A building facade parsing method using repeated texture constraints

Abstract

The present invention discloses a method, system and electronic device for building facade parsing using repetitive texture constraints. The method includes: Step 1, data preparation, preparing a set of building facade images and performing semantic label annotation on the images to obtain a semantic map; Step 2, constructing a Multi-DRCNet network using local and row-column direction repetitive texture feature constraints, mainly including a multi-layer dilated convolution module MD-Module and a multi-scale row-column convolution module MRC-Module that can extract local repetitive texture and row-column repetitive texture features; Step 3, training the Multi-DRCNet network using publicly available building facade data and saving the model parameters as pre-trained weights; Step 4, loading the pre-trained weights saved in Step 3 and fine-tuning the Multi-DRCNet network using the building facade data prepared in Step 1. The fine-tuned Multi-DRCNet network can be used to parse building facade images. The present invention can effectively solve the occlusion problem caused by objects such as trees to the building facade, improve the facade parsing accuracy, and achieve a better visual interpretation effect.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision semantic segmentation, and in particular relates to a building facade analysis technology using repeated texture constraints, which is mainly used in the fields of building facade segmentation, building attribute information acquisition, building accessory object detection, etc. Background Art

[0002] Building facade image analysis, as a key task in understanding urban scenes, focuses on the accurate segmentation of elements such as doors, windows, and balconies. Its derived applications cover areas such as building 3D reconstruction, autonomous driving, and urban planning, and are of outstanding importance in the field of computer vision. However, building facade analysis is not a simple semantic segmentation task. There are two main difficulties: ① There are various shapes of trees, cars and other objects around the building that will block the building itself. These occlusion information is a difficult problem to solve in the building facade analysis task; ② The existing building facade datasets are relatively small, and deep learning methods are prone to overfitting on small datasets.

[0003] Traditional building facade parsing methods are mainly based on the grammatical rules of buildings or feature extraction based on computer vision. Since the shape of buildings is artificially designed and has strong repeatability, although parsing regular building facades based on these properties can achieve good results, it is impossible to achieve high-precision and automated parsing effects for large-scale scenes with more complex occlusions. Deep learning-based methods are commonly used in computer vision semantic segmentation tasks and can be directly applied to building facade parsing. However, existing methods rarely design algorithms based on the structural characteristics of the building itself, and general deep learning methods cannot take advantage of the data advantage of the repeated texture of the building facade to solve the occlusion problem. In addition, it is difficult for deep learning technology to take advantage of small sample data. Summary of the invention

[0004] The technical goal of the present invention is to solve the problem of achieving automatic and high-precision analysis on building facade data with occlusion, and to provide a building facade analysis method using repeated texture constraints. The core of this method is to design a neural network Multi-DRCNet that uses local and row-column repeated texture feature constraints, mainly using large-scale dilated convolution kernels and rectangular convolution kernels to construct multi-layer dilated convolution modules MD-Module and multi-scale row-column convolution modules MRC-Module, to achieve local repeated texture and row-column repeated texture feature extraction, solve the occlusion problem, and improve the analysis accuracy of repetitive objects such as doors, windows, and balconies.

[0005] Compared with the existing methods, the present invention can achieve the best results in the building facade parsing for both occluded and unoccluded cases. Therefore, the present invention has very important practical value and broad application prospects in the direction related to building facade parsing.

[0006] The present invention provides a building facade parsing method using repeated texture constraints, and the method includes the following steps:

[0007] Step 1, obtain the building facade image and the corresponding semantic label map, and perform data preprocessing;

[0008] Step 2, construct a neural network Multi-DRCNet using local and row-column direction repeated texture feature constraints, including a multi-layer dilated convolution module MD-Module and a multi-scale row-column convolution module MRC-Module;

[0009] The overall framework structure of the Multi-DRCNet is consistent with that of the U-Net, adopting an encoder-decoder structure. In the encoding stage, a pre-trained basic feature extraction network is used. In the decoding stage, the ordinary convolution layers in the upsampling of the U-Net are replaced with the multi-scale row-column convolution module MRC-Module, and the MD-Module is added in the middle of the encoding and decoding;

[0010] The multi-layer dilated convolution module MD-Module includes multiple dilated convolutions, and the multi-scale row-column convolution module MRC-Module includes convolutions in multiple different directions;

[0011] Step 3, train the Multi-DRCNet network using a public building facade dataset, and perform accuracy evaluation;

[0012] Step 4, load the weights trained on the public dataset, fine-tune the trained neural network on the data prepared in Step 1, and use the fine-tuned network to parse the building facade.

[0013] Further, the specific implementation method of Step 1 is as follows:

[0014] Step 1.1, determine the training set: Given a threshold t, if the total number of images is less than or equal to t, the number of training sets is half of the total number of images; if the total number of images is greater than t, the number of training sets is t / 2, and then the total data is shuffled, and a random strategy is adopted for extraction to ensure the randomness of the data;

[0015] Step 1.2, manually or automatically annotate the dataset to obtain the categories that need to be segmented in the building facade image, and obtain the semantic labels corresponding to the building facade image;

[0016] Step 1.3, perform data augmentation operations on the data, including random rotation, scaling, translation, flipping, and finally adjust the image size.

[0017] Furthermore, the input and output of the multi-hole convolution module MD-Module in step 2 are feature maps of a specific size. The MD-Module includes 5 hole convolution operations. The convolution kernels are square with a size of 3, but their dilation rates are different, which are 1, 2, 4, 8, and 16 respectively. Finally, the input of the MD-Module and the outputs of the five hole convolution modules are connected. After adjusting the dimension of the output through a conventional convolution, it is output to the upsampling process and finally output;

[0018] There is batch normalization BN and ReLU activation after each convolution in the MD-Module.

[0019] Furthermore, the result of upsampling the MD-Module and the features extracted from the symmetric position in the encoding stage are concatenated together to form the input features of the current MRC-Module. The input features are directly used in the MRC-Module. After passing through three parallel sub-modules, the features are fused together and output through a convolution layer. Each sub-module has two convolution kernels K Row and K Column , which are used to obtain the features in the row and column directions respectively, and then the information in the row and column directions is fused through a convolution layer.

[0020] Furthermore, the sizes of the convolution kernels K Row and K Column are 3 rows by 15 columns and 15 rows by 3 columns respectively. The dilation rates of K Row and K Column in the three sub-modules are 1, 2, and 4 respectively.

[0021] Furthermore, in step 3, the specific method for training the Multi-DRCNet network and evaluating its accuracy is as follows:

[0022] Step 3.1, determine the hyperparameters: Use the Adam optimizer, set the initial learning rate, weight decay rate, batchsize, and the number of loop iterations. The learning rate is multiplied by a constant for reduction in each iteration;

[0023] Step 3.2, determine the loss function: The loss function uses cross-entropy loss, defined as where N is the number of all pixels, is the true probability distribution value of the i-th pixel, is the probability distribution value predicted by the neural network for the i-th pixel;

[0024] Step 3.3, use a publicly available building facade image dataset with sufficient and diverse samples to train the network and save the model parameters;

[0025] Step 3.4, Evaluation of the pre-trained network accuracy: The Intersection over Union (IOU), Pixel Accuracy (PA), and Overall Accuracy (Acc) are used as evaluation metrics.

[0026] Furthermore, in Step 4, the specific method for fine-tuning the Multi-DRCNet network and parsing the building facade images is as follows:

[0027] Step 4.1, Fine-tuning of the Multi-DRCNet network parameters: Load the pre-trained model parameters of the public dataset saved in Step 3.3, and continue to train the Multi-DRCNet network on the training set prepared in Step 1 according to the hyperparameters in Step 3.1;

[0028] Step 4.2, Parsing the building facade images: Use the newly trained Multi-DRCNet network to parse the building facade and output the visualized segmentation results.

[0029] Furthermore, the pre-trained basic feature extraction network is the VGG16 network trained on the ImageNet dataset.

[0030] The present invention also provides a building facade parsing system using repeated texture constraints, including the following modules:

[0031] A preprocessing module for obtaining building facade images and corresponding semantic label maps and performing data preprocessing;

[0032] A model construction module for constructing a neural network Multi-DRCNet using local and row-column direction repeated texture feature constraints, including a multi-layer dilated convolution module MD-Module and a multi-scale row-column convolution module MRC-Module;

[0033] The overall framework structure of the Multi-DRCNet is consistent with that of the U-Net, adopting an encoder-decoder structure. In the encoding stage, a pre-trained basic feature extraction network is used. In the decoding stage, the ordinary convolution layers in the upsampling of the U-Net are replaced with a multi-scale row-column convolution module MRC-Module, and an MD-Module is added in the middle of the encoding and decoding;

[0034] The multi-layer dilated convolution module MD-Module includes multiple dilated convolutions, and the multi-scale row-column convolution module MRC-Module includes convolutions in multiple different directions;

[0035] A training and rating module for training the Multi-DRCNet network using a public building facade dataset and performing accuracy evaluation;

[0036] A parsing module, configured to load weights trained on a public dataset, fine-tune a trained neural network on the prepared data, and parse a building facade using the fine-tuned network.

[0037] The present invention also provides an electronic device for parsing a building facade using repeated texture constraints, including: one or more processors;

[0038] A storage device, configured to store one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the method described in the above solution.

[0039] The beneficial effects produced by the present invention are as follows:

[0040] 1) For the building facade images, by constructing and training the Multi-DRCNet network, the present invention realizes full-automatic parsing and can obtain semantic information such as doors, windows, balconies, etc. on the building facade.

[0041] 2) By using the constructed MN-Module and MRC-Module, repeated texture information in the local and row-column directions can be obtained, which can better solve the occlusion problem in building facade parsing. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The present invention will be further described below in conjunction with the drawings and embodiments. In the drawings:

[0043] Figure 1 is the overall flowchart of the embodiment of the present invention;

[0044] Figure 2 is the framework diagram of the neural network Multi-DRCNet constructed by the present invention using local and row-column direction repeated texture feature constraints;

[0045] Figure 3 is the schematic structural diagram of the multi-layer dilated convolution module MD-Module in the Multi-DRCNet network of the present invention;

[0046] Figure 4 is the schematic structural diagram of the multi-scale row-column convolution module MRC-Module in the Multi-DRCNet network of the present invention;

[0047] Figure 5 is the comparison diagram of the building facade parsing results generated by the present invention with the true labels, Deeplabv3+, HRNet, and SegFormer.

[0048] Figure 6IOU index comparison chart of building facade parsing generated for the present invention with the accuracies of Deeplabv3+, HRNet, and SegFormer. Detailed implementation manners

[0049] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0050] As Figure 1 shown, the method for building facade parsing using repeated texture constraints in the embodiment of the present invention includes the following steps:

[0051] Step 1. The data preprocessing steps of the building facade image and the corresponding semantic label map are as follows:

[0052] Step 1.1. Determine the training set. Given the threshold t = 400, if the total number of images is less than or equal to t, the number of training sets is 50% of the total number of images; if the total number of images is greater than t, the number of training sets is 200. When determining the training set, the total data order should be shuffled, and a random strategy should be adopted for extraction to ensure the randomness of the data.

[0053] Step 1.2. Manually or automatically annotate the data set to obtain the categories that need to be segmented in the building facade image, such as: doors, windows, balconies, etc., and obtain the semantic labels corresponding to the building facade image.

[0054] Step 1.3. Perform data augmentation operations on the data, including random rotation, scaling, translation, flipping, etc., and finally adjust the image to a unified size of 512×512.

[0055] Step 2. The specific steps for constructing the neural network Multi-DRCNet using local and row-column direction repeated texture feature constraints, as well as the related multi-layer dilated convolution module MD-Module and multi-scale row-column convolution module MRC-Module, are as follows:

[0056] Step 2.1. Construct the multi-layer dilated convolution module MD-Module. As Figure 2 shown, the input and output of the MD-Module are both feature maps with a size of 32×32×512. The MD-Module has 5 dilated convolution operations, and the convolution kernels are square with a size of 3, but their dilation rates Dilation are different, which are 1, 2, 4, 8, and 16 respectively. Finally, the input of the MD-Module and the outputs of the five dilated convolution modules are connected, and after adjusting the dimension of the output through a conventional convolution, it is output to the upsampling process. After each convolution in the MD-Module, there is batch normalization BN and ReLU activation.

[0057] Step 2.2: Construct a multi-scale row-column convolution module MRC-Module. As Figure 3 shown, the result of upsampling from the previous module and the features extracted from the symmetric position in the encoding stage are concatenated together to form the input features of the current MRC-Module. The MRC-Module has three different sub-modules directly connected to the input features. After passing through the three parallel sub-modules, the features are fused and output through a 3×3 convolutional layer. Each sub-module has two convolutional kernels K Row and K Column of the same size but different directions, with sizes (3,15) and (15,3) respectively, that is, 3 rows and 15 columns and 15 rows and 3 columns, which are used to obtain features in the row and column directions respectively, and then the information in the row and column directions is fused through a 3×3 convolutional layer. To obtain the feature information in the row and column directions at different scales, the dilation rates of K Row and K Column in the three sub-modules are set to 1, 2, and 4 respectively.

[0058] Step 2.3: Construct the backbone of the neural network Multi-DRCNet that utilizes local and row-column direction repeated texture feature constraints. As Figure 4 shown, the overall framework structure of Multi-DRCNet is the same as that of U-Net, adopting an encoder-decoder structure. In the encoding stage, a pre-trained basic feature extraction network is used, such as the VGG16 network trained on the ImageNet dataset. In the decoding stage, the ordinary convolutional layers in the upsampling of U-Net are replaced with multi-scale row-column convolution modules MRC-Module. An MD-Module is added between the encoding and decoding.

[0059] Step 3: The specific implementation steps for training the Multi-DRCNet network and accuracy evaluation using a publicly available building facade dataset are as follows:

[0060] Step 3.1: Determine the hyperparameters used for network training. Use the Adam optimizer, set the initial learning rate to 0.0001, the weight decay rate to 0.0001, the batch size to 8, and iterate 300 times. Each time the iteration, the learning rate will be multiplied by a constant for reduction.

[0061] Step 3.2: Determine the loss function used for network training. The loss function uses cross-entropy loss, defined as where N is the number of all pixels, is the true probability distribution value of the i-th pixel, is the probability distribution value predicted by the neural network for the i-th pixel.

[0062] Step 3.3: Train the network using a publicly available building facade image dataset with sufficient and diverse samples and save the model parameters.

[0063] Step 3.4: Evaluate the accuracy of the Multi-DRCNet network trained using the publicly available dataset. The Intersection over Union (IOU), Pixel Accuracy (PA), and Overall Accuracy (Acc) are used as evaluation metrics.

[0064] Step 4: The specific process of network fine-tuning and building facade parsing is as follows:

[0065] Step 4.1: Fine-tune the parameters of the Multi-DRCNet network. Load the pre-trained model parameters of the publicly available dataset saved in Step 3.3 and continue to train the Multi-DRCNet network on the training set prepared in Step 1 according to the hyperparameters in Step 3.1.

[0066] Step 4.2: Parse the building facade image. Use the newly trained Multi-DRCNet network to parse the building facade and output the visualized segmentation results.

[0067] In specific implementation, the method proposed by the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. The system device for implementing the method, such as a computer-readable storage medium storing the corresponding computer program of the technical solution of the present invention and a computer device including a running corresponding computer program, should also be within the protection scope of the present invention.

[0068] On the other hand, an embodiment of the present invention provides a building facade parsing system using repeated texture constraints, including the following modules:

[0069] A preprocessing module for obtaining building facade images and corresponding semantic label maps and performing data preprocessing;

[0070] A model construction module for constructing a neural network Multi-DRCNet using local and row-column direction repeated texture feature constraints, including a multi-layer dilated convolution module MD-Module and a multi-scale row-column convolution module MRC-Module;

[0071] The overall framework structure of the Multi-DRCNet is the same as that of the U-Net, adopting an encoder-decoder structure. In the encoding stage, a pre-trained basic feature extraction network is used. In the decoding stage, the ordinary convolution layers in the upsampling of the U-Net are replaced with a multi-scale row-column convolution module MRC-Module, and an MD-Module is added between the encoding and decoding;

[0072] The multi-layer dilated convolution module MD-Module includes multiple dilated convolutions, and the multi-scale row-column convolution module MRC-Module includes convolutions in multiple different directions;

[0073] A training and rating module for training the Multi-DRCNet network using a public building facade dataset and performing accuracy evaluation;

[0074] An analysis module for loading the weights trained on the public dataset, fine-tuning the trained neural network on the prepared data, and parsing the building facade using the fine-tuned network.

[0075] Thirdly, an embodiment of the present invention further provides an electronic device for parsing a building facade using repeated texture constraints, including: one or more processors;

[0076] A storage device for storing one or more programs, which when executed by the one or more processors, cause the one or more processors to implement the method described in the above solution.

[0077] According to the specific steps of the above embodiment, taking the public dataset ENPC2014 building facade dataset as an example, the Multi-DRCNet network is constructed and trained. The hyperparameters used for network training are the same as those in step 3. The Adam optimizer is used, the initial learning rate is set to 0.0001, the weight decay rate is set to 0.0001, the batch size is set to 8, and the loop iterates 300 times. Each time the iteration is performed, the learning rate is multiplied by a constant Reduce. Compare the building facades analyzed by the present invention with the existing methods Deeplabv3+ (Chen L C, Zhu Y, Papandreou G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[C] / / Proceedings of the European conference on computer vision (ECCV). 2018: 801-818.), HRNet (Wang J, Sun K, Cheng T, et al. Deep high-resolution representation learning for visual recognition[J]. IEEE transactions on pattern analysis and machine intelligence, 2020, 43(10): 3349-3364.), and SegFormer (Xie E, Wang W, Yu Z, et al. SegFormer: Simple and efficient design for semantic segmentation with transformers[J]. Advances in Neural Information Processing Systems, 2021, 34: 12077-12090.). The visual effects are as Figure 5 shown, and the precision comparison of the IOU values is as Figure 6 shown.

[0078] As mentioned above, the above are only the preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the technical scope disclosed by the present invention shall be included within the protection scope of the invention.

Claims

1. A method for parsing building facades using repeated texture constraints, characterized in that It includes the following steps: Step 1: Obtain the building facade images and the corresponding semantic label maps, and perform data preprocessing; Step 2: Construct a neural network Multi-DRCNet that uses local and row-column direction repeated texture feature constraints, including a multi-layer dilated convolution module MD-Module and a multi-scale row-column convolution module MRC-Module; The overall framework structure of the Multi-DRCNet is the same as that of the U-Net, adopting an encoder-decoder structure. In the encoding stage, a pre-trained basic feature extraction network is used. In the decoding stage, the ordinary convolution layers in the upsampling of the U-Net are replaced with the multi-scale row-column convolution module MRC-Module, and the MD-Module is added in the middle of the encoding and decoding; The multi-layer dilated convolution module MD-Module includes multiple dilated convolutions, and the multi-scale row-column convolution module MRC-Module includes convolutions in multiple different directions; In Step 2, the input and output of the multi-layer dilated convolution module MD-Module are feature maps of a specific size. The MD-Module includes 5 dilated convolution operations. The convolution kernels are squares with a size of 3, but their dilation rates are different, which are 1, 2, 4, 8, and 16 respectively. Finally, the input of the MD-Module and the outputs of the five dilated convolution modules are connected. After adjusting the dimension of the output through a conventional convolution, it is output to the upsampling process and finally output; After each convolution in the MD-Module, there is batch normalization BN and ReLU activation; Step 3: Train the Multi-DRCNet network using a publicly available building facade dataset and perform accuracy evaluation; Step 4: Load the weights trained on the publicly available dataset, fine-tune the trained neural network on the data prepared in Step 1, and use the fine-tuned network to analyze the building facade.

2. The method for parsing building facades using repeated texture constraints according to claim 1, wherein: The specific implementation method of Step 1 is as follows: Step 1.1: Determine the training set: Given a threshold t, if the total number of images is less than or equal to t, the number of training sets is half of the total number of images; if the total number of images is greater than t, the number of training sets is t / 2. Then shuffle the total data and use a random strategy to extract to ensure the randomness of the data; Step 1.2: Manually or automatically annotate the dataset to obtain the categories that need to be segmented for the building facade images, and obtain the semantic labels corresponding to the building facade images; Step 1.3: Perform data augmentation operations on the data, including random rotation, scaling, translation, flipping, and finally adjust the image size.

3. The method for parsing building facades using repeated texture constraints according to claim 1, characterized in that: The result of upsampling the MD-Module and the features extracted from the symmetric position in the encoding stage are concatenated to form the input features of the current MRC-Module. The input features are directly used in the MRC-Module. After passing through three parallel sub-modules, the features are fused together and output through a convolutional layer. Each sub-module has two convolutional kernels K Row and K Column , which are used to obtain the features in the row and column directions respectively, and then the information in the row and column directions is fused through a convolutional layer.

4. The method for parsing building facades using repeated texture constraints according to claim 3, wherein: Convolution kernel K Row and K Column have sizes of 3 rows by 15 columns and 15 rows by 3 columns respectively. The K Row and K Column of the three sub-modules have dilation rates of 1, 2, and 4 respectively.

5. The method for parsing building facades using repeated texture constraints according to claim 1, characterized in that: In Step 3, the specific method for training the Multi-DRCNet network and its accuracy evaluation is as follows: Step 3.1, determine hyperparameters: Use the Adam optimizer, set the initial learning rate, weight decay rate, batch size, and number of loop iterations, and multiply the learning rate by a constant in each iteration for reduction; Step 3.2, determine the loss function: The loss function uses cross-entropy loss and is defined as where N is the number of all pixels, is the true probability distribution value of the i-th pixel, is the probability distribution value predicted by the neural network for the i-th pixel; Step 3.3: Use a publicly available building facade image dataset with sufficient and diverse samples to train the network and save the model parameters; Step 3.4: Evaluate the accuracy of the pre-trained network: Use the intersection over union IOU, pixel accuracy PA, and overall accuracy Acc as evaluation indicators.

6. The method for parsing building facades using repeated texture constraints according to claim 1, characterized in that: In Step 4, the specific method for fine-tuning the Multi-DRCNet network and analyzing the building facade images is as follows: Step 4.1, Fine-tuning of Multi-DRCNet network parameters: Load the pre-trained model parameters of the public dataset saved in Step 3.3, and continue to train the Multi-DRCNet network on the training set prepared in Step 1 according to the hyperparameters in Step 3.1; Step 4.2, Parsing building facade images: Use the newly trained Multi-DRCNet network to parse the building facade and output the visualized segmentation results.

7. The method for parsing building facades using repeated texture constraints according to claim 1, characterized in that: The pre-trained basic feature extraction network is the VGG16 network trained on the ImageNet dataset.

8. A building facade analysis system using repeated texture constraints, characterized in that, It includes the following modules: The preprocessing module is used to obtain the building facade images and the corresponding semantic label maps, and perform data preprocessing; The model construction module is used to construct the neural network Multi-DRCNet that uses local and row-column direction repeated texture features for constraint, including the multi-layer dilated convolution module MD-Module and the multi-scale row-column convolution module MRC-Module; The overall framework structure of the Multi-DRCNet is the same as that of the U-Net, adopting an encoder-decoder structure. In the encoding stage, a pre-trained basic feature extraction network is used. In the decoding stage, the ordinary convolution layers in the upsampling of the U-Net are replaced with the multi-scale row-column convolution module MRC-Module, and the MD-Module is added in the middle of the encoding and decoding; The multi-layer dilated convolution module MD-Module includes multiple dilated convolutions, and the multi-scale row-column convolution module MRC-Module includes multiple convolutions in different directions; The input and output of the multi-layer dilated convolution module MD-Module are both feature maps of a specific size. The MD-Module includes 5 dilated convolution operations. The convolution kernels are square with a size of 3, but their dilation rates are different, which are 1, 2, 4, 8, and 16 respectively. Finally, the input of the MD-Module and the outputs of the five dilated convolution modules are connected. After adjusting the dimension of the output through a conventional convolution, it is output to the upsampling process and finally output; After each convolution in the MD-Module, there is batch normalization BN and ReLU activation; The training and evaluation module is used to train the Multi-DRCNet network using the public building facade dataset and perform accuracy evaluation; The parsing module is used to load the weights trained on the public dataset, fine-tune the trained neural network on the prepared data, and parse the building facade using the fine-tuned network.

9. An electronic device for parsing building facades using repetitive texture constraints, characterized in that, It includes: One or more processors; A storage device for storing one or more programs, which when executed by the one or more processors, cause the one or more processors to implement the method according to any one of claims 1 to 7.

Images