Building group pattern recognition method considering second-order texture features and SE-CNN network

Abstract

The invention provides a building group pattern recognition method considering second-order texture features and an SE-CNN network, and relates to the technical field of building group pattern recognition, and the method comprises the following steps: generating a second-order texture co-occurrence matrix by using the area, roundness, elongation, minimum rectangular direction and influence area of a building group; and on the basis of the constructed SE-CNN network model, according to the generated second-order texture co-occurrence matrix, quantifying spatial structure characteristics of the building group. The method provides a practical value for analyzing the urban space structure and supporting map comprehensive decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of building group pattern recognition, in particular to a building group pattern recognition method considering second-order texture features and SE-CNN network. BACKGROUND

[0002] Buildings are important places for human living and activities. When they are arranged according to certain spatial rules, they form building groups, the characteristics of which can be expressed by spatial distribution patterns. Such patterns are shaped by the long-term interaction of human beings, society and natural environment, and reflect the influence of culture, social economy and ecological factors on geographical space. Building group recognition is a key task in map analysis, which reveals how individuals use maps as cognitive tools to perceive the world, and is widely used in spatial similarity judgment, spatial reasoning, spatial data mining and map generalization.

[0003] At present, the spatial pattern cognition and recognition of building groups mainly includes two types: building spatial clustering and building pattern recognition. Spatial clustering, as the basis of pattern recognition, aims to divide unlabeled buildings into different categories according to their inherent characteristics or similarity, so as to extract meaningful spatial distribution patterns. Researchers usually consider proximity, geometric similarity and continuous distribution in combination with Gestalt cognitive principles to group buildings, laying the foundation for subsequent pattern recognition. Related research dynamically extracts the structural characteristics of building groups through multiple connection linear patterns, proximity graphs, and area, orientation, and linear arrangement degree indicators. In terms of building pattern recognition, methods can be divided into rule-based unsupervised matching and sample-based supervised learning. The former relies on manually designed rules and templates, and performs linear or combined pattern recognition through MST algorithm or ladder-shaped, Z-shaped and H-shaped templates; the latter has gradually become the mainstream with the development of deep learning, and has shown obvious advantages in building group pattern recognition due to its adaptive feature learning and complex pattern representation capability. Some researchers try to analyze group patterns by processing vector building data through graph convolution networks, or fuse multiple source features, including building morphology, street view images, satellite images and point of interest information, to perform functional or pattern recognition. Some scholars generate color, texture, direction and curvature semantic feature maps through point cloud reprojection.

[0004] Existing research has made significant progress in building group pattern recognition. The method based on artificial rules or templates can better formalize spatial geometric relations and is suitable for the recognition of specific patterns. The deep learning method has advantages in complex pattern representation and large-scale data processing due to its strong feature learning ability. However, both methods have limitations: the former relies too much on prior knowledge and is difficult to adaptively capture multi-level and multi-form spatial structures; the latter often focuses on a single feature and fails to fully integrate texture characteristics and spatial context relationships. In the group recognition task where spatial structures are highly similar but semantic categories are different, these shortcomings can easily lead to misjudgment. SUMMARY

[0005] To solve the above problems, the purpose of the present application is to provide a building group pattern recognition method considering second-order texture features and SE-CNN network, aiming to improve the accuracy and generalization ability of building pattern recognition.

[0006] In order to achieve the above technical purpose, the present application provides a building group pattern recognition method considering second-order texture features and SE-CNN network, comprising the following steps: Generating a second-order texture co-occurrence matrix using the area, circularity, elongation, minimum rectangular direction and influence area of the building group; Quantifying the spatial structure features of the building group based on the constructed SE-CNN network model and the generated second-order texture co-occurrence matrix.

[0007] Preferably, in the generation of the second-order texture co-occurrence matrix, the proportion of the gray value combination appearing in the three adjacent pixel sequences is obtained according to the proportion of the gray value combination appearing in the three adjacent pixel sequences.

[0008] Preferably, before quantifying the spatial structure features of the building group, the building group is classified into regular building group and irregular building group, wherein the regular building group is a small building group closely gathered or a long strip building group scatteredly gathered; the irregular building group is a mixed building group of densely gathered embedded structure building group and sparse building.

[0009] Preferably, in the execution of building group pattern recognition classification, the building group is divided into small-scale building group closely gathered, long strip building group scatteredly gathered, densely gathered embedded structure building group and mixed building group containing other buildings.

[0010] Preferably, in the construction of the SE-CNN network model, the SE-CNN network model is composed of 6 layers, including: The first layer is a convolution unit with a convolution kernel size of 3x3 and a 2x2 pooling unit; The second layer is a channel attention mechanism module, comprising a global average pooling, a fully connected compression layer and a fully connected recovery layer. The third layer is a residual module, which is composed of two 3*3 convolution layers and a skip connection. The fourth layer is a second channel attention mechanism. The fifth layer is a convolution and pooling combination unit. The sixth layer is a fully connected layer.

[0011] Preferably, when the fifth layer is set, a convolution and pooling combination unit is formed by 3*3 convolution and 2*2 pooling.

[0012] Preferably, when the SE-CNN network model is constructed, the model generalization ability is improved by using BatchNorm+LeakyReLU+Dropout combination.

[0013] Preferably, when the SE-CNN network model is constructed, the AdamW optimizer is used to handle the weight decay problem.

[0014] The application also discloses a building group pattern recognition system considering second-order texture features and an SE-CNN network, which is used to implement the building group pattern recognition method considering second-order texture features and the SE-CNN network mentioned above, and comprises: A data processing module is configured to generate a second-order texture co-occurrence matrix by using the area, circularity, elongation rate, minimum rectangular direction and influence area of the building group. A quantitative recognition module is configured to quantize the spatial structure features of the building group based on the generated second-order texture co-occurrence matrix and the constructed SE-CNN network model.

[0015] The application also discloses a non-transitory computer readable storage medium storing computer instructions for causing a computer to execute the building group pattern recognition method considering second-order texture features and the SE-CNN network mentioned above.

[0016] The application discloses the following technical effects: The application provides a solution for building group pattern recognition, which has accuracy, robustness and expansibility, and provides practical value for urban space analysis and intelligent governance. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 is the structure of the overall framework of the fusion of the second-order texture feature and the SE-CNN network and its execution process diagram according to the present application; Figure 2 is a second-order texture co-occurrence matrix generation diagram according to the present application; Figure 3 is a direction angle division diagram according to the present application; Figure 4 is a building group adjacency relationship construction diagram according to the present application; Figure 5 is a SqueezeandExcitation module structure diagram according to the present application; Figure 6 is an SE-CNN network framework for building group pattern recognition according to the present application; Figure 7 is a training loss, training accuracy and validation accuracy change over time according to the present application. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0020] As shown in Figures 1-7 , the present application provides a building group pattern recognition method considering second-order texture features and SE-CNN network, specifically including the following processes: The present application constructs an overall framework of the fusion of second-order texture features and SE-CNN network. The framework covers four links of gray feature selection, second-order texture feature matrix construction, pattern recognition and evaluation and analysis, and the overall process is as shown in Figure 1The framework first converts the building vector data into Voronoi tessellation and Delaunay triangulation to construct the spatial adjacency relationship. Then, according to the area, circularity, elongation, minimum rectangular orientation, and polygon influence area, a second-order texture co-occurrence matrix is generated to quantify the spatial structure characteristics of the building group. Finally, the SE-CNN network with embedded channel attention mechanism and residual connection is used to identify four representative building group types. The present application combines texture features with deep learning to improve the recognition accuracy and generalization ability in complex spatial scenes. Specifically, the present application includes the following contents: 1. Feature selection: The pattern of the building group is different at different spatial scales, so the pattern recognition of the building group usually needs to combine the Gestalt principle and the psychological cognitive law in cognitive psychology (Lietal., 2004). As shown in Table 1, five feature indicators are used to describe the area, size, direction, and density of a single building. In order to ensure that all feature-based texture co-occurrence matrices have the same dimension, these attributes will be divided into the same number of classes. For area, circularity, elongation, and polygon influence area, the natural breakpoint method is used to divide them into classes, and for the minimum rectangular direction, they are also divided into classes. For this, the gray levels of all building feature attributes are divided into The present application uses all the second-order texture co-occurrence matrices generated by the feature indicators as feature inputs for model training.

[0021] Table 1: Building feature selection 2. Construction of the second-order texture co-occurrence matrix of the building group: 2.1. Definition of the second-order texture co-occurrence matrix: The second-order texture co-occurrence matrix mainly calculates the proportion of the gray value combination appearing in three adjacent pixel sequences. For an image with a gray range of , let , , represent three consecutive pixels, and be the center pixel. The line connecting to defines the first-order direction , and the line connecting to defines the second-order direction . Considering the combination of the first-order direction and the second-order direction , we divide and The value range of is divided into 8 directions: 0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°, a total of 8 directions, and Possible 8x8 (64) combinations, each combination corresponds to a second-order texture co-occurrence matrix, denoted as The value at each position in the matrix is represented as , whose calculation is shown in the formula: , where , , Indicates the coordinate position of the gray value of the three adjacent pixels in the three-dimensional texture co-occurrence matrix, and . Indicates the number of times the gray value combination appears when the first-order direction of the three adjacent pixels is . Indicates the total number of all combinations of three gray values corresponding to all direction angles. From the calculation relationship, is a three-dimensional matrix of , as shown in the example of Figure 2 .

[0022] Figure 2 Indicates that for the first-order direction , the second-order direction , and the gray value combination is (2, 1, 1), the number of times the gray value combination appears in this case is 2, the total number of gray value combinations under this angle direction is 24, so the value of is 2 / 24. The example of the gray combination (2, 1, 1) under the direction combination and is highlighted with a red line.

[0023] Considering the randomness of the adjacent direction of the building, the adjacent direction is divided into eight ranges, each range is centered on a representative direction, and the eight directions correspond to 0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°, as shown in Figure 3 .

[0024] 2.2, Building group adjacency relationship establishment: Convert the building group vector graphics into a image-like tessellation form, in this process, use the triangular subdivision line as the edge connection to build the adjacency relationship between the buildings, and the adjacency relationship construction graph is shown in Figure 4 ,​​Figure 4 Taking the area feature parameter as an example, the gray scale is divided into four levels of 0, 1, 2 and 3, corresponding to gray values of , the first-order direction , the second-order direction , the template combination is selected, and the template matching is performed in the building group adjacency relation graph. In the matching process, the sum of all templates at the direction angle is calculated and recorded as , the sum of the specific template combination at the direction angle is calculated as , the sum of all templates at the first-order direction Figure 3 and the second-order direction is 16, and = 16. Among these template combinations, there are 13 different combination forms, and for the combination, a total of 2 times occur, and = 2. Therefore, the second-order texture co-occurrence matrix value of the first-order direction , the second-order direction and the gray value combination (1, 1, 1) is , and the value is 2 / 16. By using the same method, the values of all template combinations at the direction angle are calculated, and the specific values are shown in Table 2. As shown in , taking the area feature as an example, when the direction combination is Figure 4 and , the generation process of each gray scale combination is given, and the numbers represent the area gray scale.

[0025] Table 2: Second-order texture co-occurrence matrix value of each gray value combination 2.3, building group pattern recognition classification: At present, building groups mainly include regular and irregular types. The present application divides the research area building group into regular and irregular building groups. The regular building group is a small building group that is closely gathered or a long strip building group that is dispersedly gathered. The irregular building group is a densely gathered embedded structure building group and a sparse mixed building group containing factories and other buildings.

[0026] Class A is mainly a small-scale building group that is closely gathered. In this type, the building shape is approximately regular geometric shape (such as square, rectangle, etc.), the arrangement shows high repeatability and consistency, and is mostly uniformly distributed, like a regular grid, with strong order feeling of spatial layout, and the spacing and orientation between building individuals are relatively uniform. This type often appears in the mature period of urban development, is related to urban high-rise buildings, and usually appears in the central area of the city.

[0027] ​Category B mainly consists of scattered, elongated building clusters. In this type, buildings are predominantly linear in shape, arranged in an orderly manner along a certain direction (such as roads, rivers, etc.), exhibiting a clear linear order. The arrangement direction is relatively uniform, with buildings extending in the length direction and relatively consistent in the width direction, forming a regular strip-shaped layout. This type often appears during the planning-driven period of urban development, especially during the stage when modern urban planning concepts have a profound influence, or during the industrialization and post-industrialization period of large-scale urban development and construction. It is mainly closely related to strict modern urban planning systems and rationalist planning concepts, pursuing efficiency, order, and standardization, facilitating land use, infrastructure support (such as more convenient pipeline laying in a regular layout), and traffic organization.

[0028] Category C mainly consists of densely clustered embedded building groups. In this type, the building forms are diverse, with no obvious unified geometric rules. The buildings vary greatly in size, shape, and orientation, and the spatial layout is flexible. There are cases of building splicing and nesting. This type often appears in the historical inheritance period of urban development, the transition from spontaneous development to planning intervention, and is often related to the inheritance of historical culture, natural terrain constraints, and gradual development models.

[0029] Category D mainly consists of sparse mixed building groups containing factories and other structures. In this type, the buildings are arranged in a more chaotic manner, with fragmented and irregular shapes. The relationships between buildings are complex, lacking clear order and traces of unified planning. They may be formed by the disorderly superposition of buildings from different periods and with different functions. The spatial connectivity and order are poor. This type often appears in the early spontaneous formation period of urban development and the period of lack of planning and control. It is often related to the lack of planning management in the early stage, the randomness of villagers' self-construction, and the spontaneous construction at the stage of low economic development. The construction did not consider long-term layout and overall order, and only met the basic functions.

[0030] 3. SE-CNN network model establishment: SqueezeandExcitationblock can improve the representational power of a network by explicitly modeling the interdependencies between convolutional feature channels, which is effective for feature learning. SqueezeandExcitationblock is mainly divided into Transformation (… Squeeze ), Excitation ) and Scale The Squeeze-and-Excitation module is structured as follows: (Four parts) Figure 5 As shown.

[0031] Transformation ) is after , the feature map is changed into a feature map , The formula can be defined as formula: wherein, is an input feature map, is an output feature map, represents a kernel learning a group of filters, represents a parameter of the th filter, represents a 2-dimensional spatial kernel, represents a convolution operation.

[0032] Squeeze( ) is a global average pooling using channels, and the purpose is to compress the feature map into a feature vector, so as to contain the context information of generated channels and statistical data, and alleviate the channel dependency problem.The formula can be defined as formula: wherein, is the th element of .

[0033] Excitation( ) is used to gather feature information by using channel compression operation, and adopts two fully connected layers to constitute a gate mechanism, the first fully connected layer compresses C channels into C / r channels to reduce the calculation amount, and then passes through a RELU nonlinear activation layer, the second fully connected layer restores the channel number to C channels, and then obtains the weight s through a Sigmoid activation, and finally the dimension of the obtained s is 1x1xC, which is used to depict the weight of C feature maps in the feature map U. The formula can be defined as formula: Scale( ) is used to weight the attention weight obtained in the foregoing to each channel feature, and the formula is defined as: The SE-CNN network model used in the application is as shown in Figure 6As shown, it is mainly composed of 6 layers. First is the input layer (not counted in the number of layers), when the feature gray scale is divided into 8 categories, this layer is composed of a three-dimensional matrix of 5*64*8*8*8. The first layer is a convolution unit with a kernel size of 3*3 and a 2*2 pooling unit; the second layer is a channel attention mechanism module, including global average pooling, fully connected compression layer and fully connected recovery layer; the third layer is a residual module, composed of two 3*3 convolution layers and a skip connection; the fourth layer is the second channel attention mechanism; the fifth layer is a convolution and pooling combined unit (3*3 convolution and 2*2 pooling); the sixth layer is a fully connected layer; finally, the recognition result is obtained through the output layer (not counted in the number of layers).

[0034] The SE-CNN network model mainly introduces a channel attention mechanism and a residual connection layer on the basis of a traditional convolutional neural network, for enhancing the weight learning of feature channels, improving the expression ability of key features and relieving the gradient vanishing problem in the learning process, and improving the network training stability; using BatchNorm+LeakyReLU+Dropout combination to improve the model generalization ability and reduce the risk of overfitting; using AdamW optimizer to better handle the weight decay problem.

[0035] The building vector data is used as the experimental sample, wherein the A-class building group has 314 groups, the B-class building group has 509 groups, the C-class building group has 426 groups, and the D-class building group has 458 groups; the experimental sample is divided into a training set and a test set in a ratio of 8:2, wherein 1366 samples are allocated to the training set, and 341 samples are used for testing; the network layer uses the AdamW optimizer and the LeakyReLU activation function, the Batch size and the learning rate are set to 32 and 0.0001 respectively, the training and test accuracy is reported once every 4 epochs, and after 200 times of training, the training and test accuracy reaches 0.9989 and 0.9669 respectively. The network model training accuracy is as shown in Figure 7 The x-axis represents the number of training rounds, and the left y-axis represents the loss value and the accuracy.

[0036] In order to verify the effectiveness of the network model proposed in the application, the model is used for comparative experiments with the GCN method, the CNN method using the first-order texture matrix as the input vector and the Transformer method. The same building group in a region is identified by the three methods. The GCN uses the five feature parameters in section 3.1 as node attributes, uses the Voronoi tessellation to construct the graph structure, and uses the Delaunay triangulation line as the edge relationship between nodes. The graph convolution is performed, and then the building group is identified. The CNN network uses the first-order texture matrix as the input vector for comparative experiments. Finally, the Transformer method uses the area, the minimum rectangular direction and the influence area as the training features. After the encoder-decoder, the building group is identified. The accuracy results of the GCN, CNN and Transformer models are shown in Table 3. As can be seen from Table 4, the method proposed in the application achieves a performance of 95% in the recall rate index. Compared with GCN, CNN and Transformer, the performance is improved by about 16%, 12% and 6% respectively. In the F1-score index, the performance is 97%. Compared with GCN, CNN and Transformer, the performance is improved by about 16%, 15% and 6% respectively.

[0037] Table 3 Accuracy results of comparative experiments Model verification In order to further verify the effectiveness and generalization ability of the SE-CNN network model proposed in the application, the model is used for building group pattern recognition in another region. The experimental region is Lanzhou City in Gansu Province. The building vector data of Lanzhou City is downloaded from the OpenStreetMap website. The original data only contains building contour information. The building vector data of Lanzhou City is manually divided using road and block data. The building group data used for experiments is divided. The number of building groups used for experiments and the artificial recognition results are shown in Table 4. The A class building in the experimental region is less, only 12.93%, which indicates that there are fewer high-rise buildings in the later stage of urban construction. The test area is mainly medium and large building groups, and B and C class building groups are common.

[0038] Table 4 Artificial identification results of building groups The recognition results are shown in Table 5. From Table 6, it can be seen that the model has a recognition accuracy of more than 93% for the four types of building groups, a recall rate of more than 95%, and an F1-score of more than 97%, proving the practical value of the model. The model recognition results of some building groups in the experimental area. The area contains a total of 80 building groups, including 3 A-class building groups, 18 B-class building groups, 46 C-class building groups, and 13 D-class building groups. The recognition results are basically consistent with manual recognition. In the real world, B-class building groups usually contain small buildings that are densely clustered, which makes them prone to being misidentified as C-class buildings, resulting in recognition errors (recognition errors are marked with a red box).

[0039] Table 5 Recognition results of SE-CNN network model Effectiveness analysis: To verify the effectiveness of the second-order texture feature and channel attention mechanism in building group pattern recognition, the present application uses recall (Recall) and F1-score as evaluation indicators and designs a series of ablation experiments based on the CNN model (denoted as SE-CNN). By gradually introducing the second-order texture feature (denoted as SE-2nd) and the channel attention mechanism (denoted as SE-Att), the system evaluates the contribution of each module to the model performance. From the experimental results in Table 6, it can be seen that when only the second-order texture feature (SE-2nd) is introduced, the recognition accuracy of the four types of building groups is improved by more than 10%. When the second-order texture feature (SE-2nd) is introduced alone, the various indicators are generally improved by more than 5%.

[0040] Table 6 Ablation experiment results ​​​​​The building group pattern recognition method fusing the second-order texture feature and the SE-CNN network has higher accuracy and robustness in the experimental results. The superior performance is not only due to the optimized design of the overall architecture, but also due to the synergistic effect of each module. Specifically, the second-order texture feature describes the spatial heterogeneity and local arrangement of the building group, providing rich mid-level visual cues for the model, so that it can maintain high discriminability when facing groups with complex shapes or similar spatial structures. At the same time, the introduced channel attention mechanism effectively enhances the weight learning of key texture channels, so that it can realize differentiated attention in the feature dimension, thereby avoiding the decline of recognition accuracy caused by feature redundancy or noise. The residual structure not only relieves the gradient vanishing problem in the training process of the deep network, but also improves the training stability of the model on a large number of samples. The experimental results show that in the recognition of four typical building groups, the recall rate of the method is more than 95%, and the F1-score is more than 97%, which is about 16%, 15% and 6% higher than that of GCN, CNN and Transformer method respectively, fully verifying the effectiveness and superiority of the method in complex spatial patterns. This advantage makes the method have application potential in smart city construction, urban spatio-temporal evolution research and geographic spatial big data updating scenes.

[0041] In summary, the building group pattern recognition framework is proposed, which effectively quantifies and describes the spatial structure features of the building group by fusing the second-order texture feature and the SE-CNN network. Specifically, the adjacency relationship of the building group is established by Voronoi tessellation and Delaunay triangulation, and the second-order texture co-occurrence matrix is constructed by combining area, circularity, elongation, minimum rectangular direction and influence area, etc. to realize the coding of local and global structure features of the group. On this basis, the SE-CNN network with channel attention mechanism and residual connection is introduced to further enhance the expression ability of key texture features. The experimental results show that this method has excellent performance in building group pattern recognition tasks in Wuhan, Guangzhou, Kunming and Lanzhou, and the recognition recall rate of four typical groups is more than 95%, and the F1-score is more than 97%, which is about 16%, 15% and 6% higher than that of GCN, CNN and Transformer method respectively, and has good generalization ability in different cities. Overall, the model performs well in building group pattern recognition tasks, providing practical value for analyzing urban spatial structure and supporting map synthesis decision.

[0042] The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other processing device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other processing device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions specified in the flowchart block or blocks or in conjunction with the flowcharts and / or block diagram. Figure 1 one or more of the blocks or steps in the flowchart or flowchart diagrams. Figure 1 one or more of the blocks or steps in the flowchart or flowchart diagrams.

[0043] In the description of the present application, it is to be understood that the terms "first", "second", "third" and the like, merely identify features belonging to distinct categories, and do not imply or imply a relative importance or a specific number thereof. Thus, a feature identified as a "first" and / or "second" feature can implicitly or explicitly include one or more of the feature. In the description of the present application, the meaning of "a plurality" is two or more, unless otherwise expressly specified.

[0044] Obviously, various modifications and changes can be made to the present application by those skilled in the art without departing from the spirit and scope of the present application. Thus, it is intended that the present application also include such modifications and changes insofar as they come within the scope of the claims of the present application and their equivalents.

Claims

1. A building group pattern recognition method considering second-order texture features and SE-CNN network, characterized in that, Includes the following steps: A second-order texture co-occurrence matrix is ​​generated using the area, roundness, elongation, minimum rectangle direction, and affected area of ​​the building group. Based on the constructed SE-CNN network model, the spatial structural features of building groups are quantified according to the generated second-order texture co-occurrence matrix.

2. The building group pattern recognition method considering second-order texture features and SE-CNN network according to claim 1, characterized in that: When generating the second-order texture co-occurrence matrix, the second-order texture co-occurrence matrix is ​​obtained based on the proportion of gray value combinations appearing in three adjacent pixel sequences.

3. The building group pattern recognition method considering second-order texture features and SE-CNN network according to claim 2, characterized in that: Before quantifying the spatial structural characteristics of building groups, the building groups are classified into regular building groups and irregular building groups. Regular building groups are either closely clustered small building groups or dispersed long strip-shaped building groups; irregular building groups are either densely clustered embedded structure building groups or mixed building groups with sparse buildings.

4. The building group pattern recognition method considering second-order texture features and SE-CNN network according to claim 3, characterized in that: When performing building group pattern recognition and classification, building groups are divided into tightly clustered small-scale building groups, dispersed long strip-shaped building groups, densely clustered embedded structure building groups, and sparse mixed building groups containing other buildings.

5. The building group pattern recognition method considering second-order texture features and SE-CNN network according to claim 1, characterized in that: When constructing the SE-CNN network model, the SE-CNN network model consists of 6 layers, including: The first layer consists of 3×3 convolutional units and 2×2 pooling units; The second layer is the channel attention mechanism module, which includes global average pooling, a fully connected compression layer, and a fully connected recovery layer; The third layer is a residual module, consisting of two 3×3 convolutional layers and skip connections; The fourth layer is the second channel attention mechanism; The 5th layer is a combination of convolution and pooling units; The 6th layer is a fully connected layer.

6. The building group pattern recognition method considering second-order texture features and SE-CNN network according to claim 5, characterized in that: When setting up the 5th layer, a combination of convolution and pooling units is formed by using 3×3 convolution and 2×2 pooling.

7. The building group pattern recognition method considering second-order texture features and SE-CNN network according to claim 1, characterized in that: When building the SE-CNN network model, the combination of BatchNorm, LeakyReLU, and Dropout is used to improve the model's generalization ability.

8. The building group pattern recognition method considering second-order texture features and SE-CNN network according to claim 1, characterized in that, When building the SE-CNN network model, the AdamW optimizer is used to handle the weight decay problem.

9. A building group pattern recognition system considering second-order texture features and SE-CNN network, for implementing the method as described in claim 1, characterized in that, include: The data processing module is used to generate a second-order texture co-occurrence matrix by utilizing the area, roundness, elongation, minimum rectangle direction, and affected area of ​​the building group. The quantization recognition module is used to quantify the spatial structural features of building groups based on the constructed SE-CNN network model and the generated second-order texture co-occurrence matrix.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing a computer to perform the method described in any one of claims 1 to 7.

Images