Visual comfort assessment method and device for virtual reality building

Abstract

The invention relates to the technical field of building design, and discloses a visual comfort assessment method and device for a virtual reality building, and the method comprises the steps: collecting building image data; labeling the building image data; performing sample generation by adopting an adversarial network algorithm generated based on building image data distribution, and expanding a building image data set; a federal learning architecture is adopted, and local building image data samples are utilized to train a visual comfort evaluation model on a plurality of servers; and the model on each server is trained on the respective building image data, and performs parameter updating interaction with the central visual comfort evaluation model to obtain the visual comfort evaluation model. According to the method, data privacy protection is realized by adopting a federated learning architecture, training of the visual comfort evaluation model is completed at the client through a distributed training mode, only model parameters are exchanged instead of original building image data, and the problem of privacy disclosure caused by sharing sensitive data among multiple devices is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of architectural design, and particularly relates to a method and device for evaluating visual comfort of virtual reality buildings. Background Art

[0002] The evaluation of visual comfort in architectural design is a key task, which is directly related to the visual experience and psychological feelings of users in the building environment. With the application of virtual reality technology, designers can simulate building scenes in a virtual environment, so as to evaluate and optimize the visual effects of buildings in the design stage.

[0003] However, since the evaluation of visual comfort involves a large amount of complex data processing, feature extraction and analysis, this task faces many technical challenges.

[0004] For example, in the task of evaluating the visual comfort of buildings, the traditional centralized model training method needs to aggregate the original data to a central server, which may lead to the leakage of sensitive building image data and is difficult to meet the requirements of privacy protection; moreover, the collection of building image data depends on the actual scene, with high cost and low efficiency, and it is difficult to ensure the comprehensiveness and diversity of the data, resulting in insufficient training samples. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention proposes a method and device for evaluating visual comfort of virtual reality buildings to solve the above technical problems.

[0006] In a first aspect, a method for evaluating visual comfort of virtual reality buildings is provided, including:

[0007] Collecting building image data;

[0008] Annotating the building image data, and according to the visual feedback and experience of the user in the virtual environment, annotating the image as one of the comfort, medium or discomfort levels;

[0009] Using an adversarial network algorithm generated based on the distribution of building image data for sample generation to expand the building image data set;

[0010] Adopting a federated learning architecture, on multiple decentralized devices or servers, using local building image data samples to train a visual comfort evaluation model;

[0011] The local visual comfort evaluation model on each device or server is trained on the building image data held by itself, and parameter update interaction is carried out with the central visual comfort evaluation model. Through multiple iterations, the finally trained visual comfort evaluation model is obtained.

[0012] Furthermore, the acquisition method of the building image data uses high-precision virtual reality helmets and environmental tracking devices in a virtual reality environment, and the acquisition content covers perspectives from different angles inside the building.

[0013] Furthermore, the sample generation is performed using an adversarial network algorithm generated based on the distribution of building image data, including:

[0014] Initialize the generator and discriminator of the generative adversarial network. Among them, the generator inputs a noise vector and outputs the generated sample, and the discriminator is a binary classification network that discriminates whether the sample comes from real building image data;

[0015] By analyzing the distribution characteristics of the building image training data, the generator realizes the transition from the source data distribution of the building image to the target data distribution;

[0016] Gradually optimize the generation quality of the generator through adversarial training, and use a loss function to constrain the training processes of the generator and the discriminator;

[0017] Repeat the steps of realizing the transition from the source data distribution of the building image to the target data distribution by analyzing the distribution characteristics of the building image training data; gradually optimizing the generation quality of the generator through adversarial training, and using a loss function to constrain the training processes of the generator and the discriminator until the generated sample meets the preset conditions.

[0018] Furthermore, the parameter update interaction between the local visual comfort evaluation model and the central visual comfort evaluation model includes:

[0019] After the local visual comfort evaluation model is trained, send the updated model parameters to the central visual comfort evaluation model;

[0020] The central visual comfort evaluation model aggregates the updated parameters from each local visual comfort evaluation model for global update;

[0021] Send the globally updated parameters back to each local visual comfort evaluation model for the next round of iterative training;

[0022] Through multiple iterations, the parameter update interaction between the local visual comfort evaluation model and the central visual comfort evaluation model finally obtains a globally optimized visual comfort evaluation model.

[0023] Furthermore, it also includes:

[0024] The local visual comfort evaluation model uses a fully connected neural network to extract the features of the virtual building image, and uses a quantum state simulation optimization algorithm to optimize the parameters to obtain optimized neural network parameters.

[0025] Second aspect, there is provided a visual comfort evaluation device for virtual reality architecture, based on the visual comfort evaluation method for virtual reality architecture described in any one of the foregoing, including:

[0026] Multiple clients, each client holds local building image data samples and trains a visual comfort evaluation model locally;

[0027] The building image data of the client is collected from a simulated building scene in a virtual reality environment, and the simulated building scene is constructed by professional designers in virtual reality software according to real building parameters;

[0028] A central server, which is used to aggregate the updated parameters of the local visual comfort evaluation models from each client, perform global updates, and send the globally updated parameters back to each client.

[0029] Furthermore, the client further includes a data acquisition module, and the data acquisition module uses high-precision virtual reality headsets and environmental tracking devices in the virtual reality environment to collect building image data.

[0030] Furthermore, the client includes an annotation module for annotating building image data, and a generation module for generating samples using an adversarial network algorithm based on the distribution of building image data;

[0031] The annotation module annotates the image as one of the comfort, medium or discomfort levels according to the visual feedback and experience of the user in the virtual environment.

[0032] Furthermore, it further includes a classifier, which is trained through the building image data after feature extraction and is used for visual comfort evaluation, and the evaluation categories include: comfort, medium, discomfort.

[0033] The invention adopting the above technical solution has the following advantages:

[0034] 1. The present invention adopts a federated learning architecture to achieve data privacy protection. Through a distributed training mode, the training of the visual comfort evaluation model is completed locally on the client, and only the model parameters are exchanged instead of the original building image data, solving the problem of potential privacy leakage caused by sharing sensitive data among multiple devices.

[0035] 2. The present invention uses a generative adversarial network for data augmentation, combines the data distribution transition and cross-domain dynamic alignment mechanisms to generate more realistic and diverse building image data, solves the problems of difficult collection and insufficient samples of building image data, and at the same time improves the phenomenon of unbalanced data categories.

[0036] 3. Based on the traditional generative adversarial network, the present invention adds mechanisms such as gradient penalty terms and KL divergence, improving the stability and diversity of data generation, avoiding the mode collapse phenomenon, and ensuring the quality of the generated data.

[0037] 4. The present invention uses quantum state simulation to optimize neural network parameters, explores the parameter space by utilizing the mechanisms of quantum state superposition and collapse, overcomes the problem that traditional gradient optimization algorithms may fall into local optima, and provides the feature extraction ability of the visual comfort evaluation model for complex building images. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the specific embodiments of the present invention, the drawings required for use in the specific embodiments will be briefly introduced below. In all the drawings, the elements or parts are not necessarily drawn to actual scale.

[0039] Figure 1 It is a flowchart of a visual comfort evaluation method for virtual reality buildings according to the present invention;

[0040] Figure 2 It is a federated learning architecture diagram in a visual comfort evaluation method for virtual reality buildings according to the present invention;

[0041] Figure 3 It is a comparison chart of FID scores of different generation algorithms in the adversarial network algorithm of a visual comfort evaluation method for virtual reality buildings according to the present invention;

[0042] Figure 4 It is a comparison chart of the standard deviation of the class distribution of samples generated by different methods in the adversarial network algorithm of a visual comfort evaluation method for virtual reality buildings according to the present invention;

[0043] Figure 5 It is a change chart of the loss function during the training process of the adversarial network algorithm in a visual comfort evaluation method for virtual reality buildings according to the present invention;

[0044] Figure 6 It is a sensitivity analysis chart of hyperparameters of the adversarial network algorithm in a visual comfort evaluation method for virtual reality buildings according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The embodiments of the technical solutions of the present invention will be described in detail below with reference to the drawings. The following embodiments are only used to illustrate the technical solutions of the present invention more clearly, so they are only examples and cannot be used to limit the protection scope of the present invention.

[0046] As Figures 1 to 6 shown, a visual comfort evaluation method for virtual reality buildings according to the present invention includes:

[0047] Step S01: Collect building image data;

[0048] Step S02: Annotate the building image data. According to the visual feedback and experience of the user in the virtual environment, label the image as one of the comfort, medium, or discomfort levels;

[0049] Step S03: Use an adversarial network algorithm generated based on the distribution of building image data for sample generation to expand the building image dataset;

[0050] Step S04: Adopt a federated learning architecture. On multiple decentralized devices or servers, use local building image data samples to train a visual comfort evaluation model;

[0051] Step S05: The local visual comfort evaluation models on each device or server are trained on the building image data they hold and interact with the central visual comfort evaluation model for parameter update. Through multiple iterations, obtain the finally trained visual comfort evaluation model.

[0052] In some embodiments, the building image data is collected by using a high-precision virtual reality headset and an environment tracking device in a virtual reality environment. The collected content covers perspectives from different angles inside the building.

[0053] Specifically, the federated learning architecture proposed by the present invention has multiple clients, labeled as Client 0, Client 1, Client 2, Client 3, etc. These clients represent independent nodes in the federated learning network, and each client has its own local building image data;

[0054] The source of the building image data for each client is a simulated building scene in a virtual reality environment. The simulated building scene is constructed by professional designers in virtual reality software according to real building parameters. The building image data is collected by using a high-precision virtual reality headset and an environment tracking device in a virtual reality environment. The collected content covers perspectives from different angles inside the building, including the entrance view, the view of the main indoor activity area, and the view of the outdoor landscape. The building image data is saved in the high-resolution PNG format.

[0055] Specifically, the building image data has a resolution of 1920x1080 pixels, and the building image data is annotated. The annotation method is manual annotation. According to the visual feedback and experience of the user in the virtual environment, label the image as one of the following three levels:

[0056] Comfortable: It means that in this environment, the visual experience gives the user a feeling of relaxation and pleasure;

[0057] Medium: It means that the visual experience of the environment is acceptable, but there are some deficiencies, and it may feel slightly depressing or not enough to cause discomfort to the user;

[0058] Uncomfortable: It means that the visual elements in the environment cause obvious discomfort, such as excessive light, improper color matching, or unreasonable spatial layout.

[0059] In some embodiments, an adversarial network algorithm based on the distribution of building image data is used for sample generation, including:

[0060] Initialize the generator and discriminator of the generative adversarial network. Among them, the generator inputs a noise vector and outputs the generated sample, and the discriminator is a binary classification network that discriminates whether the sample comes from real building image data;

[0061] By analyzing the distribution characteristics of the building image training data, the generator realizes the transition from the source data distribution of the building image to the target data distribution;

[0062] Gradually optimize the generation quality of the generator through adversarial training, and use a loss function to constrain the training processes of the generator and the discriminator;

[0063] Repeat the steps of realizing the transition from the source data distribution of the building image to the target data distribution by analyzing the distribution characteristics of the building image training data; gradually optimizing the generation quality of the generator through adversarial training, and using a loss function to constrain the training processes of the generator and the discriminator until the generated sample meets the preset conditions.

[0064] Specifically, in the task of the present invention, the acquisition, annotation, and preprocessing of the building image training data are time-consuming and laborious, and the lack of training samples easily leads to poor generalization ability of the model, and at the same time affects the accuracy of the model;

[0065] The present invention uses an adversarial network algorithm based on the distribution of building image data for sample generation, and then realizes the expansion of building image data;

[0066] Through the generalization ability of the model, generate more building image data samples helpful for training to expand the building image data set.

[0067] Specifically, the training process of the adversarial network algorithm based on data distribution is as follows:

[0068] 1. Initialize the generator and discriminator of the generative adversarial network. The generator inputs a noise vector and outputs the generated sample, and the discriminator is a binary classification network that discriminates whether the sample comes from real building image data, expressed as:

[0069]

[0070]

[0071] In the formula, G c () is the generator function, Fake samples generated by the generator; D c () is the discriminator function that outputs the authenticity probability based on the input building image data; x c is the input building image data for the discriminator; Sig() is the Sigmoid activation function; z c is the input vector sampled from the noise distribution; θ c are the parameters of the generator; are the parameters of the discriminator. The parameters of the discriminator include the weights and biases of the discriminator, W c are the weights of the discriminator, b c are the biases of the discriminator.

[0072] 2. By analyzing the distribution characteristics of the building image training data, the generator realizes the transition from the source data distribution of the building images to the target data distribution of the building images. The generator uses the distribution transition mechanism to generate new samples, making them closer to the target data distribution of the building images, and gradually optimizing the generation quality through adversarial training. The loss function constraint of this process is expressed as:

[0073]

[0074] In the formula, is the adversarial loss function for data distribution transition; p data (x c ) is the real building image data distribution; G c represents the generator; D c represents the discriminator; represents expectation; ~ means subject to a specific distribution; p z (z c ) is the noise distribution; ∥∥ is the L2 norm; λ4 is the weight of the gradient penalty term, which restricts the power of the discriminator through the gradient penalty mechanism, balances the learning ability of the generator, and avoids the mode collapse problem; is the gradient of the discriminator with respect to the input sample.

[0075] Preferably, λ4 is set to 0.1.

[0076] 3. To alleviate the class imbalance problem, cross-domain dynamic alignment of the unbalanced distribution is performed. The generator adopts a cross-domain dynamic alignment strategy, dynamically adjusts the strategy of generating samples according to the class distribution characteristics, generates more minority class samples, and thus realizes class balance. The loss function constraint of this process is expressed as:

[0077]

[0078] In the formula, is the cross-domain alignment loss function; λ1 is the weight of the first adversarial loss term; λ2 is the weight of the second adversarial loss term; pgen (x c ) is the distribution for generating samples; p target (x c ) is the distribution of the target samples; is the KL divergence function. By adopting the KL divergence and the multi-dimensional feature alignment mechanism, the comprehensive alignment of the class distribution in the generated building image data is realized; γ cs is the parameter for controlling the weight of the alignment term.

[0079] Preferably, γ cs is set to 0.2, λ1 is set to 0.1, and λ2 is set to 0.2.

[0080] 4. Adversarially train the generator and the discriminator. The generator deceives the discriminator by optimizing its own loss function to generate more realistic samples, while the discriminator improves its ability to identify fake samples by optimizing its loss function. The two are optimized through adversarial games to achieve the generation of high-quality building image data. The loss function constraint of this process is expressed as:

[0081]

[0082] In the formula, is the adversarial loss function; λ3 is the weight of the gradient regularization term; ) is the gradient of the adversarial loss function of the data distribution transition with respect to the generator parameters.

[0083] Preferably, λ3 is set to 0.5.

[0084] 5. The parameters of the generator and the discriminator are updated by optimizing the total loss function to reach the optimal state of adversarial training. In the total loss function, weighted terms of distribution transition and class alignment are adopted to improve the distribution effectiveness and class balance of the generated samples. The parameters of the generator and the discriminator are continuously optimized through backpropagation, and finally high-quality samples are generated to improve the class balance, which is expressed as:

[0085]

[0086] In the formula, is the total loss of the generative adversarial network; ← is the parameter update operation; η c is the learning rate of the generative adversarial network; represents the gradient with respect to the generator parameters; represents the gradient with respect to the discriminator parameters.

[0087] Preferably, η c is set to 0.01.

[0088] 6. Repeat the above steps iteratively until the preset stop iteration condition is satisfied, which means the model training is completed.

[0089] As Figure 3 shown, this utility aims to verify the differences in the quality of building image generation among different generation algorithms by comparing the changing trends of FID scores during the training processes of traditional GAN, WGAN-GP, AC-GAN, and this method.

[0090] The experimental results show that the FID score of this method stabilizes at 15 ± 2 after 300 iterations, significantly lower than that of traditional GAN (45 ± 3), WGAN-GP (35 ± 2), and AC-GAN (25 ± 2).

[0091] The experimental results prove that through the data distribution transition mechanism and the cross-domain dynamic alignment strategy, the building images produced by this method are closer to the real data distribution in terms of texture details and global structure.

[0092] As Figure 4 shown, to evaluate the effect of the cross-domain dynamic alignment strategy on alleviating the problem of class imbalance, the standard deviations of the sample numbers of five types of buildings (residential, commercial, industrial, public facilities, historical buildings) generated by real data, traditional GAN, and this method were compared in the experiment.

[0093] The experimental results show that the standard deviation of the data produced by this method is 18.2%, close to that of real data (13.8%), and there is a significant improvement compared to traditional GAN (37.6%). Among them, the difference in the sample numbers of the historical building category decreased from about 40% in traditional GAN to about 19%, indicating that the dynamic alignment strategy effectively improves the generation ratios of industrial buildings and scarce historical building categories through KL divergence constraints and multi-dimensional feature matching.

[0094] As Figure 5 shown, by monitoring the dynamic changes of the loss functions of the generator and the discriminator, the experiment verified the influence of the gradient penalty mechanism on the training stability.

[0095] The experimental results show that the generator loss stabilizes at 0.5 ± 0.2 after 20 rounds, and the discriminator loss remains at 0.7 ± 0.1. There is no mode collapse phenomenon commonly seen in traditional GAN. Compared with the frequent loss oscillations in traditional GAN training, the smoothness of the loss curve of this method has been greatly improved, proving that the gradient penalty term effectively balances the adversarial training game process by constraining the L2 norm gradient of the discriminator.

[0096] As Figure 6 shown, to verify the robustness of the gradient penalty weight and the alignment weight, the experiment used heatmaps to show the changes in FID scores under different parameter combinations.

[0097] The experimental results show that when the gradient penalty weight ∈ [0.08, 0.12] and the alignment weight ∈ [0.18, 0.22], the FID score stabilizes in the optimized range of 15.5 - 17.2, which is significantly better than other parameter combinations, proving the scientificity of the parameter settings. At the same time, it reveals that the gradient penalty weight is more sensitive to the impact on the generation quality compared to the alignment weight.

[0098] In some embodiments, the preset stopping iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0099] After the training of the building image data augmentation model is completed, the trained building image data augmentation model is used to increase the number of samples.

[0100] In some embodiments, if the original collected samples are 800 and the building image data augmentation model augments and generates 200 samples, then the augmented building image data set contains 1000 samples.

[0101] In some embodiments, the parameter update interaction between the local visual comfort evaluation model and the central visual comfort evaluation model includes:

[0102] After the local visual comfort evaluation model is trained, the updated model parameters are sent to the central visual comfort evaluation model;

[0103] The central visual comfort evaluation model aggregates the updated parameters from each local visual comfort evaluation model for global update;

[0104] The globally updated parameters are sent back to each local visual comfort evaluation model for the next round of iterative training;

[0105] Through multiple iterations, the parameter update interaction between the local visual comfort evaluation model and the central visual comfort evaluation model finally obtains a globally optimized visual comfort evaluation model.

[0106] Specifically, each client has a local visual comfort evaluation model. These local visual comfort evaluation models are trained on the building image data of various clients, and there is an interaction between the local visual comfort evaluation model and the central visual comfort evaluation model, that is, the update flow of the visual comfort evaluation model parameters.

[0107] θ fed ′ represents the updated visual comfort evaluation model parameters after local training, and θ fed represents the parameters received from the central visual comfort evaluation model;

[0108] Specifically, the central visual comfort evaluation model aggregates the updates from the local visual comfort evaluation models. The update of the global visual comfort evaluation model is the way that the federated learning process aggregates individual updates to create a new improved global visual comfort evaluation model. After the update of the global visual comfort evaluation model is completed, the finally trained visual comfort evaluation model can be obtained.

[0109] The parameters θ of the visual comfort evaluation model fed0 to θ fed3 are exchanged between the local visual comfort evaluation model and the central visual comfort evaluation model. The exchange method allows the central visual comfort evaluation model to aggregate the updates and enables the local visual comfort evaluation model to receive the new, aggregated parameters;

[0110] Specifically, through multiple iterations, the local visual comfort evaluation model is trained, the updates are sent to the central visual comfort evaluation model, the global update is performed, and then the updated parameters are sent back to the local visual comfort evaluation model;

[0111] Based on this federated learning architecture, privacy protection of building image data can be achieved. During the training process, the original building image data is not shared between clients or with the central server, but only the parameters of the visual comfort evaluation model are exchanged for updates. That is, in the way that the central visual comfort evaluation model synchronizes the parameters to each client, θ fed ′ is assigned to each client.

[0112] In some embodiments, it further includes:

[0113] The local visual comfort evaluation model uses a fully connected neural network to extract the features of the virtual building image and uses a quantum state simulation optimization algorithm to optimize the parameters, obtaining the optimized neural network parameters.

[0114] Specifically, the visual comfort evaluation model of each client uses a 5-layer fully connected neural network to extract the features of the virtual building image and uses a quantum state simulation optimization algorithm to optimize the parameters during the neural network training process;

[0115] Inspired by the principles of state superposition and state collapse in quantum mechanics, regarding each neural network parameter as a quantum state that can exist in multiple states, through simulating the processes of quantum state superposition, measurement, and collapse, the quantum state simulation optimization algorithm explores the parameter space and optimizes the network performance. Compared with the traditional gradient optimization algorithm, the quantum state simulation optimization algorithm does not need to rely on gradient information, can jump out of the local optimal solution, and through the properties of quantum state superposition and collapse, widely explores the parameter space, thereby improving the convergence speed and enhancing the diversity of solutions, and has the potential to find the global optimal solution;

[0116] The neural network optimized based on quantum state simulation can effectively improve the performance and accuracy of the model on complex building image data, and is suitable for the application scenarios of automatic processing and analysis of virtual building images that require highly accurate and detailed feature recognition.

[0117] Specifically, the training process of the neural network optimized based on quantum state simulation is as follows:

[0118] 1. Perform parameter quantization, initialize the quantum state for each neural network parameter, and simulate various possibilities of the parameters. Each weight parameter and bias are initialized as a superposition form of quantum states. This multi-state nature of the parameters can effectively enhance the flexibility and adaptability of the neural network when processing complex and variable virtual building image data. The initialization method is expressed as:

[0119]

[0120] In the formula, is the k-th weight quantum state of the neural network; is the k-th bias quantum state of the neural network; is the k-th weight of the neural network; is the k-th bias of the neural network; is the k-th weight complex probability amplitude of the neural network; is the k-th bias complex probability amplitude of the neural network; K is the number of neurons in the neural network.

[0121] Furthermore, the weight complex probability amplitude of the neural network and the bias complex probability amplitude of the neural network satisfy the normalization condition, expressed as:

[0122]

[0123] In the formula, || is the absolute value symbol.

[0124] 2. Define a Hamiltonian to describe the energy of the system. Its eigenvalues are associated with the loss function of the network, thus transforming the problem into solving the ground state problem of the quantum system to describe the energy state of the system. This can enable the network to more precisely adjust the weights and biases during the feature extraction of training virtual building images, especially when dealing with building images with high dimensions and complex geometric structures, and can more effectively find the optimal parameter configuration, expressed as:

[0125]

[0126] In the formula, w p is the weight of the neural network, b p is the bias of the neural network; L p () is the loss function of the neural network; |w p ,bp > is the right ket of the quantum state corresponding to the neural network weights and biases; <w p , b p | is the left ket of the quantum state corresponding to the neural network weights and biases.

[0127] Furthermore, the calculation method of the loss function is expressed as:

[0128]

[0129] In the formula, ∥∥ is the L2 norm; y p is the target output of the neural network, that is, the class probability obtained by performing a preset Softmax calculation on the output features of the neural network; σ p is the Sigmoid function based on non-linear adjustment.

[0130] Let the input of the Sigmoid function based on non-linear conditions be z p , and the calculation method is expressed as:

[0131]

[0132] In the formula, γ pe is the hyperparameter for adjusting the non-linear strength.

[0133] 3. According to the principles of quantum dynamics, let the system evolve according to the Hamiltonian, and this process is simulated by quantum gates to explore the parameter space, so that during the feature extraction process, the network can select the optimal solution from multiple possible parameter states, thus better adapting to the details and features in the virtual building images, expressed as:

[0134]

[0135] In the formula, t is the number of iterations, h pae is the reduced Planck constant; |Ψ p (t)> is the quantum state of the system at the t-th iteration; |Ψ p (0)> is the quantum state of the system at the initial iteration; i * is the imaginary unit.

[0136] Furthermore, the unit time evolution operator is defined as This evolution process describes the non-linear interaction between parameters through high-order Hamiltonian terms, and the evolution operator is expanded using the Taylor series, expressed as:

[0137]

[0138] In the formula, I is the identity matrix.

[0139] 4. Measure the quantum system at the evolved specific vertex time point. The measurement result collapses the quantum state into specific states, which correspond to possible parameter values. Through measurement, the system collapses to a certain state, determining the update direction of the parameters, expressed as:

[0140] M p =|Ψ p (t)><Ψ p (t)|

[0141] In the formula, M p is the quantum measurement result; |Ψ p (t)> is the system quantum state at the t-th iteration; <Ψ p (t)| is the conjugate of the system quantum state at the t-th iteration.

[0142] 5. Evaluate the parameter state after measurement, calculate its fitness to judge the performance of the state. In the feature extraction of virtual building images, the network is allowed to dynamically adjust parameters according to the complexity of image features to improve the recognition accuracy and processing speed. The calculation method of fitness is expressed as:

[0143] Fitness p =<Ψ p (t)|H p |Ψ p (t)>

[0144] In the formula, Fitness p is the fitness function of the neural network, measuring the energy state of the quantum state under the current configuration.

[0145] 6. Update the quantum state parameters according to the measurement result and fitness evaluation. If the measurement result shows performance improvement, continue to evolve in this direction; otherwise, change the evolution path by adjusting the Hamiltonian. The update method is expressed as:

[0146]

[0147] In the formula, γ pae is the learning rate of the neural network; is the gradient of the fitness function with respect to the neural network weights; is the gradient of the fitness function with respect to the neural network biases; is the update increment of the neural network weight parameters; is the update increment of the neural network bias parameters; T p (t) is the temperature parameter at the t-th iteration.

[0148] Preferably, γ pae is set to 0.01.

[0149] Furthermore, architectural images often contain complex geometric structures and diverse texture information. The algorithm needs to be able to accurately capture key features while maintaining efficient search. By adopting an adaptive quantum annealing scheduling strategy to optimize the dynamic parameter adjustment during the process, monitoring the loss fluctuations during the training process, and adaptively adjusting the temperature of the quantum system, the search process can be more finely controlled, and the convergence speed and efficiency of the algorithm can be optimized. The calculation method of the temperature parameter is expressed as:

[0150]

[0151] where T initial is the initial temperature; α p is the annealing rate coefficient; is the difference between the current loss and the loss of the previous iteration.

[0152] Preferably, T initial is set to 0.95, and α p is set to 0.1.

[0153] 7. Repeat the above steps until the termination condition is met. The termination conditions include reaching the preset number of iterations N p or the loss function L p drops to the threshold (usually 10 -4 );

[0154] Through cyclic iteration, the quantum state simulation optimization algorithm can efficiently search the high-dimensional complex parameter space and achieve the optimization of the neural network.

[0155] After feature extraction is completed, a preset classifier is used to evaluate visual comfort. The classifier is trained with the architectural image data after feature extraction. For example, a random forest or a support vector machine can be used to implement visual comfort evaluation. The evaluation categories include: comfortable, medium, and uncomfortable.

[0156] In some other embodiments, a visual comfort evaluation device for virtual reality architecture is provided, based on a visual comfort evaluation method for virtual reality architecture according to any one of the preceding items, including:

[0157] Multiple clients, each client holds a local architectural image data sample and trains a visual comfort evaluation model locally;

[0158] The architectural image data of the client is collected from a simulated architectural scene in the virtual reality environment. The simulated architectural scene is constructed by professional designers in virtual reality software according to real architectural parameters;

[0159] A central server, which is used to aggregate the updated parameters of the local visual comfort evaluation models from each client, perform global updates, and send the globally updated parameters back to each client.

[0160] In some embodiments, the client further includes a data acquisition module, and the data acquisition module uses a high-precision virtual reality helmet and an environment tracking device in the virtual reality environment to acquire building image data.

[0161] In some embodiments, the client includes an annotation module for annotating building image data, and a generation module for generating samples using an adversarial network algorithm generated based on the distribution of building image data;

[0162] The annotation module annotates the image as one of the comfort, medium, or discomfort levels according to the visual feedback and experience of the user in the virtual environment.

[0163] In some embodiments, a classifier is further included, and the classifier is obtained by training with the building image data after feature extraction and is used for visual comfort evaluation. The evaluation categories include: comfort, medium, and discomfort.

[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered by the scope of the claims and the description of the present invention.

Claims

1. A visual comfort assessment method for virtual reality buildings, characterized in that: include: Collect building image data; Annotating the building image data, and annotating the image as one of a comfortable, medium or uncomfortable level according to the visual feedback and experience of the user in the virtual environment; The adversarial network algorithm based on the distribution of architectural image data is used to generate samples to expand the architectural image dataset; Using a federated learning architecture, the visual comfort assessment model is trained using local building image data samples on multiple distributed devices or servers; The local visual comfort assessment model on each device or server is trained on the architectural image data it holds, and interacts with the central visual comfort assessment model for parameter updates. After multiple iterations, the final trained visual comfort assessment model is obtained.

2. The visual comfort evaluation method for virtual reality buildings according to claim 1 is characterized in that: The building image data is collected by using a high-precision virtual reality helmet and an environment tracking device in a virtual reality environment, and the collected content covers different perspectives inside the building.

3. The visual comfort evaluation method for virtual reality buildings according to claim 1 is characterized in that: The sample generation using the adversarial network algorithm based on the distribution generation of building image data includes: Initialize the generator and discriminator of the generative adversarial network, where the generator inputs a noise vector and outputs a generated sample, and the discriminator is a binary classification network that determines whether the sample comes from real building image data; By analyzing the distribution characteristics of architectural image training data, the generator achieves the transition from the source data distribution of architectural images to the target data distribution; Gradually optimize the generation quality of the generator through adversarial training, and use the loss function to constrain the training process of the generator and discriminator; By repeatedly analyzing the distribution characteristics of architectural image training data, the generator achieves a transition from the source data distribution of architectural images to the target data distribution; the generation quality of the generator is gradually optimized through adversarial training, and the loss function is used to constrain the steps of the training process of the generator and discriminator until the generated samples meet the preset conditions.

4. The visual comfort assessment method for virtual reality buildings according to claim 1 is characterized in that: The parameter updating interaction between the local visual comfort assessment model and the central visual comfort assessment model includes: After the local visual comfort assessment model is trained, the updated model parameters are sent to the central visual comfort assessment model; The central visual comfort assessment model aggregates the updated parameters from each local visual comfort assessment model and performs global update; The globally updated parameters are then sent back to each local visual comfort assessment model for the next round of iterative training; Through multiple iterations, the parameter update interactions between the local visual comfort assessment model and the central visual comfort assessment model eventually obtain a globally optimized visual comfort assessment model.

5. The visual comfort assessment method for virtual reality buildings according to claim 1 is characterized in that: Also includes: The local visual comfort evaluation model adopts a fully connected neural network to realize feature extraction of virtual building images, and adopts a quantum state simulation optimization algorithm to perform parameter optimization to obtain optimized neural network parameters.

6. A visual comfort evaluation device for virtual reality buildings, based on a visual comfort evaluation method for virtual reality buildings according to any one of claims 1 to 5, characterized in that: include: Multiple clients, each client holds local building image data samples and trains the visual comfort assessment model locally; The architectural image data of the client is collected from a simulated architectural scene in a virtual reality environment, and the simulated architectural scene is constructed in the virtual reality software by a professional designer according to real architectural parameters; The central server is used to aggregate the updated parameters of the local visual comfort assessment model from each client, perform global update, and send the globally updated parameters back to each client.

7. The visual comfort assessment device for virtual reality buildings according to claim 6, characterized in that: The client also includes a data acquisition module, which uses a high-precision virtual reality helmet and an environment tracking device in a virtual reality environment to collect building image data.

8. The visual comfort assessment device for virtual reality buildings according to claim 7, characterized in that: The client includes a labeling module for labeling building image data, and a generation module for generating samples using an adversarial network algorithm based on building image data distribution generation; The labeling module labels the image as one of comfortable, medium or uncomfortable levels according to the user's visual feedback and experience in the virtual environment.

9. The visual comfort assessment device for virtual reality buildings according to claim 6, characterized in that: It also includes a classifier, which is trained by the architectural image data after feature extraction and is used for visual comfort assessment. The assessment categories include: comfortable, medium, and uncomfortable.

Images