on the two-dimensional ground image I(x,y). fft (x,y); Step 3: Based on diffusion theory, analyze the two-dimensional Fourier spectrum I fft (x,y) is subjected to diffusion processing to obtain the diffused two-dimensional Fourier spectrum I. fft_new (x,y); Step 4: Based on the diffused two-dimensional Fourier spectrum I fft_new (x,y) Obtain the diffused normal image and output it; Step 5: Perform feature point detection on the diffused normal image and generate a feature point description image; Step 6: Match the feature point description image with the pre-stored georeferenced image, and output the matching result after converting the location result.

Description

Technical Field

[0001] This invention relates to visual enhancement technology and image matching technology in the field of image processing, and particularly to a scene matching and positioning method for unmanned aerial vehicles (UAVs), specifically an image registration method based on image diffusion features. Background Technology

[0002] Navigation technology is crucial in unmanned systems, serving as the foundation for ensuring the system can execute missions and return safely. In recent years, with the development of computer technology, image processing technology, sensor technology, and artificial intelligence technology, scene matching visual navigation, as an autonomous, passive, electronically interference-resistant, and high-resolution visual navigation method, has been increasingly widely used in unmanned systems.

[0003] Currently, most methods for coarse localization of UAVs using aerial view matching, both domestically and internationally, rely on traditional image feature extraction. Some researchers have addressed template matching in aerial images using techniques based on scale-invariant feature descriptors. For example, some researchers have studied UAV localization in GPS-free environments and utilized optical flow to determine the UAV's position. They used inter-frame transformation for pose tracking and histograms of orientation gradient features for registration on Google Maps, and then employed particle filtering for finer localization. A relatively mature approach utilizes SIFT features extracted from aerial images as a solution to the UAV image matching problem. Other approaches employ deep learning for end-to-end image matching to achieve UAV localization; however, this method relies on large amounts of labeled data, lacks interpretability, and demands significant computational power. Summary of the Invention

[0004] Objective of the Invention: This invention provides an image registration method based on image diffusion features for UAV visual positioning. The method includes two steps: real-time image preprocessing and matching. The UAV system acquires real-time overhead image information from the ground via an electro-optical pod, then converts the image to the frequency domain and calculates diffusion features based on the weak decomposition of the diffusion equation. These diffusion features are then used to enhance the image information. Finally, the inverse of the fast Fourier transform is used to output the enhanced result, thus obtaining the preprocessed image output. This invention is low-cost, simple to operate, has low computational complexity, and achieves good image registration results.

[0005] The technical solution of this invention is:

[0006] An image registration method based on image diffusion features includes the following steps:

[0007] Step 1: Acquire real-time 2D ground images. The acquired 2D ground images are denoted as I(x,y), where (x,y) are the coordinates of the pixels in the image.

[0008] Step 2: Obtain the two-dimensional Fourier spectrum I based on the two-dimensional ground image I(x,y). fft (x,y);

[0009] Step 3: Based on diffusion theory, analyze the two-dimensional Fourier spectrum I fft (x,y) is subjected to diffusion processing to obtain the diffused two-dimensional Fourier spectrum I. fft_new (x,y);

[0010] Step 4: Based on the diffused two-dimensional Fourier spectrum I fft_new (x,y) is used to obtain the normal image after diffusion, and then output it.

[0011] Step 5: Perform feature point detection on the diffused normal image and generate a feature point description image;

[0012] Step 6: Match the feature point description image with the pre-stored georeferenced image, and output the matching result after converting the location result.

[0013] Furthermore, in step one, a dedicated UAV optoelectronic pod device is used to collect real-time two-dimensional ground images.

[0014] Furthermore, in step three, the diffusion processing includes: calculating the mapping relationship between the diffusion characteristics of the image formation process and the features of the formed image in the Sobolev space, thereby performing image diffusion processing.

[0015] Furthermore, in step three, the diffusion process includes the following steps:

[0016] Step a): Calculate the two-dimensional Fourier spectrum I fft The parameter values ​​used for feature extraction at each coordinate (x, y) in (x, y) are given by the following formula:

[0017]

[0018] Where, x max y is the image height; max α represents the image width; α, β, λ, and k are all algorithm parameters, and all are positive numbers.

[0019] Step b) Superimpose the parameter values ​​used for feature extraction at each coordinate (x, y) into the original two-dimensional Fourier spectrum I. fft In (x,y), the formula is as follows:

[0020] I fft_new (x,y)=I fft (x,y)*[T(x,y)+1]

[0021] Among them, I fft_new(x,y) is the two-dimensional Fourier spectrum after diffusion processing.

[0022] Furthermore, in step a), the parameter value T(0,0) = γ at the DC component (0,0);

[0023] Where γ is an algorithm parameter and is a positive number.

[0024] Furthermore, in step two, a two-dimensional discrete Fourier transform is performed on the ground two-dimensional image I(x,y) to obtain its two-dimensional Fourier spectrum I. fft (x,y).

[0025] Furthermore, in step four, the diffused two-dimensional Fourier spectrum I... fft_new The normal image after diffusion is obtained by performing a two-dimensional discrete Fourier inverse transform on (x,y).

[0026] Furthermore, in step five, the SIFT algorithm is used to detect feature points in the diffused normal image.

[0027] The present invention has the following beneficial effects:

[0028] (1) The image registration method based on image diffusion features of the present invention broadens the problem-solving ideas in similar fields. It is the first realization of this theory in this field and has high research value.

[0029] (2) The image registration method based on image diffusion features of the present invention focuses on signal enhancement of valuable features, which can provide a more robust solution for visual positioning under severe weather conditions and has great research potential. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the image registration method based on image diffusion features proposed in this invention;

[0031] Figure 2 This is a schematic diagram illustrating the matching operation between the feature point description image obtained in this invention and a pre-stored geographic reference image.

[0032] Figure 3 This is a schematic diagram of diffusion processing of an RGB three-channel image in Embodiment 1 of the present invention. Detailed Implementation

[0033] The present invention will now be described in detail with reference to specific implementation processes. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0034] This invention takes the ground top view image obtained by the UAV during flight as the input object, calculates the stable physical quantities with feature information carried in the image, and explores how to use physical quantities with feature information to enhance the semantic information of the image, thereby achieving better image matching effect. The image enhancement algorithm adopted is mainly based on the formation mechanism of image diffusion.

[0035] To address the issue of drones using image matching for localization in the event of GPS interruption or interference, this invention designs and implements a new image enhancement algorithm based on the image diffusion formation mechanism. This algorithm aims to achieve better top-view image matching performance in general image datasets and practical applications, ultimately enabling drones to achieve good localization in real-world environments using only the collected top-view images.

[0036] Combination Figure 1 The image registration method based on image diffusion features of the present invention will be described in detail, such as... Figure 1 As shown, it includes the following steps:

[0037] S11: Use a dedicated UAV electro-optical pod to collect real-time ground top view images, and then read the ground top view images into the system. That is, read the original images collected by the camera of the electro-optical pod into the system. Here, the collected two-dimensional ground image is denoted as I(x,y), where (x,y) are the coordinates of the pixels in the image.

[0038] S12: Obtain the two-dimensional Fourier spectrum of the input image using a two-dimensional discrete Fourier transform, denoted as I. fft (x,y);

[0039] S13: The two-dimensional Fourier spectrum obtained from S12 fft (x,y) is processed using diffusion theory and certain algorithms;

[0040] In this step, the core technique used in processing the two-dimensional Fourier spectrum using diffusion theory is to calculate the mapping relationship between the diffusion characteristics of the image formation process and the features of the formed image in the Sobolev space. This leads to a new image processing algorithm based on diffusion, which reduces the problem to the inner product space and solves it in the frequency domain. The specific algorithm is as follows:

[0041] 1) Calculate the parameter values ​​used for feature extraction at each coordinate (x, y), using the following formula:

[0042]

[0043] Where, x max y is the image height; max α represents the image width; α, β, λ, and k are all algorithm parameters, and all are positive numbers.

[0044] In addition, let the parameter value at the DC component be:

[0045] T(0,0)=γ

[0046] Here, γ is also an algorithm parameter.

[0047] 2) Finally, the calculated parameter values ​​used for feature extraction are superimposed into the original two-dimensional Fourier spectrum, as shown in the following formula:

[0048] I fft_new (x,y)=I fft (x,y)*[T(x,y)+1]

[0049] Among them, I fft_new (x,y) is the processed two-dimensional Fourier spectrum.

[0050] S14: Processed two-dimensional Fourier spectrum I fft_new (x,y) The normal image is obtained by using the two-dimensional discrete Fourier inverse transform;

[0051] S15: Output the image processed by the diffusion algorithm; thus, the image processed by the diffusion algorithm can be obtained.

[0052] S16: Use the SIFT algorithm to detect feature points in the image obtained in S15 and generate a feature point description image. Compared with conventional feature point detection, the feature points in the acquired two-dimensional image are more prominent after the diffusion processing described in this invention. As a result, the feature point description image after detection by the SIFT algorithm will be more explicit, which increases the accuracy of subsequent matching and improves efficiency.

[0053] S17: Perform a matching operation between the image information obtained in S16 and the pre-stored georeferenced images;

[0054] S18: Perform location transformation on the matching results and output them.

[0055] Example 1

[0056] Combination Figure 2 and Figure 3The process involves processing a ground image captured by a drone using a diffusion equation and then performing image registration on a global map. First, for diffusion feature enhancement, the original image is read using the RGB color space and split into R, G, and B channels. Then, for each channel's two-dimensional image, a discrete two-dimensional Fourier transform is performed, followed by corresponding operations in the frequency domain. Finally, an inverse discrete two-dimensional Fourier transform is applied. The processed two-dimensional images are then merged to obtain the image after diffusion feature enhancement preprocessing. Next, for image matching, the SIFT algorithm from the field of image processing is used to detect feature points in the preprocessed image and generate corresponding feature point descriptions. Finally, this image is matched with a pre-stored georeferenced image to achieve drone localization.

[0057] The embodiments disclosed herein are merely preferred embodiments of the present invention. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the present invention, and are not intended to limit the present invention. Any modifications and variations made by those skilled in the art within the scope of this specification should fall within the protection scope of the present invention.

Claims

1. An image registration method based on image diffusion features, characterized in that: The method includes the following steps: Step 1: Acquire real-time 2D ground images, and record the acquired 2D ground images as follows: ,in These are the coordinates of pixels in the image; Step 2: Based on the 2D ground image Its two-dimensional Fourier spectrum was obtained. ; Step 3: Analysis of the two-dimensional Fourier spectrum based on diffusion theory Diffusion processing was performed to obtain the diffused two-dimensional Fourier spectrum. ; Step 4: Based on the diffused two-dimensional Fourier spectrum Perform a two-dimensional discrete Fourier inverse transform to obtain the diffused normal image and output it; Step 5: Use the SIFT algorithm to detect feature points in the diffused normal image and generate a feature point description image; Step 6: Match the feature point description image with the pre-stored georeferenced image, and output the matching result after converting the location result.

2. The image registration method based on image diffusion features according to claim 1, characterized in that: In step one, a dedicated UAV optoelectronic pod device is used to collect real-time two-dimensional ground images.

3. The image registration method based on image diffusion features according to claim 1, characterized in that: In step three, the diffusion processing includes: calculating the mapping relationship between the diffusion characteristics of the image formation process and the features of the formed image in the Sobolev space, thereby performing image diffusion processing.

4. The image registration method based on image diffusion features according to claim 3, characterized in that: In step three, the diffusion process includes the following steps: Step a): Calculate the two-dimensional Fourier spectrum Each coordinate in The parameter values ​​used for feature extraction are given by the following formula: in, Image height; Image width; , , , All are algorithm parameters, and all are positive numbers; Step b) Set each coordinate The parameter values ​​used for feature extraction are superimposed into the original two-dimensional Fourier spectrum. In Chinese, the formula is as follows: in, This is the two-dimensional Fourier spectrum after diffusion processing.

5. The image registration method based on image diffusion features according to claim 4, characterized in that: In step a), the parameter value at (0,0) of the DC component. ; in, These are algorithm parameters and must be positive numbers.

6. The image registration method based on image diffusion features according to claim 1, characterized in that: In step two, the ground two-dimensional image Its two-dimensional Fourier spectrum is obtained by performing a two-dimensional discrete Fourier transform. .

Images