CN-122024072-A - Remote sensing image change detection method and device based on edge contour guidance

CN122024072ACN 122024072 ACN122024072 ACN 122024072ACN-122024072-A

Abstract

The invention discloses a remote sensing image change detection method and device based on edge contour guidance, wherein the method comprises the steps of obtaining a remote sensing image; the method comprises the steps of obtaining a profile graph according to a remote sensing graph in combination with a Canny operator, wherein the remote sensing graph comprises an early remote sensing graph and a late remote sensing graph which have preset time intervals, the profile graph comprises an early profile graph and a late profile graph, multi-time-step noise adding is carried out on the remote sensing graph and the profile graph, a preset multi-scale feature extraction is carried out by combining a denoising diffusion probability model, multi-time-step feature graphs of the remote sensing graph and the profile graph under each scale are obtained respectively, enhancement fusion is carried out on the multi-time-step feature graphs of the remote sensing graph and the profile graph under each scale respectively, and enhancement fusion results under each scale are fused in sequence according to the scale from small to large, so that a change detection result graph is obtained.

Inventors

YANG BIN
TANG ZIWEN
WANG ZEPING

Assignees

湖南大学

Dates

Publication Date: 20260512
Application Date: 20260408

Claims (9)

1. The remote sensing image change detection method based on edge contour guidance is characterized by comprising the following steps of: The method comprises the steps of obtaining a remote sensing image, combining the remote sensing image with a Canny operator to obtain a profile image according to the remote sensing image, wherein the remote sensing image comprises an early remote sensing image and a late remote sensing image with preset time intervals, and the profile image comprises an early profile image and a late profile image; Carrying out multi-time-step noise adding on the remote sensing image and the profile image, and carrying out preset multi-scale feature extraction by combining a denoising diffusion probability model to respectively obtain multi-time-step feature images of the remote sensing image and the profile image under each scale; and respectively carrying out enhancement fusion on the multi-time-step feature images of the remote sensing image and the profile image under each scale, and sequentially fusing enhancement fusion results under each scale from small to large according to the scale to obtain a change detection result image.
2. The edge contour guidance-based remote sensing image change detection method according to claim 1, wherein performing enhanced fusion on the multi-time-step feature map of the remote sensing map and the contour map at each scale comprises: And under the condition that the two continuous convolution layers and the ReLU activation function are used for respectively carrying out depth feature extraction and fusion on the early enhancement representation and the late enhancement representation to obtain early feature representation and late feature representation under each scale, and calculating absolute differences of the early feature representation and the late feature representation to obtain an initial change feature map under each scale.
3. The method for detecting the change of the remote sensing image based on edge contour guidance according to claim 2, wherein the step of sequentially fusing the enhanced fusion results under each scale from small to large to obtain a change detection result graph comprises the following steps: Under the condition of the minimum scale, carrying out self-adaptive calibration on an initial change feature map to obtain output features of the minimum scale, under the condition of the remaining scales, carrying out element-by-element addition feature fusion and self-adaptive calibration on first input of each scale and the initial change feature map in sequence according to the sequence from small scale to large scale to obtain output features of each scale, wherein the first input of each scale comprises a result of carrying out up-sampling twice on the output features of one scale on each scale, and carrying out 3X 3 convolution, reLU activation function processing and 3X 3 convolution on the output features of the maximum scale in sequence to obtain a change detection result map.
4. The edge contour guidance-based remote sensing image change detection method of claim 3, wherein the adaptive calibration comprises: Carrying out channel dimension self-adaptive calibration and space dimension self-adaptive calibration on input features in parallel to respectively obtain a first result and a second result; and adding the first result and the second result to obtain the output of the self-adaptive calibration.
5. The edge contour guided remote sensing image change detection method of claim 4, wherein the adaptive calibration of channel dimensions comprises: Sequentially carrying out global average pooling, full-connection network processing, sigmoid activation function processing and element-by-element multiplication processing on the input features to obtain the first result; the full-connection network processing comprises a dimension reduction part and a dimension increase part which are cascaded, wherein the dimension reduction part sequentially comprises 1 multiplied by 1 convolution and ReLU activation function processing, and the dimension increase part comprises 1 multiplied by 1 convolution.
6. The edge contour guidance-based remote sensing image change detection method according to claim 5, wherein the adaptive calibration of the spatial dimension comprises: And carrying out 1×1 convolution, sigmoid activation function processing and element-by-element multiplication processing on the input features in sequence to obtain the second result.
7. The edge contour guidance-based remote sensing image change detection method according to claim 6, wherein performing multi-time step denoising on the remote sensing map and the contour map comprises: According to preset noise scheduling, carrying out noise adding processing on the remote sensing image and the profile image for a preset number of time steps: ; ; ; ; wherein the time step = 、、 ... ; Is the maximum time step; representing an early telemetry map; representing a late telemetry graph; Representing an early profile; Representing a late profile; Representing time steps The noise adding result of the early remote sensing image is carried out; Representing time steps Adding a noise result to the lower late remote sensing graph; Representing time steps The noise adding result of the lower early profile; Representing time steps Adding a noise result to the lower late remote sensing graph; Representing a gaussian distribution from the standard A random noise tensor of the mid samples; Representing the fidelity coefficient.
8. The method for detecting the change of the remote sensing image based on edge contour guidance according to claim 7, wherein the step of extracting the preset multi-scale features by combining the denoising diffusion probability model to obtain the multi-time step feature map of the remote sensing map and the contour map under each scale respectively comprises the following steps: The method comprises the steps of inputting multi-time step noise adding results of a remote sensing image and a profile image into a U-Net module of a denoising diffusion probability model respectively, carrying out a down-sampling process and an up-sampling process according to an input image by the U-Net module, inputting output results of each scale of the down-sampling process into the up-sampling process of the same scale in the down-sampling process and the up-sampling process, splicing the output of the last scale in the up-sampling process, and then carrying out up-sampling of the same scale; and respectively processing the multi-time-step noise adding results of the remote sensing map and the profile map, and then storing the output results of each scale in the up-sampling process, namely the multi-time-step characteristic map of the remote sensing map and the profile map under each scale.
9. A remote sensing image change detection device based on edge profile guidance for use in the method of any one of claims 1 to 8, wherein the device is adapted to implement the method of any one of claims 1 to 8.

Description

Remote sensing image change detection method and device based on edge contour guidance Technical Field The invention relates to the technical field of remote sensing image processing, in particular to a remote sensing image change detection method and device based on edge contour guidance. Background The change detection algorithm accurately identifies the dynamic change of the earth surface coverage and the environment by comparing the multi-temporal remote sensing images, and has key effects in the fields of urban expansion monitoring, disaster emergency response, ecological environment assessment, agricultural management and the like. The method not only provides scientific basis for policy making and resource planning, but also promotes the remote sensing technology to develop automatically and intelligently, and is a core technology for realizing dynamic updating of geographic information and sustainable development decision support. The core of the change detection is to accurately capture the dynamic change of the earth surface. With the exponential growth of the scale of remote sensing data and the complexity of application scenes, the conventional change detection method has difficulty in coping with the dual challenges of high precision and high robustness. Especially under complex scenes such as illumination change, noise interference, dynamic evolution of ground features and the like, the traditional method often has the problems of high false alarm rate, boundary positioning misalignment and the like, and becomes a key bottleneck for limiting the practical application of the traditional method. Although the existing neural network model remarkably improves detection performance through end-to-end feature learning, two main challenges are faced, namely, firstly, the dimension perception capability is insufficient, the capturing capability of fine edge changes (such as road extension and vegetation boundary migration) is limited, key geometric information is easy to lose under strong noise interference, and secondly, structural continuity is lost, boundary fracture or discontinuous phenomenon often occurs in a generated change graph, and the topological relation between a land feature change area and a background is distorted. The root cause of the problems is that the prior knowledge of the space structure of the existing model is not utilized enough, and the cooperative modeling capability of the multi-scale characteristic and the topological relation of the ground object is lacking. Therefore, a new technical solution is needed to solve the technical problem of how to perform high-precision and high-robustness remote sensing image change detection. Disclosure of Invention The invention provides a remote sensing image change detection method and device based on edge contour guidance, which are used for solving the technical problem of how to perform high-precision and high-robustness remote sensing image change detection. In order to achieve the above object, the present invention provides a remote sensing image change detection method based on edge contour guidance, including: The method comprises the steps of obtaining a remote sensing map, obtaining a profile map by combining the remote sensing map with a Canny operator, wherein the remote sensing map comprises an early remote sensing map and a late remote sensing map which have preset time intervals, and the profile map comprises an early profile map and a late profile map; carrying out multi-time step denoising on the remote sensing image and the profile image, and carrying out preset multi-scale feature extraction by combining a denoising diffusion probability model to respectively obtain multi-time step feature images of the remote sensing image and the profile image under each scale; and respectively carrying out enhancement fusion on the multi-time-step feature images of the remote sensing image and the profile image under each scale, and sequentially fusing enhancement fusion results under each scale from small to large according to the scale to obtain a change detection result image. Preferably, the performing enhanced fusion on the multi-time-step feature map of the remote sensing map and the profile map under each scale comprises: Under the condition that all scales are the same, splicing and fusing the multi-time-step feature graphs of the remote sensing graph and the profile graph in the channel dimension to obtain early enhancement representation and late enhancement representation under all scales, respectively carrying out depth feature extraction and fusion on the early enhancement representation and the late enhancement representation through two groups of continuous convolution layers and ReLU activation functions to obtain early feature representation and late feature representation under all scales, and calculating absolute differences of the early feature representation and the late feature representation to obtain an in