CN-120013923-B - Heterogeneous remote sensing image building change detection method based on mutual distillation and differential evolution

CN120013923BCN 120013923 BCN120013923 BCN 120013923BCN-120013923-B

Abstract

The invention particularly relates to a construction change detection method of a heterogeneous remote sensing image based on mutual distillation and differential evolution, which comprises the steps of obtaining visible light and remote sensing images in the same region under different time, establishing a pseudo twin encoder change detection network for independently extracting heterogeneous image features, establishing a structural perception mutual distillation module, realizing cross-modal alignment of the heterogeneous features by carrying out knowledge mutual distillation on structural characterization of model difference robustness in the heterogeneous features, enhancing the feature extraction performance of an encoder, establishing a differential dynamic evolution extraction module, modeling mapping relations among the heterogeneous features as a dynamic evolution process in a time domain, accurately extracting characterization related to change based on modal differences of a constant region and distinguishing characteristics of the variation region differences in the dynamic evolution process, effectively enhancing the accuracy degree of prediction of the variation region, and helping to reduce omission.

Inventors

WANG QI
BAI HAICHEN
LI QIANG

Assignees

西北工业大学

Dates

Publication Date: 20260508
Application Date: 20250217

Claims (9)

1. The utility model provides a heterogeneous remote sensing image building change detection method based on mutual distillation and differential evolution, which is characterized in that the method comprises the following steps: Obtaining visible light and SAR remote sensing images in the same region at different times, marking the region where the building change occurs, establishing a heterogeneous remote sensing image change detection dataset, and dividing the dataset into a training set and a testing set; The method comprises the steps of constructing a pre-trained heterogeneous change detection network, wherein the heterogeneous change detection network comprises a pseudo-twin encoder and a decoder, two branches of the pseudo-twin encoder have the same structure but do not share parameters, the characteristics of a heterogeneous image are respectively extracted and the dimension is gradually reduced, each layer of the pseudo-twin encoder comprises a structured mutual distillation module for realizing cross-modal characteristic alignment and a differential dynamic evolution extraction module for differentially extracting relevant characterization of a real change region, the differential dynamic evolution extraction module inputs the extracted relevant differential characterization of the real change region into a layer with the corresponding dimension in the decoder in a jump connection mode and is used for guiding the decoder to generate a change region prediction mask with more precision and fine granularity, and the structured mutual distillation module specifically comprises the following processing steps: Respectively cutting and blocking the heterogeneous features output by the pseudo-twin encoder, respectively and independently performing Fourier transform and high-pass filtering on the segmented heterogeneous features in an image block, then performing Fourier inverse transform, splicing the image block back to the original input feature distribution, and outputting independent structural characterization of the heterogeneous features; Mapping each element of the structural representation into a node, mapping the position relation between different elements into edges connected between the nodes, constructing a grid map, and carrying out smoothing treatment on the structural representation through a two-layer graph convolution neural network; masking the smoothed structured representation by adopting a mutual attention mechanism of transposed matrix multiplication; Training the heterogeneous change detection network by using a training set, and testing the trained heterogeneous change detection network by using a testing set to obtain a heterogeneous image data building change detection result.
2. The method for detecting building changes of a heterologous remote sensing image based on mutual distillation and differential evolution according to claim 1, wherein the method further comprises data augmentation of a training set, wherein the data augmentation comprises cutting, overturning and shifting.
3. The method for detecting the building change of the heterologous remote sensing image based on mutual distillation and differential evolution according to claim 1, wherein knowledge mutual distillation is carried out on the masked structured representation by adopting a loss function combining weighting and weight increment factors.
4. The method for detecting the building change of the heterogeneous remote sensing image based on mutual distillation and differential evolution according to claim 1, wherein the differential dynamic evolution extraction module specifically comprises the following processing steps: based on the absolute value of the difference between the visible light characteristic and the SAR characteristic and the visible light characteristic, the characteristic difference and the SAR characteristic are spliced in sequence, and then convolution is carried out to obtain an intermediate transition characteristic from the visible light characteristic to the SAR characteristic; Recombining the visible light characteristic, the intermediate transition characteristic and the SAR characteristic according to the time sequence conversion angle to obtain a time domain evolution characteristic; capturing integral difference characterization of the time domain evolution features on the time domain by adopting 3-dimensional convolution, wherein the integral difference characterization comprises modal differences of unchanged areas and change differences of changed areas; carrying out time domain average pooling on time domain evolution features, and effectively distinguishing features which are displayed in the time domain pooling process based on modal differences and variation differences; Reconstructing a convolution kernel of the 3-dimensional convolution, and performing difference extraction on time domain pooling features by adopting the 3-dimensional convolution after the reconstruction of the convolution kernel to obtain a modal difference characterization; Subtracting the integral difference characterization from the modal difference characterization to obtain the difference characterization only related to the change region, and further refining and optimizing the difference characterization related to the change region by adopting 3-dimensional convolution.
5. The method for detecting the building change of the heterologous remote sensing image based on mutual distillation and differential evolution according to claim 1, wherein the training of the heterologous change detection network by using the training set specifically comprises the steps of selecting AdamW as an optimizer, and optimizing a network model by adopting structural mutual distillation loss and change loss.
6. The method for detecting the building change of the heterologous remote sensing image based on mutual distillation and differential evolution according to claim 5, wherein the optimizing the network model by adopting the structured mutual distillation loss and the change loss comprises the following steps: constructing MSE loss as a loss function for supervising the full interaction of the structural characterization knowledge of each layer of encoder, and fusing the mutual distillation loss corresponding to each layer of encoder in a weighted mode : Wherein, the And Representing the actual value and the predicted value, respectively, of the ith pixel, N being the total encoder layer number, Is the weight factor of each layer of knowledge distillation loss; constructing variation losses in combination with DICE losses and BCE losses The problem of class unbalance between a change region and a constant region in heterogeneous data is solved: Wherein, the Is a super parameter; and through the combination of the structural mutual distillation loss and the change loss, guiding the network model to accurately predict the change area.
7. A storage medium having stored thereon a computer program which, when executed by a processor, implements the method for detecting a construction change of a heterologous remote sensing image based on mutual distillation and differential evolution as claimed in any one of claims 1 to 6.
8. A computer program product comprising a computer program, characterized in that the computer program, when executed by a processor, implements the method for detecting a construction change of a heterologous remote sensing image based on mutual distillation and differential evolution according to any one of claims 1 to 6.
9. An electronic device, comprising: and a memory for storing executable instructions of the processor; wherein the processor is configured to perform the inter-distillation and differential evolution based heterogeneous remote sensing image construction change detection method of any one of claims 1 to 6 via execution of the executable instructions.

Description

Heterogeneous remote sensing image building change detection method based on mutual distillation and differential evolution Technical Field The invention relates to the technical field of change detection, in particular to a construction change detection method of a heterogeneous remote sensing image based on mutual distillation and differential evolution, which realizes accurate change region detection of a construction region in heterogeneous remote sensing image data of a Synthetic Aperture Radar (SAR) and a visible light image. Background The change detection has great significance in aspects of disaster assessment, city planning, environmental protection and the like, can detect the change condition of an object of interest, and can provide comprehensive star action references for planners or leaders. Most of the existing change detection methods utilize homologous data acquired by the same type of sensor to carry out change detection, and when the requirements on timeliness and reliability are extremely high in face of disaster assessment and the like, and tasks of the same mode data can not be acquired in a short time, the homologous change detection method is difficult to effectively cope with. Therefore, research on the heterogeneous change detection method is important to immediately detect the change of the data mode after the disaster occurs, so that accurate and reliable data support is provided for the evaluation of the disaster and the development of subsequent rescue tasks. It should be noted that the information disclosed in the above background section is only for enhancing understanding of the background of the invention and thus may include information that does not form the prior art that is already known to those of ordinary skill in the art. Disclosure of Invention The invention provides a construction change detection method of a heterogeneous remote sensing image based on mutual distillation and differential evolution, which is used for solving the defect of depth dependence of the existing change detection method on homologous data and realizing accurate detection on a construction change area of the heterogeneous remote sensing image data. Other features and advantages of the invention will be apparent from the following detailed description, or may be learned by the practice of the invention. According to a first aspect of the present invention, there is provided a method for detecting a building change of a heterogeneous remote sensing image based on mutual distillation and differential evolution, the method comprising: Obtaining visible light and SAR remote sensing images in the same region at different times, marking the region where the building change occurs, establishing a heterogeneous remote sensing image change detection dataset, and dividing the dataset into a training set and a testing set; Constructing a pre-trained heterogeneous change detection network, wherein the heterogeneous change detection network comprises a pseudo-twin encoder and a decoder, two branches of the pseudo-twin encoder have the same structure but do not share parameters, and the characteristics of a heterogeneous image are respectively extracted and the dimension is gradually reduced; the method comprises the steps that each layer of a pseudo-twin encoder comprises a structured mutual distillation module for realizing cross-modal feature alignment and a differential dynamic evolution extraction module for differentially extracting relevant characterization of a real change region, wherein the differential dynamic evolution extraction module inputs the extracted relevant differential characterization of the real change region into a layer with a corresponding size in a decoder in a jump connection mode and is used for guiding the decoder to generate a change region prediction mask with more precision and fine granularity; Training the heterogeneous change detection network by using a training set, and testing the trained heterogeneous change detection network by using a testing set to obtain a heterogeneous image data building change detection result. In some exemplary embodiments, the method further comprises data augmentation of the training set, the data augmentation comprising cropping, flipping, and shifting. In some exemplary embodiments, the structured mutual distillation module specifically comprises the following processing steps: Respectively cutting and blocking the heterogeneous features output by the pseudo-twin encoder, respectively and independently performing Fourier transform and high-pass filtering on the segmented heterogeneous features in an image block, then performing Fourier inverse transform, splicing the image block back to the original input feature distribution, and outputting independent structural characterization of the heterogeneous features; Mapping each element of the structural representation into a node, mapping the position relation between different elements in