A remote sensing image building change detection method based on matching optimization

Abstract

The application provides a remote sensing image building change detection method based on matching optimization, comprising the following steps: step 1: high-precision building recognition is performed on two images before and after, a graph model is constructed according to image change intensity information and edge intensity characteristics, and a building change candidate area of two time phases before and after is obtained; step 2: a method based on motion statistical feature matching is used to estimate the roof relationship, and a fast robust filtering strategy is used for judgment and correction, so that the stability of the matching process is improved; step 3: a fully connected conditional random field model is established, building change information of the candidate area of two time phases before and after and roof matching results are integrated together for cooperative optimization, and the influence of position difference on building change detection is eliminated. The application can accurately capture potential building change areas, and overcomes the problem that building change detection of high-resolution remote sensing images is difficult to find and complete under complex and diverse scenes.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and more specifically, to a method for detecting building changes in remote sensing images based on matching optimization. Background Technology

[0002] Changes in built-up land use are often considered a key indicator of urban development. Urban expansion and population growth frequently lead to the misuse of natural land and illegal construction, which severely damages natural ecosystems and disrupts 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 use are crucial for protecting natural resources, maintaining ecological balance, and promoting sustainable urban development.

[0003] Building change detection using remote sensing imagery is a fundamental task in natural resource surveys and monitoring, widely applied in areas such as illegal construction investigation, urban spatial planning, and land cover monitoring, and has significant application value. With the significant improvement in satellite imaging technology and resolution, using high-resolution dual-temporal satellite imagery for building change detection has become the preferred strategy. However, high-resolution remote sensing images are rich in detail, have complex land cover, and are often subject to interference from radiometric differences between images. Furthermore, for complex scenes, phenomena such as image point displacement caused by terrain undulations and differences in building positions between images due to projection differences make building change detection still challenging. With the rapid development of satellite remote sensing technology, the massive increase in high-resolution remote sensing data has brought more possibilities for enriching the scale of surface observation and refining land cover change detection. At the same time, as image spatial resolution increases, the spectral separability of land cover decreases, increasing the uncertainty of change detection results. Although traditional remote sensing image change detection methods have achieved many results on low-to-medium resolution and hyperspectral remote sensing images, they often struggle to achieve ideal results when dealing with high-resolution remote sensing images.

[0004] In recent years, deep learning technology has rapidly emerged and flourished, and has been successfully applied to various visual tasks, greatly improving the ability to quickly detect and identify changes, providing a systematic and efficient new technological 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 unique characteristics of remote sensing data also bring great challenges to the application of deep learning technology. For example, high-rise buildings can cause severe projection differences during orthorectification, resulting in geometric offsets of buildings at the same location in two different images. Directly comparing features from two images to detect changes often produces a very high false alarm rate. In addition, the existence of different types of sensors and radiometric differences, as well as the complexity and diversity of ground features, mean that current change detection algorithms are still far from meeting practical needs. Summary of the Invention

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

[0006] To achieve the above objectives, the technical solution adopted in this application is as follows:

[0007] A method for detecting building changes in remote sensing images based on matching optimization, comprising:

[0008] Step 1: High-precision building identification is performed on the two phases of images by multi-task semantic segmentation through contour distance perception connection network. The image change intensity information is obtained by combining object-oriented method. Based on the image change intensity information and edge intensity features, a graph model is constructed to obtain the candidate areas of building changes in the two phases.

[0009] Step 2: For candidate areas of building changes, the roof relationship is estimated using a method based on motion statistical feature matching, including feature detection based on phase information and histogram matching based on phase consistency direction. A fast and robust filtering strategy is used to judge and correct the results, thereby improving the stability of the matching process.

[0010] 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 time phases for collaborative optimization, thereby eliminating the impact of location differences on building change detection.

[0011] Furthermore, step 3 includes: using spatial overlay analysis to classify building changes into three types: "new additions", "demolitions", and "alterations".

[0012] Furthermore, in step 1, multi-task semantic segmentation of the two phases is performed using a contour distance-aware connection network.

[0013] The high-precision building identification of the images is specifically performed as follows:

[0014] First, a contour distance-aware multi-task network model is used to extract semantic information of buildings. Based on feature encoding, a global attention module with strip pooling is used to perform deep optimization of the top-level features 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 expressive power of the decoding module. Finally, the upsampled features are decomposed into three tasks: building segmentation, contour extraction, and distance estimation. An overall loss function is constructed to complete the prediction, thereby obtaining a more accurate and complete building information.

[0015] Furthermore, in step 1, image change intensity information is obtained using an object-oriented approach, and a graph model is constructed based on the image change intensity information and edge intensity features to obtain candidate areas for building changes in the two time phases:

[0016] Step 1.1: Using an object-oriented approach, the spectral and textural features of the image are obtained by extracting multispectral histograms and Histogram of Oriented Gradient - Local Binary Pattern (HOG-LBP). In the multispectral image, the segmentation object histograms for each band are calculated, and then... and Representation Object Spectral and textural structural features:

[0017] (6)

[0018] (7)

[0019] in, Indicates the first Objects in the band The grayscale histogram, It is the number of image bands; and These represent the gradient direction and rotation-invariant LRI histograms, respectively.

[0020] Step 1.2: Calculate Change Confidence (CC) based on cosine similarity measure; use cosine similarity to measure the distance difference between objects for both spectral features and texture structure features. :

[0021] (8)

[0022] (9)

[0023] and Representing the objects in the two phases of images respectively Features;

[0024] Step 1.3: Spectral characteristic distance Distance to texture structure features Obtain information on the intensity of change:

[0025] (10)

[0026] Step 1.4: Obtain image edge intensity features through gradient calculation, combine the obtained change intensity information and the buildings extracted by semantic segmentation, construct a graph cut optimization model and use the maximum flow minimum cut theory to obtain the global optimal solution, and generate candidate areas for building changes in the early and late phases respectively.

[0027] The edge strength is expressed as ,in and Representing adjacent and The pixel value.

[0028] Furthermore, step 2 includes:

[0029] Step 2.1: Dilate the generated candidate regions of building changes. For the dilated candidate regions of building changes, perform precise feature point detection based on phase consistency after extracting local frequency information through log-Gabor wavelets.

[0030] Step 2.2: Histogram matching based on phase consistency direction, specifically including:

[0031] Step 2.21: During the HOPC construction process, first select a template window of a certain size on the image;

[0032] Step 2.22: Calculate the phase consistency feature value and feature vector for each pixel within the window to obtain the feature information required by HOPC;

[0033] Step 2.23: Divide the template window into n blocks, each block containing m cells, to form the basic architecture of HOPC;

[0034] Step 2.24: Calculate the phase consistency direction histogram for each unit;

[0035] Step 2.25: To eliminate the effects of illumination and contrast variations, normalize the phase consistency orientation histogram of each cell within each overlapping block;

[0036] Step 2.26: Combine the histogram vectors within all blocks to form the HOPC descriptor;

[0037] Step 2.3: Improve the stability of the matching process by determining the correction through a fast and robust filtering strategy, specifically including:

[0038] Step 2.3.1: Use NCC as the similarity measure to perform matching within the limited range of detection points;

[0039] (16)

[0040] In the formula, and Representing template images and HOPC descriptor, and They represent and The mean;

[0041] Step 2.3.2: Utilize After finding a match between candidate building regions with dual temporal variations, a fast and robust filtering strategy based on grid motion statistics is used to evaluate the number of matching pairs contained in the neighboring regions of the feature points of the matching pair to be judged, distinguishing between correct and incorrect matches, and generating the final geometrically consistent matching result.

[0042] Furthermore, step 3 specifically includes:

[0043] Step 3.1: Calculate the joint change probability required for the univariate potential by using the candidate areas of building change and the matched roof points. ;

[0044] The energy function of the model is expressed as follows:

[0045] (17)

[0046] in, It is the set of all pixels in the image. It is a univariate potential function, representing the label assigned to No. The cost of each node is given by the following formula:

[0047] (18)

[0048] (19)

[0049] In the formula, It is estimated based on the matching results. The magnitude of the change at a point, i.e., the joint probability of change. The maximum distance is set; It is any point To the nearest matching point distance, Indicates matching point The connected region it is located in and Pixels in the candidate region for building changes in two consecutive time phases The value;

[0050] Step 3.2: Calculate the joint edge intensity of the two images;

[0051] It is a binary potential function used to calculate the label and To ensure spatial continuity in the segmentation results, the costs assigned to the two nodes are determined separately. To further enhance the segmentation, objects that are close to each other and have similar features are assigned the same label. Two Gaussian kernel functions are used here, as shown in the following equation:

[0052] (20)

[0053] in, The weights are the values ​​for two kernel functions, which are defined as follows:

[0054] (twenty one)

[0055] (twenty two)

[0056] (twenty three)

[0057] In the formula, and There are two kernel functions. The edge intensity of the joint image is obtained by extracting the maximum color contrast from the two images. and express Two pixels in the time phase and spectral values; and These represent the corresponding pixel coordinates, and the relationship between them is determined by parameters. Control Used to control the effect of color similarity Used to control the size of the spatial range. Used to eliminate smaller, isolated areas.

[0058] Step 3.3: Based on all candidate areas of building changes in the two phases of images, optimize using a fully connected conditional random field to obtain the final building change detection results.

[0059] Compared with the prior art, the beneficial effects of this application are:

[0060] 1. Based on semantic segmentation and object-oriented technology, this method accurately captures potential building change areas by combining high-precision building identification and change saliency, overcoming the challenge of difficult and incomplete detection of building changes in high-resolution remote sensing images under complex and diverse scenarios;

[0061] 2. By integrating a motion statistics feature matching optimization module, the corresponding relationship between roofs in images is explored, which solves the interference of geometric displacement factors caused by projection difference and improves the accuracy of building change detection. Attached Figure Description

[0062] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a flowchart of a remote sensing image building change detection method based on matching optimization according to this application;

[0064] Figure 2 This is a schematic diagram illustrating the principle of a remote sensing image building change detection method based on matching optimization according to this application.

[0065] Figure 3 (a) is the pre-image used for building change extraction, (b) is the semantic building label map of the pre-image, (c) is the post-image used for building change extraction, and (d) is the semantic building label map of the post-image.

[0066] Figure 4 This is a diagram of the multi-task network model for contour distance perception in this application;

[0067] Figure 5 (a) is a segmentation map of the joint image, and (b) is a change intensity information map of the joint image;

[0068] Figure 6 (a) shows the generated candidate area of ​​building changes from the previous image, and (b) shows the generated candidate area of ​​building changes from the later image.

[0069] Figure 7 (a) shows the candidate areas of building changes with minimum bounding rectangles in the earlier image, and (b) shows the candidate areas of building changes with minimum bounding rectangles in the later image.

[0070] Figure 8 According to The result of feature matching;

[0071] Figure 9(a) is the buffer zone map of matching points in the previous image, (b) is the buffer zone map of matching points in the later image, (c) is the probability map of the combined change between the two images, (d) is the edge intensity information map of the combined image, (e) is the candidate map of all building changes in the two images, and (f) is the optimized building change detection result map.

[0072] Figure 10 This is a schematic diagram of spatial overlay analysis. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of this application, but not all embodiments.

[0074] like Figure 1 and 2 As shown, a method for detecting building changes in remote sensing images based on matching optimization includes:

[0075] Step 1: High-precision building identification is performed on the two phases of images by multi-task semantic segmentation through contour distance perception connection network. The image change intensity information is obtained by combining object-oriented method. Based on the image change intensity information and edge intensity features, a graph model is constructed to obtain the candidate areas of building changes in the two phases.

[0076] A multi-task semantic segmentation method using contour distance-aware connected networks is employed to segment two consecutive 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...) is used. Figure 4 As shown, semantic information of buildings is extracted. Based on feature encoding, a global attention module with strip pooling is used to deeply optimize the top-level features, enabling feature perception of different receptive fields. Then, a residual module is introduced to upsample the fused deep and shallow features, enhancing the nonlinear expressive power of the decoding module. Finally, the upsampled features are decomposed into three tasks: building segmentation, contour extraction, and distance estimation. An overall loss function is constructed to complete the prediction, thus obtaining a relatively accurate and complete building (Semantic Building, SBL) result. Figure 3 As shown in (b) and (d).

[0077] The strip pooling global attention module is based on stacked multiple dilated convolutional modules with different strides, and enhances features through strip pooling attention. The stacked convolutional blocks consist of dilated convolutions with strides of 1, 2, and 4. The combination of dilated convolutions enables feature perception across different receptive fields, with a maximum receptive field of 15*15 pixels, essentially covering the 16*16 pixels of the top-level features. Strip pooling is a global feature optimization method that acquires long-range contextual information with minimal computation. By combining horizontal and vertical pooling, it effectively expands the network's receptive field.

[0078] The residual module introduces a branch containing a 1×1 convolution kernel on the basis of the initial residual block. While maintaining the receptive field range, it does not increase the number of parameters too much. The main branch uses a 1×1 convolution kernel to reduce the dimensionality, and then performs upsampling through a 3×3 deconvolution kernel. Finally, the upsampled features are enhanced to improve the channel representation.

[0079] The loss function consists of segmentation mask loss, contour loss, and distance loss. Segmentation masking 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 squared error loss. The calculations are as follows:

[0080] (1)

[0081] (2)

[0082] (3)

[0083] (4)

[0084] (5)

[0085] In the formula, , and These are segmentation mask loss, contour loss, and distance loss, respectively. The loss for a binary classification task is represented by the binary cross-entropy loss. With Dice loss constitute, and They are the first The predicted and actual values ​​for each pixel.

[0086] As a further implementation, step 1, which combines object-oriented methods to obtain image change intensity information and constructs a graph model based on the image change intensity information and edge intensity features to obtain candidate areas for building changes in the two time phases, specifically involves:

[0087] Step 1.1: Using an object-oriented approach, the spectral and textural features of the image are obtained by extracting multispectral histograms and Histogram of Oriented Gradient - Local Binary Pattern (HOG-LBP). In the multispectral image, the segmentation object histograms for each band are calculated, and then... and Representation Object Spectral and textural structural features:

[0088] (6)

[0089] (7)

[0090] in, Indicates the first Objects in the band The grayscale histogram, This refers to the number of image bands; to reduce the number of colors, the color of each band is quantized, using 12 different values ​​to represent 256 colors, thus compressing the spectral features to 36 dimensions; and These represent the gradient direction and the local binary pattern histogram with rotation invariance, respectively.

[0091] Before adopting an object-oriented approach, each band of the two sets of remote sensing images was separated, compared and analyzed separately, and the maximum gray value of the corresponding band pixels was obtained. The generated bands were then re-overlaid to produce a joint image. The SLIC (Simple Linear Iterative Clustering) algorithm was used to segment the joint image, ensuring a one-to-one correspondence. Figure 5 As shown in (a).

[0092] Step 1.2: Calculate Change Confidence (CC) based on cosine similarity measure; use cosine similarity to measure the distance difference between objects for both spectral features and texture structure features. :

[0093] (8)

[0094] (9)

[0095] and Representing the objects in the two phases of images respectively Features;

[0096] Step 1.3: Spectral characteristic distance Distance to texture structure features Obtain information on the intensity of change, such as Figure 5 As shown in (b):

[0097] (10)

[0098] Among them, spectral feature distance Distance to texture structure features It can be obtained from formulas (8) and (9).

[0099] Step 1.4: Obtain image edge intensity features through gradient calculation, combine the obtained change intensity information and the buildings extracted by semantic segmentation, construct a graph cut optimization model and use the maximum flow minimum cut theory to obtain the global optimal solution, and generate candidate areas for building changes in the early and late phases respectively.

[0100] To ensure the completeness of potential changed building region extraction, building change detection is modeled as two binary classification problems. The extraction of changed building objects for each data period is a binary classification problem, with changed buildings as foreground and unchanged areas as background. Gradient calculation is used to obtain image edge intensity features. Combined with the acquired change intensity information and semantically segmented buildings, a graph cut optimization model is constructed, and the global optimal solution is obtained using the maximum flow minimum cut theory. This generates changed building proposals (CBPs) for two different time periods, as follows: Figure 6 As shown in (a) and (b).

[0101] The energy function is defined as follows:

[0102] (11)

[0103] (12)

[0104] (13)

[0105] in, Corresponding to unchanged and changed markers, It is a collection of all pixels, data item Represents pixels The cost of being marked as foreground or background Represents pixels The probability of a building changing can be determined by... Calculated. Smoothing term. Used to penalize adjacent pixels and Discontinuities between them encourage spatial consistency. These are weighting coefficients. It is the set of all adjacent pixel pairs in an image, usually taken as a 4-neighborhood or 8-neighborhood. It is a pixel Pixel values, This is the edge strength. Represents pixels in SBL The tag, Indicates the spatial distance between adjacent pixels. It is determined by the intensity of image noise.

[0106] (14)

[0107] in Using the mean value of the entire map area, two candidate regions for building changes across different time phases are generated. Simultaneously, the minimum bounding rectangle corresponding to each region is obtained, such as... Figure 7 As shown in (a) and (b).

[0108] Step 2: For candidate areas of building changes, the roof relationship is estimated using a method based on motion statistical feature matching, including feature detection based on phase information and histogram matching based on phase consistency direction. A fast and robust filtering strategy is used to judge and correct the results, thereby improving the stability of the matching process.

[0109] In the generated candidate areas for buildings with dual temporal changes, many falsely identified buildings are incorrectly identified due to the influence of image point displacement caused by terrain undulations and the differences in building positions between images caused by different projection differences. Therefore, it is necessary to further determine the correspondence between roofs in the two consecutive images to constrain the detection of building changes. A method based on motion statistical feature matching is used to estimate roof relationships, mainly including feature detection based on phase information and feature matching based on the histogram of oriented phase congruency (HOPC) structural descriptor. Since buildings within a certain area (such as a small community) have similar scale, height, outline shape, and structure, their projection differences on the images and the differences between images also have a consistent relationship. Therefore, a fast and robust filtering strategy can be used to judge and correct this, improving the stability of the matching process.

[0110] As a further implementation, step 2 includes:

[0111] Step 2.1: Dilate the generated candidate regions of building changes. For the dilated candidate regions of building changes, perform precise feature point detection based on phase consistency after extracting local frequency information through log-Gabor wavelets.

[0112] Human visual perception is influenced by phase information in the frequency domain. This phase information is also relatively resistant to interference from factors such as image scale, illumination, and contrast. Therefore, feature extraction based on phase information is a method consistent with visual characteristics. In the frequency domain, various image features (such as endpoints, corners, and edge lines) exhibit maximum phase consistency, which can be used to detect feature points in the image. Considering that the log-Gabor wavelet has a large bandwidth and provides richer local frequency information, it is suitable for phase consistency analysis. Furthermore, its good locality in the spatial domain can improve feature localization accuracy. Therefore, based on local frequency information extraction using the log-Gabor wavelet, feature points are accurately detected using phase consistency. To ensure that as many feature points as possible are included, the generated candidate regions for building changes are appropriately expanded, and feature points are extracted from these regions using phase consistency-based feature point detection.

[0113] Step 2.2: Histogram matching based on phase consistency direction, specifically including:

[0114] Step 2.21: During the HOPC construction process, first select a template window of a certain size on the image;

[0115] Step 2.22: Calculate the phase consistency feature value and feature vector for each pixel within the window to obtain the feature information required by HOPC;

[0116] The HOPC descriptor is made invariant to illumination and contrast by introducing phase consistency features. It also calculates the eigenvalues ​​and eigendirections of phase consistency and constructs a feature space vector representing the geometric structure of the image by drawing on the idea of ​​gradient direction histogram.

[0117] Step 2.23: Divide the template window into n blocks, each block containing m cells, to form the basic architecture of HOPC;

[0118] Step 2.24: Calculate the phase consistency direction histogram for each unit;

[0119] Considering the interference caused by image grayscale inversion, when creating the orientation histogram within the statistical unit, the phase consistency feature orientation needs to be transformed to the range [0, 180]. The calculation is as follows:

[0120] (15)

[0121] In the formula, Indicates the direction of phase consistency characteristics. The direction of the phase consistency feature after transformation.

[0122] Step 2.25: To eliminate the effects of illumination and contrast variations, normalize the phase consistency orientation histogram of each cell within each overlapping block;

[0123] Step 2.26: Combine the histogram vectors within all blocks to form the HOPC descriptor;

[0124] Step 2.3: Improve the stability of the matching process by determining the correction through a fast and robust filtering strategy, specifically including:

[0125] Step 2.3.1: Use the normalized cross-correlation (NCC) as a similarity measure to perform matching within the limited range of detection points;

[0126] (16)

[0127] In the formula, and Representing template images and HOPC descriptor, and They represent and The mean;

[0128] Specifically, at the detected feature points, take The template window, each block is composed of It consists of 10 units, each containing 10 units. The system calculates HOPC feature descriptors by dividing the data into 9 pixels and using NCC as a similarity measure to perform matching within the detection point limit. This means that based on prior knowledge, two points that are far apart are definitely not matched.

[0129] Step 2.3.2: Utilize After finding matches between candidate building regions exhibiting dual temporal variations, a fast and robust filtering strategy based on grid motion statistics is employed to evaluate the number of matching pairs within the neighboring regions of the feature points of the match to be judged, distinguishing between correct and incorrect matches. This generates the final geometrically consistent matching result, i.e., the early and later matching points, with the corresponding connections shown as follows: Figure 8 As shown.

[0130] To further improve the effectiveness and robustness of the method, the overlapping area and the number of matching points in the minimum bounding rectangles of the corresponding positions in the two periods are statistically analyzed. In a specific implementation, if the overlapping area exceeds 50% and the number of matching points is more than 20, it is considered that there has been no change.

[0131] 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 time phases for collaborative optimization, thereby eliminating the impact of location differences on building change detection.

[0132] To provide high-quality building change information, a fully connected CRF (conditional random field) model is adopted. This model integrates building change information from two temporal candidate regions with roof matching results for collaborative optimization, transforming change detection into a binary labeling problem. 0 represents an unchanged object, and 1 represents a changed building.

[0133] As a further implementation method, step 3 specifically includes:

[0134] Step 3.1: Match candidate areas of building changes with rooftop points (e.g., Figure 9 (as shown in (a) and (b)) the joint change probability required to calculate the univariate potential ;

[0135] The energy function of the model is expressed as follows:

[0136] (17)

[0137] in, It is the set of all pixels in the image. It is a univariate potential function, representing the label assigned to No. The cost of each node is given by the following formula:

[0138] (18)

[0139] (19)

[0140] In the formula, It is estimated based on the matching results. The magnitude of the point change, i.e., the joint probability of change, such as Figure 9 As shown in (c), The maximum distance is set; It is any point To the nearest matching point distance, Indicates matching point The connected region it is located in and For the changes in the candidate area of ​​the building between the two time phases The value;

[0141] Step 3.2: Calculate the joint edge intensity of the two images;

[0142] It is a binary potential function used to calculate the label and To ensure spatial continuity in the segmentation results, the costs assigned to the two nodes are determined separately. To further enhance the segmentation, objects that are close together and have similar features are assigned the same label. This paper employs two Gaussian kernel functions, as shown in the following equation:

[0143] (20)

[0144] in, The weights are the values ​​for two kernel functions, which are defined as follows:

[0145] (twenty one)

[0146] (twenty two)

[0147] (twenty three)

[0148] In the formula, and There are two kernel functions. For joint edge strength, such as Figure 9 As shown in (d), the image is obtained by extracting the maximum color contrast from the two images. and express Two pixels in the time phase and spectral values; and These represent the corresponding pixel coordinates, and the relationship between them is determined by parameters. Control; Used to control the effect of color similarity Used to control the size of the spatial range. Used to eliminate smaller, isolated areas.

[0149] Step 3.3: Identify all candidate areas for building changes in the two imagery periods (e.g., ...). Figure 9 Based on the results shown in (e), a fully connected conditional random field is used for optimization to obtain the final building change detection results, as shown in (e). Figure 9 As shown in (f).

[0150] As a further implementation, step 3 is followed by: using spatial overlay analysis to classify building changes into three types: “new additions,” “demolitions,” and “modifications.”

[0151] The fully connected conditional random field joint optimization method can obtain the final building change objects. By spatially overlaying these objects with the initial building change candidate areas from both periods, the changed buildings can be further classified into three types: "newly added," "demolished," and "altered," as shown in Table 1. Figure 10 As shown.

[0152] Table 1 Determination of Building Change Types

[0153]

[0154] This application explores a method for detecting building changes in remote sensing images based on matching optimization. On one hand, it combines semantic segmentation with object-oriented techniques to calculate the saliency of building changes, while incorporating edge intensity information. A graph model is used to obtain initial potential candidate regions for building changes in both preceding and subsequent images, overcoming the problem of incomplete or difficult detection in complex and diverse scenes. On the other hand, it mines the correspondence between rooftops in the two images based on motion statistics through feature matching, constraining the building change detection. All information is integrated into a fully connected conditional random field model to eliminate the influence of location differences on building change detection, improving the accuracy and reliability of the results.

[0155] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting building changes in remote sensing images based on matching optimization, characterized in that, include: Step 1: High-precision building identification is performed on the two phases of images by multi-task semantic segmentation through contour distance perception connection network. The image change intensity information is obtained by combining object-oriented method. Based on the image change intensity information and edge intensity features, a graph model is constructed to obtain the candidate areas of building changes in the two phases. Step 2: For candidate areas of building changes, the roof relationship is estimated using a method based on motion statistical feature matching, including feature detection based on phase information and histogram matching based on phase consistency direction. A fast and robust filtering strategy is used to judge and correct the results, thereby improving 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 time phases for collaborative optimization, thereby eliminating the impact of location differences on building change detection; In step 1, the high-precision building identification of the two images in the previous and subsequent phases is achieved through multi-task semantic segmentation of the contour distance perception network: First, the semantic information of the buildings is extracted using a multi-task network model with contour distance perception. Based on feature encoding, the top-level features are deeply optimized through a global attention module with strip pooling to achieve feature perception of different receptive fields. Then, a residual module is introduced to upsample the fused deep and shallow features, enhancing the nonlinear expressive power of the decoding module. Finally, the upsampled features are decomposed into three tasks: building segmentation, contour extraction, and distance estimation. An overall loss function is constructed to complete the prediction, thereby obtaining a more accurate and complete building.

2. The method for detecting building changes in remote sensing images based on matching optimization as described in claim 1, characterized in that, Step 3 is followed by: using spatial overlay analysis to classify building changes into three types: "new additions", "demolitions", and "alterations".

3. The method for detecting building changes in remote sensing images based on matching optimization as described in claim 2, characterized in that, In step 1, image change intensity information is obtained using an object-oriented approach, and a graph model is constructed based on this image change intensity information and edge intensity features to obtain candidate areas for building changes in the two time phases: Step 1.1: Using an object-oriented approach, the spectral and texture structure features of the image are obtained by extracting multispectral histograms and gradient orientation-local binary mode joint histograms; in the multispectral image, the segmentation object histograms for each band are calculated, and then... and Representation Object Spectral and textural structural features: (6) (7) in, Indicates the first Objects in the band The grayscale histogram, It is the number of image bands; and These represent the gradient direction and rotation-invariant LRI histograms, respectively. Step 1.2: Calculate the change intensity information based on the cosine angle similarity measure; use the cosine angle similarity measure to measure the distance difference between objects for both spectral features and texture structure features. : (8） (9) and Representing the objects in the two phases of images respectively Features; Step 1.3: Spectral characteristic distance Distance to texture structure features Obtain information on the intensity of change: (10) Step 1.4: Obtain image edge intensity features through gradient calculation, combine the obtained change intensity information and the buildings extracted by semantic segmentation, construct a graph cut optimization model and use the maximum flow minimum cut theory to obtain the global optimal solution, and generate candidate areas for building changes in the early and late phases respectively. The edge strength is expressed as ,in and Representing adjacent and The pixel value.

4. The method for detecting building changes in remote sensing images based on matching optimization as described in claim 3, characterized in that, Step 2 includes: Step 2.1: Dilate the generated candidate regions of building changes. For the dilated candidate regions of building changes, perform precise feature point detection based on phase consistency after extracting local frequency information through log-Gabor wavelets. Step 2.2: Histogram matching based on phase consistency direction, 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 phase consistency feature value and feature vector for each pixel within the window to obtain the feature information required by HOPC; Step 2.23: Divide the template window into n blocks, each block containing m units, to form the basic architecture of HOPC; Step 2.24: Calculate the phase consistency direction histogram for each unit; Step 2.25: To eliminate the effects of illumination and contrast variations, normalize the phase consistency orientation histogram of each cell within each overlapping block; Step 2.26: Combine the histogram vectors within all blocks to form the HOPC descriptor; Step 2.3: Improve the stability of the matching process by determining the correction through a fast and robust filtering strategy, specifically including: Step 2.3.1: Use NCC as the similarity measure to perform matching within the limited range of detection points; （16） In the formula, and Representing template images and HOPC descriptor, and They represent and The mean; Step 2.3.2: Utilize After finding a match between candidate building regions with dual temporal variations, a fast and robust filtering strategy based on grid motion statistics is used to evaluate the number of matching pairs contained in the neighboring regions of the feature points of the matching pair to be judged, distinguishing between correct and incorrect matches, and generating the final geometrically consistent matching result.

5. The method for detecting building changes in remote sensing images based on matching optimization as described in claim 4, characterized in that, Step 3 specifically involves: Step 3.1: Calculate the joint change probability required for the univariate potential by using the candidate areas of building change and the matched roof points. ; The energy function of the model is expressed as follows: (17) in, It is the set of all pixels in the image. It is a univariate potential function, representing the label assigned to No. The cost of each node is given by the following formula: (18) (19) In the formula, It is estimated based on the matching results. The magnitude of the change at a point, i.e., the joint probability of change. The maximum distance is set; It is any point To the nearest matching point distance, Indicates matching point The connected region it is located in and For the changes in the candidate area of ​​the building between the two time phases The value; Step 3.2: Calculate the joint edge intensity of the two images; It is a binary potential function used to calculate the label and To ensure spatial continuity in the segmentation results, the costs assigned to the two nodes are determined separately. To further enhance the segmentation, objects that are close together and have similar features are assigned the same label. This paper employs two Gaussian kernel functions, as shown in the following equation: (20) in, The weights are the values ​​for two kernel functions, which are defined as follows: (21) (22) (23) In the formula, and There are two kernel functions. To achieve joint edge intensity, the maximum color contrast is extracted from the two consecutive images. and express Two pixels in the time phase and spectral values; and These represent the corresponding pixel coordinates, and the relationship between them is determined by parameters. Control Used to control the effect of color similarity Used to control the size of the spatial range. Used to eliminate smaller, isolated areas; Step 3.3: Based on all candidate areas of building changes in the two phases of images, optimize using a fully connected conditional random field to obtain the final building change detection results.

Images