Urban 360-degree panoramic building segmentation method and system

Abstract

The present application relates to the technical field of computer vision and image processing, and particularly relates to a city 360-degree panoramic building segmentation method and system, which comprises the following steps: firstly, obtaining building image dataset; secondly, extracting non-distorted building features; thirdly, generating building images matched with input images based on a conditional diffusion model, and performing image segmentation based on a decoder network; finally, comparing the prediction results output by the decoder network with real segmentation masks, calculating a loss function, adjusting system parameters through an optimization algorithm, testing the trained panoramic building segmentation model to evaluate the segmentation effect and performance of the model in real panoramic scenes, which can not only improve the precision and quality of building image processing, but also significantly reduce processing time and cost, and can adapt to various complex scenes and image types, and is widely used in various industries.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and image processing, and in particular to a city 360-degree panoramic building segmentation method and system. BACKGROUND

[0002] With the acceleration of urbanization, accurate city building segmentation has become a key technology in the fields of intelligent city management, automatic driving navigation, virtual reality and augmented reality. However, traditional building segmentation methods are mainly based on two-dimensional plane images, which have obvious limitations in processing city 360-degree panoramic images. Especially due to the wide-angle characteristics of panoramic images, severe geometric distortion is easily generated, thereby affecting the accuracy and reliability of segmentation.

[0003] As a generative model, the diffusion model has shown great potential in image synthesis and enhancement, but its application in building segmentation has not been thoroughly explored. Especially in the context of 360-degree panoramic images, how to effectively use the diffusion model for high-precision building segmentation is still a technical problem to be solved.

[0004] In addition, traditional panoramic image feature extraction methods are usually difficult to handle the distortion problem in the image, resulting in inaccurate building features. In order to solve the above problems, there is an urgent need for a technology that can extract non-distorted features in 360-degree panoramic images, and at the same time combine the diffusion model to achieve an efficient and accurate building segmentation method. SUMMARY

[0005] The present application aims to solve the problems of insufficient building image segmentation accuracy and panoramic image distortion, and provides a city 360-degree panoramic building segmentation method and system. By constructing a foreground and background knowledge base, combining conditional diffusion model, Transformer and deformable convolution (DeformConv) technology, high-quality building image segmentation and non-distorted feature extraction are achieved.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] In a first aspect, the present application provides a city 360-degree panoramic building segmentation method, which specifically includes the following steps:

[0008] S1, obtaining building image dataset: obtaining a 360-degree panoramic building image dataset, the dataset being a Cityscapes, ADE20K or Matterport3D panoramic segmentation dataset;

[0009] S2, extracting distortion-free building features: using a Transformer network to extract the global features of the building, and combining the deformable convolution technology to adjust the geometric structure of the extracted building global features, eliminate the distortion problem caused by the projection of the panoramic image, and obtain the distortion-free building features;

[0010] S3, generating images based on a conditional diffusion model: using a conditional diffusion model to add noise to the distortion-free building features obtained in step S2 and generate a building image that matches the input image;

[0011] S4, image segmentation based on a decoder network: inputting the building image obtained in step S3 into the decoder network through the distortion-free building features extracted in step S2 for accurate segmentation;

[0012] S5, loss calculation and optimization: comparing the prediction results output by the decoder network with the real segmentation mask, calculating the loss function, and adjusting the system parameters through an optimization algorithm to obtain a trained panoramic building segmentation model;

[0013] S6, model testing: testing the trained panoramic building segmentation model to evaluate the segmentation effect and performance of the model in real panoramic scenarios.

[0014] As a further technical solution of the present application, the specific process of step S2 is as follows:

[0015] S21, given an input 360-degree panoramic building image , the image is represented as a matrix form of image data , where H is the image height, W is the image width, and C is the number of color channels, the input image is divided into multiple non-overlapping patches, each patch is flattened into a vector and mapped to a high-dimensional space, specifically: set the size of the patch to , then the input image is divided into patches, and the flattened vector of each patch is represented as , and these vectors are mapped to a feature space with dimension D through linear transformation: , where is the linear transformation matrix, is the bias term, is the sequence length, and the initial feature representation is obtained.

[0016] S22, initial feature representation The input is fed into a Transformer network for processing. The Transformer consists of multiple layers of self-attention mechanisms and feedforward neural networks (FFN). The output of each Transformer layer undergoes deformable convolution to adjust its local geometric features. Specifically, for the L-th layer of the Transformer, the features represented by the (L-1)-th layer are first processed... Through self-attention mechanism Calculate the updated feature representation : Then the updated feature representation Local geometric adjustments are made using deformable convolution for each location. The output of deformable convolution is represented as: ,in, It is the first The kernel weights of layer deformable convolutions, It is the sampling position index of the convolution kernel. This represents the number of sampling locations; It is the position offset of the standard convolution kernel; It is the first The offsets learned by the DeformConv layer are used to obtain the feature representation after deformable convolution. Then The input is fed into a feedforward neural network (FFN) for further processing to obtain the output of the Lth layer: .

[0017] As a further technical solution of the present invention, the specific process of step S3 is as follows:

[0018] S31. Input building image The feature representation is converted using step S2. Image features It contains both global and local features of the image, along with the input segmentation mask. Encoded as a representation vector with the same dimensions as the feature representation. and with image features Together, these are input as conditions into the diffusion model, where image features... and the encoded segmentation mask Combining to form a condition vector : ;

[0019] S32, Diffusion model in feature space Adding noise, we get: ,in, is the feature representation at time step t; is the scaling coefficient at time step t; is the noise sampled from a standard normal distribution;

[0020] S33, reverse reasoning in the feature space, the model gradually removes the noise, and the feature representation is restored to a state closer to the original input feature, and this process is performed by the reverse diffusion model step: wherein: is the feature representation at time step t-1; is the feature representation at time step t; is the noise item predicted by the diffusion model at time step t, which is based on the input feature representation and the feature at the current time step ; this process gradually removes the noise from a preset time step T, and finally at time step t=0, a clear, high-quality final feature representation is restored, which matches the original input image feature;

[0021] S34, map the final feature representation back to the image space to generate the final building image : wherein, is a feature mapping function, implemented by a decoding network (up-sampling network).

[0022] As a further technical solution of the present application, the specific process of step S4 is:

[0023] S41, the generated building image is repeated step S2 to extract the distortion-free building feature to obtain the feature representation and input into the decoder network;

[0024] S42, the feature representation is processed by multiple layers of the decoder network, and the segmented building image is gradually restored and output to obtain the predicted building segmentation result : wherein, denotes the decoder network.

[0025] As a further technical solution of the present application, the loss function of step S5 is: wherein, is the segmentation loss, represents a specific loss calculation method, i.e. cross-entropy loss, is the predicted segmentation result, is the true segmentation mask.

[0026] As a further technical solution of the present invention, the specific process of step S6 is as follows: first, load the unseen 360-degree panoramic building image and the corresponding segmentation mask (ground truth) from the test dataset, and preprocess the image. The preprocessing includes removing noise from the image, adjusting the image resolution, and performing geometric correction so that the input data meets the requirements of the model. Then, input the preprocessed panoramic building image into the already trained segmentation model, and the model generates building segmentation results based on the input image.

[0027] Secondly, the present invention provides a 360-degree panoramic building segmentation system for cities, comprising:

[0028] The data acquisition module is used to acquire 360-degree panoramic images of buildings;

[0029] The feature extraction module is used to extract distortion-free building features from panoramic images;

[0030] The image generation module is used to generate matching building images based on building features and segmentation masks;

[0031] The image segmentation module is used to input the generated building images into the decoder network for accurate segmentation;

[0032] The optimization module is used to calculate the loss function between the segmentation result and the real mask, and to adjust the system parameters through the optimization algorithm;

[0033] The model testing module is used to perform performance testing on the optimized model.

[0034] Compared with existing technologies, this invention achieves high-precision building image segmentation and generation in complex scenes through a conditional diffusion model and distortion-free feature extraction technology for 360-degree panoramic images. Specifically, this invention constructs a conditional diffusion model that can automatically extract information from large-scale and diverse building image data. During image generation, the conditional diffusion model gradually reduces noise in the image, making the generated building images more accurate and clear in detail. Furthermore, by using a Transformer network to extract global features from the panoramic image and combining it with Deformable Convolution technology for distortion-free feature extraction, the geometric structure of the building can be effectively preserved from distortion in 360-degree panoramic images, ensuring the consistency and realism of the building's form from different viewpoints. This not only improves the accuracy and quality of building image processing but also significantly reduces processing time and cost. It can adapt to various complex scenes and image types, and has extremely high promotional value in practical applications. It can be widely applied in various industries, thus providing strong support for the development of building image processing technology. Attached Figure Description

[0035] Figure 1 This is a schematic diagram illustrating the workflow of the urban 360-degree panoramic building segmentation method provided by the present invention.

[0036] Figure 2 The structural block diagram of the urban 360-degree panoramic building segmentation system provided by the present invention. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the scope of the invention is not limited in any way.

[0038] Example 1: As Figure 1 As shown, this embodiment provides a method for 360-degree panoramic building segmentation in a city, specifically including the following steps:

[0039] S1. Obtain the building image dataset: Obtain a 360-degree panoramic building image dataset. In this embodiment, the building image data sequence and corresponding background dataset are obtained from the database via a wired or wireless connection. The building image dataset stored in the database is represented as follows. Each image , represented as a three-dimensional tensor (height H, width W, number of color channels C). As an example, building image datasets can be commonly used panoramic segmentation datasets, that is, labeled multi-view, multi-lighting image datasets of buildings and their environment. Such datasets can include commonly used panoramic segmentation datasets such as Cityscapes, ADE20K, or Matterport3D. These datasets provide rich scene diversity and annotation information, and are suitable for feature extraction and segmentation tasks of panoramic images.

[0040] S2. Extracting Distortion-Free Building Features: A Transformer network is used to extract global features of the building. Combined with Deformable Convolution (DeformConv) technology, the extracted global features are geometrically adjusted to eliminate distortion caused by panoramic image projection, resulting in distortion-free building features. Specifically, the Transformer network is first used to extract global features from the 360-degree panoramic image to capture the overall structure and shape of the building. Then, Deformable Convolution (DeformConv) technology is used to adjust the geometric structure of the extracted building features, eliminating distortion caused by panoramic image projection. The resulting distortion-free building features are used for subsequent image segmentation operations.

[0041] S21. Given an input 360-degree panoramic image. The image is represented as a matrix of image data. Where H is the image height, W is the image width, and C is the number of color channels. The input to a Transformer network is typically a sequence of feature vectors. Where N represents the sequence length and D represents the dimension of each vector.

[0042] To adapt to Transformer, first convert the input image... The data is divided into multiple non-overlapping patches. Each patch is flattened into a vector and mapped to a high-dimensional space. The size of each patch is set to [value missing]. Then the input image can be divided into There are several small blocks, and the flattened vector representation of each block is... These vectors are mapped to a feature space of dimension D through a linear transformation: ,in, It is a linear transformation matrix. It is a bias term. Given the sequence length, we obtain the initial feature representation received by the Transformer. ;

[0043] S22. To eliminate distortion caused by panoramic image projection, DeformableConvolution (DeformConv) technology is used. Unlike standard convolution, DeformConv allows the position of the convolution kernel to be dynamically adjusted on the feature map to adapt to geometric changes in the image, thus capturing the geometric structure of buildings more flexibly. In this embodiment, the offset between the ERP coordinates and the projected image coordinates of each pixel is first calculated. These offsets represent the geometric deformation caused by projection distortion and are applied as the convolution kernel offset in DeformableConvolution. Specifically, given the initial feature representation of the input... The data is input into a Transformer network for processing. During convolution operations, each convolutional kernel dynamically adjusts its sampling position according to its corresponding offset to adapt to the geometric deformation of the image and eliminate distortion caused by projection distortion in panoramic images. In this way, Deformable Convolution adaptively adjusts the position of the convolutional kernel at each feature location to ensure that the extracted building features are accurate and distortion-free. Typically, a Transformer consists of multiple layers of self-attention mechanisms and feedforward neural networks (FFNs). In the embodiments of this invention, the output of each Transformer layer is processed by DeformConv to adjust its local geometric features. For the Lth layer of the Transformer, the processing is as follows: The feature representation input to the Lth layer... The updated feature representation is calculated using a self-attention mechanism. : Then, the feature representation updated by the self-attention mechanism is... Local geometry adjustments are performed using DeformConv for each location. The output of DeformConv is represented as: ,in: These are the convolutional kernel weights of the Lth layer, DeformConv. It is the sampling position index of the convolution kernel. This represents the number of sampling locations; It is the position offset of the standard convolution kernel; This is the offset learned by DeformConv at layer L; next, the feature representation after DeformConv processing... The input is fed into a feedforward neural network (FFN) for further processing to obtain the output of the Lth layer: .

[0044] This adjustment process ensures that the building features extracted from the 360-degree panoramic image remain spatially consistent, eliminating distortion problems caused by projection in the panoramic image. Finally, the output distortion-free feature map is used for subsequent building image segmentation operations, ensuring the accuracy and stability of the segmentation. By combining the Transformer network and DeformableConvolution technology, this embodiment can effectively extract distortion-free global features of buildings from 360-degree panoramic images, significantly improving the effect and quality of image segmentation.

[0045] S3. Image generation based on conditional diffusion model: The distortion-free building features obtained in step S2 are subjected to noise addition and reverse inference using the conditional diffusion model to generate a building image that matches the input image; specifically:

[0046] First, the input building image. Converted into feature representation This feature representation, generated by a Transformer network modified by DeformConv, contains both global and local features of the image: Meanwhile, the input segmentation mask Encoded as a representation vector with the same dimensions as the feature representation. and with image features Together, they are input as conditions into the diffusion model: , For image features and the encoded segmentation mask combination The resulting conditional vector;

[0047] Next, the diffusion model performs a feature space analysis. Adding noise allows us to utilize higher-level feature representations and masking information, making the generation process more stable and flexible. ,in: It is the feature representation at time step t; It is the scaling factor for time step t; It is noise sampled from a standard normal distribution;

[0048] When performing backward reasoning in the feature space, the model progressively removes noise and refines the feature representation. The process of restoring the data to a state closer to the original input features is carried out through the steps of the backdiffusion model: ,in: It is the feature representation at time step t-1; It is the feature representation at time step t; It is the noise term predicted by the diffusion model at time step t, which is based on the feature representation of the input. and the characteristics of the current time step Specifically, the goal of this formula is to gradually remove noise starting from a preset time step T. In this embodiment, T=50, and finally, at time step t=0, a clear, high-quality feature representation is restored. This feature indicates a match with the original input image features;

[0049] Once the reverse reasoning process is complete, the final feature representation is obtained. It will be mapped back to image space to generate the final building image. This process is achieved through a reverse feature mapping function, which transforms the high-dimensional feature representation back into the pixel-level representation of the original image: ,in, It is a feature mapping function, usually implemented by a decoding network (upsampling network); It is the final feature representation; It is a generated image of a building.

[0050] In this way, the conditional diffusion model can add and remove noise in the feature space to generate building images that match the input conditions.

[0051] S4. Image segmentation based on decoder network: The building image obtained in step S3 is processed by extracting distortion-free building features from the building image in step S2, and then input into the decoder network for accurate segmentation; specifically: the generated building image... The input is used to extract distortion-free building features in step 2 to obtain the feature representation. and will The input is fed into the decoder network. After multi-layer processing by the decoder network, the segmented building image is gradually recovered and output. : ,in, Indicates the decoder network, It is the predicted result of building segmentation.

[0052] S5. Loss Calculation and Optimization: Compare the predicted results output by the decoder network with the actual segmentation mask, calculate the loss function, and adjust the system parameters through optimization algorithms to obtain a trained panoramic building segmentation model. To optimize the segmentation effect, the predicted results output by the decoder network need to be optimized. Compared to the real segmentation mask Compare and calculate the loss function The loss function is defined as follows: ,in, It is a segmentation loss. This represents a specific loss calculation method (cross-entropy loss). It is the predicted segmentation result. It is the actual segmentation mask.

[0053] S6. Model Testing: The trained panoramic building segmentation model is tested to evaluate its segmentation performance in real panoramic scenes. First, unseen 360-degree panoramic building images and their corresponding segmentation masks (ground truth) are loaded from the test dataset. To ensure the quality and consistency of the input data, the images need to be preprocessed. This includes removing noise, adjusting image resolution, and performing geometric correction to ensure the input data meets the model's requirements. The preprocessed panoramic building images are then input into the trained segmentation model. The model will generate building segmentation results based on the input images. In this process, the model uses the features and weights learned above to infer the probability that each pixel in the panoramic image belongs to a building.

[0054] The conditional diffusion model provided in this embodiment possesses stepwise reasoning capabilities, effectively controlling image quality during image generation and avoiding blurring and distortion problems common in traditional generation methods. Simultaneously, the combination of Transformer and Deformable Convolution ensures more accurate extraction of building features from panoramic images, maintaining image clarity and distortion-free characteristics even under large-scale viewpoint changes. This combination of techniques significantly improves the quality of image segmentation and generation, particularly in building detail and edge processing. It enables more efficient image segmentation and generation when handling complex building scenes and exhibits greater adaptability. Whether in urban planning, architectural design, virtual reality environment construction, or other fields requiring high-precision building image generation, the techniques disclosed herein can significantly improve the effectiveness and efficiency of image processing.

[0055] Example 2: As Figure 2 As shown, this embodiment provides a 360-degree panoramic building segmentation system for cities, including:

[0056] The data acquisition module is used to acquire 360-degree panoramic images of buildings;

[0057] The feature extraction module is used to extract distortion-free building features from panoramic images;

[0058] The image generation module is used to generate matching building images based on building features and segmentation masks;

[0059] The image segmentation module is used to input the generated building images into the decoder network for accurate segmentation;

[0060] The optimization module is used to calculate the loss function between the segmentation result and the real mask, and to adjust the system parameters through the optimization algorithm;

[0061] The model testing module is used to perform performance testing on the optimized model.

[0062] In further embodiments, the following is also provided:

[0063] An electronic device includes a memory and a processor, and carries computer instructions that run on the processor. When executed by the processor, these instructions can precisely implement the method steps described in Embodiment 1. For the sake of simplicity, the detailed method description will not be repeated here.

[0064] It should be clarified that the processor in this embodiment is not limited to a central processing unit (CPU). It can also be a general-purpose processor core, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices (PLDs), discrete gate circuits or transistor logic components, discrete hardware modules, etc. Furthermore, the general-purpose processor can be a microprocessor (MPU) or any conventional computing core.

[0065] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0066] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0067] The method in Example 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0068] The algorithms, systems, and network structures not described in detail in this invention are all general technologies in the field.

[0069] Those skilled in the art will recognize that the units, i.e., algorithm steps, of the various examples described in connection with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

Claims

1. A method for segmenting buildings in a 360-degree panoramic view of a city, characterized in that, Specifically, the following steps are included: S1. Obtain building image dataset: Obtain a 360-degree panoramic building image dataset, such as Cityscapes, ADE20K, or Matterport3D panoramic segmentation dataset. S2. Extract distortion-free building features: Use the Transformer network to extract global features of the building image, and combine deformable convolution technology to adjust the geometric structure of the extracted global building features to eliminate the distortion problem caused by panoramic image projection and obtain distortion-free building features. The specific process is as follows: S21. Given an input 360-degree panoramic image of a building. Image data is represented as a matrix. Where H is the image height, W is the image width, and C is the number of color channels, the input image is... The data is divided into multiple non-overlapping blocks. Each block is flattened into a vector and mapped to a high-dimensional space. Specifically, the size of each block is set to... The input image is then divided into There are several small blocks, and the flattened vector representation of each block is... These vectors are mapped to a feature space of dimension D through a linear transformation: ,in, It is a linear transformation matrix. It is a bias term. Given the sequence length, the initial feature representation is obtained. ; S22, Initial Feature Representation The input is fed into a Transformer network for processing. The Transformer consists of multiple layers of self-attention mechanisms and feedforward neural networks stacked together. The output of each Transformer layer undergoes deformable convolution to adjust its local geometric features. Specifically, for the L-th layer of the Transformer, the features represented by the (L-1)-th layer are first processed... Through self-attention mechanism Calculate the updated feature representation : Then the updated feature representation Local geometric adjustments are made using deformable convolution for each location. The output of deformable convolution is represented as: ,in, It is the first The kernel weights of layer deformable convolutions, It is the sampling position index of the convolution kernel. This represents the number of sampling locations; It is the positional offset of the standard convolution kernel; It is the first The offsets learned by the DeformConv layer are used to obtain the feature representation after deformable convolution. Then The input is fed into a feedforward neural network (FFN) for further processing to obtain the output of the Lth layer: ; S3. Generate images based on conditional diffusion model: Use the conditional diffusion model to add noise to the distortion-free building features obtained in step S2 and perform reverse reasoning to generate building images that match the input image. S4. Image segmentation based on decoder network: Extract the distortion-free building features from the building image obtained in step S3 through step S2, and input it into the decoder network for accurate segmentation. S5. Loss Calculation and Optimization: Compare the predicted results output by the decoder network with the actual segmentation mask, calculate the loss function, and adjust the system parameters through optimization algorithms to obtain the trained panoramic building segmentation model. S6. Model Testing: Test the trained panoramic building segmentation model to evaluate its segmentation effect and performance in real panoramic scenes.

2. The urban 360-degree panoramic building segmentation method according to claim 1, characterized in that, The specific process of step S3 is as follows: S31. Input building image The feature representation is converted using step S2. Image features It contains both global and local features of the image, along with the input segmentation mask. Encoded as a representation vector with the same dimensions as the feature representation. and with image features Together, these are input as conditions into the diffusion model, where image features... and the encoded segmentation mask Combining to form a condition vector : ; S32, Diffusion model in feature space Adding noise, we get: ,in, It is a feature representation at time step t; It is the scaling factor for time step t; It is noise sampled from a standard normal distribution; S33. Performing reverse reasoning in the feature space, the model progressively removes noise and represents the features. The process of restoring the data to a state closer to the original input features is carried out through the steps of the backdiffusion model: ,in: It is the feature representation at time step t-1; It is the feature representation at time step t; It is the noise term predicted by the diffusion model at time step t, which is based on the feature representation of the input. and the characteristics of the current time step The process begins by gradually removing noise at a preset time step T, and finally restores a clear, high-quality final feature representation at time step t=0. This feature indicates a match with the original input image features; S34. Represent the final features Mapping back to image space generates the final building image. : ,in, It is a feature mapping function, implemented by a decoding network, which uses an upsampling network.

3. The urban 360-degree panoramic building segmentation method according to claim 2, characterized in that, The specific process of step S4 is as follows: S41. Generate the building image Repeat step S2 to extract the features of the distortion-free buildings and obtain the feature representation. And input it into the decoder network; S42, Feature Representation After multi-layer processing by the decoder network, the segmented building image is gradually recovered and output, resulting in the predicted building segmentation result. : ,in, This represents the decoder network.

4. The urban 360-degree panoramic building segmentation method according to claim 3, characterized in that, The loss function described in step S5 is: ,in, It is a segmentation loss. The representative loss calculation method is cross-entropy loss. It is the predicted segmentation result. It is the actual segmentation mask.

5. The urban 360-degree panoramic building segmentation method according to claim 4, characterized in that, The specific process of step S6 is as follows: First, load the unseen 360-degree panoramic building image and the corresponding segmentation mask from the test dataset, and preprocess the image. The preprocessing includes removing noise from the image, adjusting the image resolution, and performing geometric correction to make the input data meet the requirements of the model. Then, input the preprocessed panoramic building image into the already trained segmentation model. The model generates building segmentation results based on the input image.

6. A 360-degree panoramic building segmentation system for cities, characterized in that, Implementing the method as described in any one of claims 1-5 includes: The data acquisition module is used to acquire 360-degree panoramic images of buildings; The feature extraction module is used to extract distortion-free building features from 360-degree panoramic building images; The image generation module is used to generate matching building images based on building features and segmentation masks; The image segmentation module is used to input the generated building images into the decoder network for accurate segmentation; The optimization module is used to calculate the loss function between the segmentation result and the real mask, and to adjust the system parameters through the optimization algorithm; The model testing module is used to perform performance testing on the optimized model.

Images