Remote sensing image building change detection method based on matching optimization

Abstract

The invention provides a remote sensing image building change detection method based on matching optimization, and the method comprises the steps: 1, carrying out the high-precision building recognition of an image in a previous period and an image in a later period, building a graph model according to the image change intensity information and edge intensity features, and obtaining building change candidate regions in the previous period and the later period; 2, estimating a roof relationship by adopting a method based on motion statistical feature matching, and judging and correcting through a fast robust filtering strategy to improve the stability of a matching process; and step 3, establishing a full-connection conditional random field model, integrating the building change information and roof matching results of the candidate regions of the two time phases in the front and rear periods together for collaborative optimization, and eliminating the influence of position difference on building change detection. According to the method, the potential building change area is accurately captured, and the problem that high-resolution remote sensing image building change detection is difficult and incomplete under the condition of complex and diverse scenes is solved.

Description

Technical Field

[0001] The present application relates to the technical field of remote sensing image processing, and in particular to a remote sensing image building change detection method based on matching optimization. Background Art

[0002] Changes in built-up land are often seen as a key indicator of urban development. Urban expansion and population growth often lead to improper use of natural land and illegal construction, which can seriously damage natural ecosystems and disrupt ecological balance. These disturbances not only reduce environmental quality, but also threaten human habitats and health. Therefore, timely monitoring and management of built-up land is essential to protect natural resources, maintain ecological balance, and promote sustainable urban development.

[0003] As a basic work of natural resource survey and monitoring, building change detection based on remote sensing images is widely used in the fields of illegal building investigation and punishment, urban space planning, and surface cover monitoring, and has important application value. With the significant improvement of satellite imaging technology and resolution, the use of high-resolution dual-temporal satellite images for building change detection has become the preferred strategy. However, high-resolution remote sensing images are rich in detail information, complex ground object coverage, and images are often interfered by radiation differences. At the same time, for complex scenes, there are phenomena such as image point displacement caused by terrain undulations and differences in building positions between images caused by different projection differences, making building change detection still challenging. With the rapid development of satellite remote sensing technology, the massive growth of high-resolution remote sensing data has brought more possibilities for enriching the scale of surface observations and refining ground object change detection. At the same time, with the improvement of image spatial resolution, the spectral separability of ground objects decreases, increasing the uncertainty of change detection results. Although traditional remote sensing image change detection methods have achieved many results in medium- and low-resolution and high-spectral remote sensing images, it is often difficult to achieve ideal results when facing high-resolution remote sensing images.

[0004] In recent years, deep learning technology has rapidly emerged and flourished, and has been successfully applied to a variety of visual tasks, greatly improving the ability to quickly discover and identify change information, and providing a systematic and efficient new technical means for urban planning, construction, and operation management. However, the use of deep learning for change detection is still in the exploratory stage. The complexity and particularity of remote sensing data also pose great challenges to the application of deep learning technology. For example, high-rise buildings will cause serious projection differences during the orthorectification process, resulting in geometric offset problems for buildings at the same location in the two images. Directly comparing the features of the two images to detect changes usually results in extremely high false alarms. In addition, the existence of different types of sensors and radiation differences, as well as the complexity and diversity of objects, also make the current change detection algorithm far from actual needs. Summary of the invention

[0005] In view of this, the present application provides a remote sensing image building change detection method based on matching optimization to overcome the problems of difficulty in discovering and incomplete detection of building changes in complex and diverse scenes.

[0006] To achieve the above objectives, the technical solutions adopted in this application are as follows: A remote sensing image building change detection method based on matching optimization, comprising: Step 1: Perform high-precision building recognition on the previous and next images respectively through multi-task semantic segmentation of contour distance-aware connection network, obtain image change intensity information in an object-oriented manner, and construct a graph model based on the image change intensity information and edge intensity features to obtain candidate building change areas in the previous and next phases; Step 2: For the candidate areas of building changes, the roof relationship is estimated by using a method based on motion statistical feature matching, including feature detection based on phase information and directional histogram matching based on phase consistency, and correction is made through a fast and robust filtering strategy to improve the stability of the matching process; Step 3: Establish a fully connected conditional random field model to integrate the building change information and roof matching results of the candidate areas in the early and late phases for collaborative optimization to eliminate the impact of location differences on building change detection.

[0007] Furthermore, the step 3 further includes: using spatial overlay analysis to classify building changes into three types: "new addition", "demolition" and "change".

[0008] Furthermore, in step 1, the multi-task semantic segmentation of the contour distance-aware connection network is used to compare the previous and next two phases. The images are used for high-precision building recognition as follows: Firstly, a contour distance-aware multi-task network model is used to extract building semantic information. Based on feature encoding, the top-level features are deeply optimized through the strip-pooled global attention module to achieve feature perception of different receptive fields. Then, a residual module is introduced to upsample the fused deep and shallow features to enhance the nonlinear expression ability of the decoding module. Finally, the upsampled features are decomposed into three tasks: building segmentation, contour extraction, and distance estimation. The overall loss function is constructed to complete the prediction, thereby obtaining a more accurate and complete building.

[0009] Furthermore, in step 1, the image change intensity information is obtained in an object-oriented manner, and a graph model is constructed based on the image change intensity information and the edge intensity feature to obtain the candidate building change areas in the early and late phases: Step 1.1: Using an object-oriented approach, we extract the multispectral histogram and the joint histogram of oriented gradient-local binary pattern (HOG-LBP) to obtain the image spectral and texture structural features; in the multispectral image, we obtain the segmentation object histogram of each band and use and Representation Object Spectral and textural characteristics of: (6) (7) in, Indicates Objects on Band The grayscale histogram of is the number of image bands; and Represent the gradient direction and rotation invariant LRI histogram respectively; Step 1.2: Calculate the change confidence (CC) based on the cosine angle similarity measure; use the cosine angle similarity measure method to measure the distance difference between objects for spectral features and texture structure features respectively : (8) (9) and Respectively represent the image objects of the previous and next periods Features; Step 1.3: Distance by spectral feature Distance from texture structure features Get change intensity information: (10) Step 1.4: Obtain the image edge intensity features through gradient calculation, combine the acquired change intensity information and the buildings extracted by semantic segmentation, build a graph cut optimization model and use the maximum flow minimum cut theory to obtain the global optimal solution, and generate the candidate building change areas in the early and late phases respectively; The edge strength is expressed as ,in and Respectively represent adjacent and The pixel value of .

[0010] Furthermore, the step 2 comprises: Step 2.1: dilate the generated candidate building change area, and for the dilated candidate building change area, perform precise feature point detection based on phase consistency on the basis of extracting local frequency information through log-Gabor wavelet; Step 2.2: Matching based on phase consistency direction histogram, specifically including: Step 2.21: During the HOPC construction process, first select a template window of a certain size on the image; Step 2.22: Calculate the eigenvalue and eigenvector of phase consistency for each pixel in the window to obtain the feature information required by HOPC; Step 2.23: Divide the template window into n blocks, each block contains m cells, to form the basic architecture of HOPC; Step 2.24: Count the phase consistency direction histogram of each unit; Step 2.25: To eliminate the effects of illumination and contrast changes, normalize the phase consistency direction histogram of the units within each overlapping block; Step 2.26: Combine the histogram vectors in all blocks to form the HOPC descriptor; Step 2.3: Improve the stability of the matching process by using a fast and robust filtering strategy to determine the correction, including: Step 2.3.1: Use NCC as the similarity measure to match within the limited range of the detection point; (16) In the formula, and Represents the template image and The HOPC descriptor, and Respectively and The mean of Step 2.3.2: Exploitation After finding a match between candidate building areas with dual temporal changes, a fast and robust filtering strategy based on grid motion statistics is used to evaluate the number of matching points in the neighborhood of the feature points of the matching pair to be judged, distinguish correct and incorrect matches, and generate the final geometric consistency matching result.

[0011] Furthermore, the step 3 is specifically as follows: Step 3.1: Calculate the joint change probability required for the unary potential by using the building change candidate area and the matching roof points ; The energy function of the model is expressed as follows: (17) in, is the set of all pixels in the image. is a univariate potential function, which means that the label assigned to No. The cost of a node is as follows: (18) (19) In the formula, It is estimated based on the matching results. The size of the point change, that is, the probability of joint change, is the maximum distance set; Any point To the nearest matching point The distance Indicates matching point The connected area where and is the pixel in the candidate area of ​​building change in the previous and next two phases The value of Step 3.2: Calculate the joint edge strength of the two images before and after; is a binary potential function, which is used to calculate the label and The cost assigned to the two nodes is that in order to make the segmentation result have spatial continuity, the same label will be assigned to objects with close distance and similar features. Two Gaussian kernel functions are used here, as shown in the following formula: (20) in, is the weight of the two kernel functions, and the two kernel functions are defined as follows: (twenty one) (twenty two) (twenty three) In the formula, and are two kernel functions, The edge strength of the joint image is obtained by extracting the maximum color contrast between the two images. and express Two pixels in phase and Spectral value of and Represents the corresponding pixel coordinates, and the relationship between them is controlled by the parameter control, Used to control the influence of color similarity, Used to control the size of the spatial range. Useful for culling small isolated areas.

[0012] Step 3.3: Based on all the candidate building change areas in the previous and next images, the fully connected conditional random field is used for optimization to obtain the final building change detection result.

[0013] Compared with the prior art, the beneficial effects of this application are: 1. Based on semantic segmentation and object-oriented technology, the potential building change areas are accurately captured by combining high-precision building recognition and change significance, overcoming the difficulty of detecting building changes in high-resolution remote sensing images under complex and diverse scenes; 2. The motion statistical feature matching optimization module is integrated to explore the corresponding relationship between roofs in images, solve the interference of geometric displacement factors caused by projection differences, and improve the accuracy of building change detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 A flow chart of a remote sensing image building change detection method based on matching optimization is provided for this application; Figure 2 A schematic diagram of a remote sensing image building change detection method based on matching optimization for this application; Figure 3 (a) is the early image used for building change extraction, (b) is the semantic building labeling map of the early image, (c) is the late image used for building change extraction, and (d) is the semantic building labeling map of the late image; Figure 4 This is a multi-task network model diagram for contour distance perception in this application; Figure 5 (a) is the segmentation map of the joint image, and (b) is the change intensity information map of the joint image; Figure 6 (a) is the candidate area map of building changes generated in the early image, and (b) is the candidate area map of building changes generated in the later image; Figure 7(a) is the candidate area map of building changes with the minimum bounding rectangle frame in the early image, and (b) is the candidate area map of building changes with the minimum bounding rectangle frame in the later image; Figure 8 Based on The result graph of feature matching; Fig. 9 (a) is the buffer zone range map of the matching points in the early image, (b) is the buffer zone range map of the matching points in the late image, (c) is the joint change probability map of the two images, (d) is the joint edge strength information map of the two images, (e) is the candidate area map of all building changes in the two images, and (f) is the optimized building change detection result map. Fig.10 This is a schematic diagram of spatial overlay analysis. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0017] like Figure 1 and 2 As shown, a remote sensing image building change detection method based on matching optimization includes: Step 1: Perform high-precision building recognition on the previous and next images respectively through multi-task semantic segmentation of contour distance-aware connection network, obtain image change intensity information in an object-oriented manner, and construct a graph model based on the image change intensity information and edge intensity features to obtain candidate building change areas in the previous and next phases; The multi-task semantic segmentation method of the contour distance-aware connection network is used to analyze the previous and next images, such as Figure 3 As shown in (a) and (c), high-precision building recognition is performed. First, a contour distance-aware multi-task network model (such as Figure 4 As shown in the figure, the semantic information of the building is extracted. On the basis of feature encoding, the top-level features are deeply optimized through the strip pooling global attention module to realize the feature perception of different receptive fields. Then, the residual module is introduced to upsample the fused deep and shallow features to enhance the nonlinear expression ability of the decoding module. Finally, the upsampled features are decomposed into three tasks: building segmentation, contour extraction and distance estimation. The overall loss function is constructed to complete the prediction, and then a relatively accurate and complete building (Semantic Building, SBL) is obtained, as shown in the figure. Figure 3 As shown in (b) and (d).

[0018] Among them, the strip pooling global attention module is based on stacking multiple dilated convolution modules with different step sizes, and performs feature enhancement through strip pooling attention. The stacked convolution block is composed of dilated convolutions with step sizes of 1, 2, and 4. The combination of dilated convolutions realizes feature perception of different receptive fields. The maximum receptive field is 15*15 pixels, which basically covers the 16*16 pixels of the top-level features. Strip pooling is a global feature optimization method that obtains remote context information with minimal computation. It effectively expands the network's receptive field by combining horizontal and vertical pooling.

[0019] The residual module introduces a branch containing a 1×1 convolution kernel on the basis of the initial residual block, which maintains the receptive field range without increasing too many parameters. The main branch uses a 1×1 convolution kernel to reduce the dimension, and upsamples through a deconvolution with a convolution kernel of 3×3, and then enhances the channel expression of the upsampled features.

[0020] The loss function consists of segmentation mask loss, contour loss and distance loss. Segmentation mask and contour extraction are both binary classification tasks, using the sum of binary cross entropy loss and Dice loss; distance estimation is a regression task, using mean square error loss. The calculation is as follows:

[0021] (1) (2) (3) (4) (5) In the formula, , and They are segmentation mask loss, contour loss and distance loss respectively. Represents the binary classification task loss by binary cross entropy loss With Dice loss constitute, and They are The predicted and true values ​​of each pixel.

[0022] As a further implementation method, in step 1, the image change intensity information is obtained in an object-oriented manner, and a graph model is constructed based on the image change intensity information and the edge intensity feature to obtain the candidate building change areas in the early and late phases: Step 1.1: Using an object-oriented approach, we extract the multispectral histogram and the joint histogram of oriented gradient-local binary pattern (HOG-LBP) to obtain the image spectral and texture structural features; in the multispectral image, we obtain the segmentation object histogram of each band and use and Representation Object Spectral and textural characteristics of: (6) (7) in, Indicates Objects on Band The grayscale histogram of is the number of image bands; in order to reduce the number of colors, the color of each band is quantized, 256 colors are represented by 12 different values, and the spectral features are compressed to 36 dimensions; and Respectively represent the gradient direction and the local binary pattern histogram with rotation invariance; Before adopting the object-oriented method, the bands of the two phases of remote sensing images are separated and compared and analyzed respectively to obtain the maximum grayscale value of the corresponding band pixels. The generated bands are re-overlapped to generate a joint image. The SLIC (Simple Linear Iterative Clustering) algorithm is used to segment the joint image so that it has a one-to-one correspondence, such as Figure 5 As shown in (a).

[0023] Step 1.2: Calculate the change confidence (CC) based on the cosine angle similarity measure; use the cosine angle similarity measure method to measure the distance difference between objects for spectral features and texture structure features respectively : (8) (9) and Respectively represent the image objects of the previous and next periods Features; Step 1.3: Distance by spectral feature Distance from texture structure features Get change intensity information, such as Figure 5 As shown in (b): (10) Among them, the spectral characteristic distance Distance from texture structure features It can be obtained according to formulas (8) and (9).

[0024] Step 1.4: Obtain the image edge intensity features through gradient calculation, combine the acquired change intensity information and the buildings extracted by semantic segmentation, build a graph cut optimization model and use the maximum flow minimum cut theory to obtain the global optimal solution, and generate the candidate building change areas in the early and late phases respectively; In order to ensure the integrity of the extraction of potential changed building areas, the building change detection is modeled as two binary classification problems, in which the extraction of building change objects in each period of data is a binary classification problem, the changed buildings are the foreground, and the unchanged areas are the background. The image edge intensity features are obtained by gradient calculation, and the graph cut optimization model is constructed by combining the obtained change intensity information and the buildings extracted by semantic segmentation. The maximum flow minimum cut theory is used to obtain the global optimal solution, and the changed building proposals (CBP) of the two phases are generated respectively, such as Figure 6 As shown in (a) and (b).

[0025] The energy function is defined as follows: (11) (12) (13) in, Corresponding to the unchanged and changed marks, is the set of all pixels, data item Represents pixels The cost of being marked as foreground or background, Represents pixels The probability that a building changes can be obtained by Calculated. Smooth term Used to penalize neighboring pixels and discontinuities between to encourage spatial consistency, is the weight coefficient, It is the set of all adjacent pixel pairs in the image, usually a 4-neighborhood or an 8-neighborhood. It's pixels The pixel value of is the edge strength, Represents a pixel in SBL Tags, represents the spatial distance between adjacent pixels, Determined by the intensity of image noise.

[0026] (14) in Refers to the mean value of the entire image range, generates two candidate areas of building changes in the time phase, and obtains the corresponding minimum enclosing rectangle according to each area, such as Figure 7 As shown in (a) and (b).

[0027] Step 2: For the candidate areas of building changes, the roof relationship is estimated by using a method based on motion statistical feature matching, including feature detection based on phase information and directional histogram matching based on phase consistency, and correction is made through a fast and robust filtering strategy to improve the stability of the matching process; In the generated candidate areas of dual-phase changing buildings, many pseudo-changing buildings are misidentified due to the displacement of image points caused by terrain undulations and the influence of building position differences between images caused by different projection differences. Therefore, it is necessary to further determine the corresponding relationship between the roofs of the previous and next images to constrain the detection of building changes. The method based on motion statistical feature matching is used to estimate the roof relationship, which mainly includes feature detection based on phase information and feature matching based on the histogram of orientated phase congruency (HOPC) structure descriptor. Since the scale, height, outline shape and structure of buildings within a certain area (such as a community) are relatively similar, their projection differences on the image and the differences between images also have a consistent relationship, so they can be further judged and corrected through a fast and robust filtering strategy to improve the stability of the matching process.

[0028] As a further implementation, the step 2 comprises: Step 2.1: dilate the generated candidate building change area, and for the dilated candidate building change area, perform precise feature point detection based on phase consistency on the basis of extracting local frequency information through log-Gabor wavelet; The perception of human visual features is affected by the phase information in the frequency domain. At the same time, phase information has a certain resistance to interference from factors such as image scale, illumination and contrast. Feature extraction based on phase information is a method that conforms to visual characteristics. In the frequency domain, various image features (such as endpoints, corners and edge lines) will show the greatest phase consistency, which can be used to detect feature points in the image. Considering that the log-Gabor wavelet has a large bandwidth and can provide richer local frequency information, the analysis of phase consistency is more appropriate. At the same time, it has good locality in the spatial domain and can improve the accuracy of feature positioning. Therefore, feature points are accurately detected based on phase consistency on the basis of extracting local frequency information through log-Gabor wavelet. In order to ensure that all feature points are included as much as possible, the generated building change candidate area is appropriately expanded, and feature points are extracted in these areas using feature point detection based on phase consistency.

[0029] Step 2.2: Matching based on phase consistency direction histogram, specifically including: Step 2.21: During the HOPC construction process, first select a template window of a certain size on the image; Step 2.22: Calculate the eigenvalue and eigenvector of phase consistency for each pixel in the window to obtain the feature information required by HOPC; The HOPC descriptor is made illumination and contrast invariant by introducing phase consistency features. The eigenvalues ​​and feature directions of phase consistency are calculated at the same time, and the idea of ​​gradient directional histogram is used to construct a feature space vector representing the geometric structure of the image.

[0030] Step 2.23: Divide the template window into n blocks, each block contains m cells, to form the basic architecture of HOPC; Step 2.24: Count the phase consistency direction histogram of each unit; Taking into account the interference caused by the grayscale inversion of the image, when calculating the direction histogram within the statistical unit, the phase consistency feature direction needs to be transformed to between [0,180], and the calculation is as follows: (15) In the formula, represents the phase consistency characteristic direction, is the phase consistency characteristic direction after transformation.

[0031] Step 2.25: To eliminate the effects of illumination and contrast changes, normalize the phase consistency direction histogram of the units within each overlapping block; Step 2.26: Combine the histogram vectors in all blocks to form the HOPC descriptor; Step 2.3: Improve the stability of the matching process by using a fast and robust filtering strategy to determine the correction, including: Step 2.3.1: Use the Normalized Cross-Correlation (NCC) as a similarity measure to match within the limited range of the detection point; (16) In the formula, and Represents the template image and The HOPC descriptor, and Respectively and The mean of Specifically, on the detected feature points, take The template window, each block consists of units, each unit contains The HOPC feature descriptor is calculated by dividing the pixels into 9 directions, and NCC is used as the similarity measure to match within the limited range of the detection points. That is, it is assumed that two points that are far apart must not match.

[0032] Step 2.3.2: Exploitation After finding a match between the candidate building areas with dual temporal changes, the fast and robust filtering strategy based on grid motion statistics is used to evaluate the number of matching points contained in the neighboring area of ​​the matching feature point to be judged to distinguish correct and incorrect matches, and generate the final geometric consistency matching result, that is, the early matching points and the late matching points. The corresponding relationship connection line is as follows: Figure 8 shown.

[0033] In order to further improve the effectiveness and robustness of the method, the overlapping area and the number of matching points in the minimum circumscribed rectangle of the corresponding positions in the two periods are counted. In a specific implementation method, if the overlapping area exceeds 50% and the number of matching points is more than 20, it is considered that there is no change.

[0034] Step 3: Establish a fully connected conditional random field model to integrate the building change information and roof matching results of the candidate areas in the early and late phases for collaborative optimization to eliminate the impact of location differences on building change detection.

[0035] In order to provide high-quality building change information, a fully connected CRF (conditional random field) model is used to integrate the building change information and roof matching results of the two phase candidate areas for collaborative optimization, and transform change detection into a binary labeling problem, that is, , 0 represents unchanged objects, and 1 represents changed buildings.

[0036] As a further implementation, the step 3 is specifically: Step 3.1: Compare the building change candidate areas with the matching roof points (such as Fig. 9 (a) and (b)) to calculate the joint change probability required for the unary potential ; The energy function of the model is expressed as follows: (17) in, is the set of all pixels in the image. is a univariate potential function, which means that the label assigned to No. The cost of a node is as follows: (18) (19) In the formula, It is estimated based on the matching results. The size of the point change, that is, the probability of joint change, such as Fig. 9 As shown in (c), is the maximum distance set; Any point To the nearest matching point The distance Indicates matching point The connected area where and is the pixel in the candidate area of ​​the building that changes in the previous and next two phases The value of Step 3.2: Calculate the joint edge strength of the two images before and after; is a binary potential function, which is used to calculate the label and The cost assigned to the two nodes is that in order to make the segmentation result have spatial continuity, the same label will be assigned to objects with close distance and similar features. This paper uses two Gaussian kernel functions, as shown in the following formula: (20) in, is the weight of the two kernel functions, and the two kernel functions are defined as follows: (twenty one) (twenty two) (twenty three) In the formula, and are two kernel functions, is the joint edge strength, such as Fig. 9 As shown in (d), the maximum color contrast is obtained by extracting the maximum color contrast between the previous and next two images. and express Two pixels in phase and Spectral value of and Represents the corresponding pixel coordinates, and the relationship between them is controlled by the parameter control; Used to control the influence of color similarity, Used to control the size of the spatial range. Useful for culling small isolated areas.

[0037] Step 3.3: All candidate building change areas in the previous and next two images (such as Fig. 9 Based on the results shown in (e) in the figure, the fully connected conditional random field is used for optimization to obtain the final building change detection result, as shown in Fig. 9 As shown in (f).

[0038] As a further implementation method, step 3 further includes: using spatial overlay analysis to classify building changes into three types: "new addition", "demolition" and "change".

[0039] Based on the joint optimization method of fully connected conditional random fields, the final building change object can be obtained. By spatially overlaying it with the initial building change candidate areas before and after, the changed buildings can be further divided into three types: "newly added", "demolition" and "changed", as shown in Table 1 and Fig.10 shown.

[0040] Table 1 Determination of building change types

[0041] This application explores a remote sensing image building change detection method based on matching optimization. On the one hand, semantic segmentation is combined with object-oriented technology to calculate the significance of building changes, while incorporating edge strength information, and using a graph model to obtain the initial potential candidate areas of building changes in the previous and next periods, overcoming the problem of difficulty in finding and incompleteness in complex and diverse scenes; on the other hand, the corresponding relationship between the roofs of the previous and next images is mined through feature matching based on motion statistics, constraining building change detection, and integrating all information into a fully connected conditional random field model to eliminate the influence of position differences on building change detection and improve the accuracy and reliability of the results.

[0042] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A remote sensing image building change detection method based on matching optimization, characterized in that: include: Step 1: Perform high-precision building recognition on the previous and next images respectively through multi-task semantic segmentation of contour distance-aware connection network, obtain image change intensity information in an object-oriented manner, and construct a graph model based on the image change intensity information and edge intensity features to obtain candidate building change areas in the previous and next phases; Step 2: For the candidate areas of building changes, the roof relationship is estimated by using a method based on motion statistical feature matching, including feature detection based on phase information and directional histogram matching based on phase consistency, and correction is made through a fast and robust filtering strategy to improve the stability of the matching process; Step 3: Establish a fully connected conditional random field model to integrate the building change information and roof matching results of the candidate areas in the early and late phases for collaborative optimization to eliminate the impact of location differences on building change detection.

2. A remote sensing image building change detection method based on matching optimization as claimed in claim 1, characterized in that: The step 3 also includes: using spatial overlay analysis to classify building changes into three types: "new addition", "demolition" and "change".

3. A remote sensing image building change detection method based on matching optimization as claimed in claim 1 or 2, characterized in that: In step 1, high-precision building recognition is performed on the previous and next images respectively through multi-task semantic segmentation of the contour distance-aware connection network, specifically: Firstly, a contour distance-aware multi-task network model is used to extract building semantic information. Based on feature encoding, the top-level features are deeply optimized through a strip-pooled global attention module to achieve feature perception of different receptive fields. Then, a residual module is introduced to upsample the fused deep and shallow features to enhance the nonlinear expression ability of the decoding module; finally, the upsampled features are decomposed into three tasks: building segmentation, contour extraction, and distance estimation, and an overall loss function is constructed to complete the prediction, thereby obtaining a more accurate and complete building.

4. A remote sensing image building change detection method based on matching optimization as claimed in claim 3, characterized in that: In step 1, the image change intensity information is obtained in an object-oriented manner, and a graph model is constructed based on the image change intensity information and the edge intensity feature to obtain the candidate building change areas in the early and late phases: Step 1.1: Using an object-oriented approach, obtain the image spectrum and texture structure features by extracting the multispectral histogram and the gradient direction-local binary pattern joint histogram; in the multispectral image, obtain the segmentation object histogram of each band respectively, and use and Representation Object Spectral and textural characteristics of: (6) (7) in, Indicates Objects on Band The grayscale histogram of is the number of image bands; and Represent the gradient direction and rotation invariant LRI histogram respectively; Step 1.2: Calculate the change intensity information based on the cosine angle similarity measure; use the cosine angle similarity measure method to measure the distance difference between objects for spectral features and texture structure features respectively : (8） (9) and Respectively represent the image objects of the previous and next periods Features; Step 1.3: Distance by spectral feature Distance from texture structure features Get change intensity information: (10) Step 1.4: Obtain the image edge intensity features through gradient calculation, combine the acquired change intensity information and the buildings extracted by semantic segmentation, build a graph cut optimization model and use the maximum flow minimum cut theory to obtain the global optimal solution, and generate the candidate building change areas in the early and late phases respectively; The edge strength is expressed as ,in and Respectively represent adjacent and The pixel value of .

5. The remote sensing image building change detection method based on matching optimization as claimed in claim 4, characterized in that: The step 2 comprises: Step 2.1: dilate the generated candidate building change area, and for the dilated candidate building change area, perform precise feature point detection based on phase consistency on the basis of extracting local frequency information through log-Gabor wavelet; Step 2.2: Matching based on phase consistency direction histogram, specifically including: Step 2.21: During the HOPC construction process, first select a template window of a certain size on the image; Step 2.22: Calculate the eigenvalue and eigenvector of phase consistency for each pixel in the window to obtain the feature information required by HOPC; Step 2.23: Divide the template window into n blocks, each block contains m units, to form the basic architecture of HOPC; Step 2.24: Count the phase consistency direction histogram of each unit; Step 2.25: To eliminate the effects of illumination and contrast changes, normalize the phase consistency direction histogram of the units within each overlapping block; Step 2.26: Combine the histogram vectors in all blocks to form the HOPC descriptor; Step 2.3: Improve the stability of the matching process by using a fast and robust filtering strategy to determine the correction, including: Step 2.3.1: Use NCC as the similarity measure to match within the limited range of the detection point; （16） In the formula, and Represents the template image and The HOPC descriptor, and Respectively and The mean of Step 2.3.2: Exploitation After finding matches between candidate building areas with dual temporal changes, a fast and robust filtering strategy based on grid motion statistics is used to evaluate the number of matching points contained in the neighborhood of the feature points of the matching pair to be judged, distinguish correct and incorrect matches, and generate the final geometric consistency matching result.

6. A remote sensing image building change detection method based on matching optimization as claimed in claim 5, characterized in that: The step 3 is specifically as follows: Step 3.1: Calculate the joint change probability required for the unary potential by using the building change candidate area and the matching roof points ; The energy function of the model is expressed as follows: (17) in, is the set of all pixels in the image. is a univariate potential function, which means that the label assigned to No. The cost of a node is as follows: (18) (19) In the formula, It is estimated based on the matching results. The size of the point change, that is, the probability of joint change, is the maximum distance set; Any point To the nearest matching point The distance Indicates matching point The connected area where and is the pixel in the candidate area of ​​the building that changes in the previous and next two phases The value of Step 3.2: Calculate the joint edge strength of the two images before and after; is a binary potential function, which is used to calculate the label and The cost assigned to the two nodes is that in order to make the segmentation result have spatial continuity, the same label will be assigned to objects with close distance and similar features. This paper uses two Gaussian kernel functions, as shown in the following formula: (20) in, are the weights of the two kernel functions, which are defined as follows: (21) (22) (23) In the formula, and are two kernel functions, To combine edge strength, the maximum color contrast is extracted from the previous and next images. and express Two pixels in phase and Spectral value of and Represents the corresponding pixel coordinates, and the relationship between them is controlled by the parameter control, Used to control the influence of color similarity, Used to control the size of the spatial range. Used to remove small isolated areas; Step 3.3: Based on all the candidate building change areas in the previous and next images, the fully connected conditional random field is used for optimization to obtain the final building change detection result.

Images