A multi-dimensional intelligent evaluation method and computer system for architectural space visual comfort based on model knowledge transfer

Abstract

This invention discloses a multidimensional intelligent evaluation method for architectural space visual comfort based on model knowledge transfer. This method forms a target dataset by combining existing open-source image evaluation datasets with newly added data based on annotation rules. Evaluation labels are standardized and normalized using a fitted normal distribution. A multi-gated hybrid expert model is employed, with pre-training and distillation learning performed via a teacher-student model. Through the transfer of model knowledge and the updating of normal distribution parameters, a comprehensive multidimensional evaluation of image color, lighting, and architectural layout is achieved. The advantages of this method include improved evaluation accuracy and interpretability, the integration of subjective and objective evaluations, and the efficient application of knowledge transfer, providing an intelligent and comprehensive evaluation tool for the design of architectural space visual comfort.

Description

Technical Field

[0001] The present invention relates to the field of architectural design and comfort evaluation, and in particular to a multi-dimensional intelligent evaluation method and computer system for architectural space visual comfort based on model knowledge migration. Background Art

[0002] With the continuous development of intelligent technology and the deepening of people-oriented concepts in architectural space design, the quality requirements for public building spaces are gradually increasing. Visual comfort, as one of the key factors affecting the design quality of public building spaces, is directly related to physical and mental health in daily life. Traditional visual comfort evaluation is usually based on a single dimension of quantitative physical parameters such as light intensity and color temperature, which cannot fully reflect the subjective visual experience of users in complex architectural spaces. At the same time, due to the diversity of spatial functions and the variability of user needs, the evaluation methods based on small amounts of objective data through sampling methods such as questionnaires in the past are unable to adapt to the visual comfort needs of different scenarios.

[0003] With the continuous advancement of intelligent technology, in the fields of computer vision and image evaluation, image-based intelligent evaluation technology for architectural scene quality continues to develop. Subjective scoring and grading algorithms for different dimensions of images, such as color, lighting, and layout, are becoming increasingly widely used. More and more image evaluation datasets are being open-sourced, but evaluation datasets of different dimensions often suffer from problems such as incomplete labels and inconsistent label dimensions, making it difficult to uniformly apply them to the visual comfort evaluation of architectural scene images. The main problems are:

[0004] 1) Migrating image datasets with different evaluation dimensions and building unified labeled datasets: Existing data on lighting, color, layout, and other aspects comes from diverse sources and is often heterogeneous, often developed for a single evaluation dimension or for different evaluation purposes. Key challenges in building evaluation models include pre-training on existing image datasets, efficiently integrating existing open-source data through knowledge transfer, normalizing different evaluation labels, and introducing new datasets.

[0005] 2) Interpretability of the Multi-Dimensional Intelligent Evaluation Algorithm for Visual Comfort: Previous evaluation algorithms often focused on a single dimension, using prompts such as "Which scene looks more comfortable" to prompt users to select an image and thereby categorize comfort levels. However, visual comfort evaluation is clearly influenced by multiple dimensions, and implementing a comprehensive multi-dimensional evaluation presents another challenge.

[0006] This is an area where this application needs to focus on improvement. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a multi-dimensional intelligent evaluation method and computer system for the visual comfort of architectural space based on model knowledge transfer, aiming to improve the quality of architectural space design and meet the comfort requirements of modern public spaces. It not only improves the accuracy and interpretability of the evaluation, but also realizes the fusion of subjective and objective evaluation and the efficient application of knowledge transfer, providing an intelligent and comprehensive evaluation tool for the visual comfort design of architectural space.

[0008] In order to solve the above technical problems, the present invention provides a multi-dimensional intelligent evaluation method for architectural space visual comfort based on model knowledge transfer, comprising the following steps:

[0009] Step S1: Establishment of image multi-dimensional evaluation dataset;

[0010] Collect existing open-source image evaluation datasets, including datasets for image color, image lighting, and different dimensions of building layouts, and continuously expand the datasets by supplementing them with on-site architectural space images.

[0011] Step S2: label division of the newly added data set;

[0012] Extract two different architectural space images from the dataset to be evaluated. Based on the prompt "Which of the following images performs better in terms of color, lighting, architectural layout, etc.", the corresponding evaluation dimension of the selected image is rewarded or penalized accordingly. After multiple rounds of scoring, the final label of the current dataset is formed.

[0013] By fitting a normal distribution To reflect the evaluation distribution of the current building picture, where μ i Indicates the rating of the current image on a certain evaluation dimension. A larger value indicates a higher rating. Indicates the stability of the current rating. The smaller the value, the more stable the evaluation of the current image. The larger the value, the more unstable and controversial the evaluation of the object.

[0014] Step S3: Label normalization of the migration dataset;

[0015] Since existing evaluation datasets are often evaluated on different dimensions, the label design is also different. Some are based on fitting normal distribution or other probability distribution forms, while others are directly divided into 1, 2, 3...n levels. In order to migrate datasets with different labels to the same model, these datasets are allowed to be evaluated on only one dimension, but their labels must be normalized to the normal distribution. On; similarly, μ i Indicates the rating value of the current image on a certain evaluation dimension, corresponding to the original label evaluation value of the migration dataset; Indicates the stability of the current score. However, since the migration dataset may not have this label, a default value is set based on the dataset size for those without this label.

[0016] Step S4: pre-training the multi-gated hybrid expert model using the migration dataset;

[0017] Pre-training the multi-gated mixture of experts model using the normalized migration dataset in step S3;

[0018] A multi-gated mixture of experts model is used to fit the μ of the input image in each dimension i ,σ i The framework of the multi-gated hybrid expert model includes the following aspects: first, after the data is input, it passes through an attention mechanism layer, and then enters a gating network and several expert layers respectively. Among them, each expert layer is composed of the encoder module of the Transformer. Each expert layer receives the same information from the attention mechanism layer, and then the output parameters are weighted and combined by the weight values ​​output by the gating network; the gating network is composed of several layers of neural networks, and the terminal activation function selects softmax. The weighted and reorganized data are then input into each discriminant layer corresponding to the evaluation dimension. The discriminant layer is also composed of several layers of neural networks. Finally, each discriminant layer outputs the evaluation value μ of the corresponding dimension. i ,σ i ;

[0019] The base loss value of the multi-gated hybrid expert model is composed of the KL divergence of the model output and input distribution. Since there are multiple evaluation dimensions, the weighted average is taken as the final base loss value;

[0020] Loss ini =∑ i＝1 w i D KL (In i |Out i ),∑ i＝1 w i =1;

[0021] Step S5: The multi-gated hybrid expert model learns the newly added data set and realizes the multi-dimensional evaluation of visual comfort;

[0022] The setting of the student model is the same as the multi-gated hybrid expert model structure described in step S4. Because the scene objects described by the migration dataset are not necessarily all architectural spaces, but are pure color combination pictures or some pictures that can reflect the lighting effects, the migration dataset often cannot be directly trained together with the new dataset. However, the migration dataset does have guiding value for evaluating the color, lighting, layout and other dimensions of the pictures.

[0023] Use the pre-trained model in step S4 as the teacher model, train a new multi-gated hybrid expert model of the same framework as the student model to learn the newly added data set marked in step S2, input the newly added data set into the teacher model and the student model, and obtain the output μ on different evaluation dimensions. T , and μ S , And calculate the KL divergence D between the two sets of outputs KL (T|S), used to measure the distillation loss of the teacher model and the student model;

[0024]

[0025] This part of the loss value is weighted with the loss value set by the original model to coordinate the training of the model, realize the migration of multivariate data and the multi-dimensional evaluation of visual comfort, and the distillation loss Loss between the input teacher model and the student model output is calculated. dist Multiplying by the scaling factor α, the final overall loss value is:

[0026] Loss total =Loss ini +αLoss dist .

[0027] The scaling factor α is set based on the initial values ​​of the two loss values ​​and is dynamically adjusted as the model is trained. The purpose is to keep the two loss values ​​of the same order of magnitude and avoid serious bias and oscillation during training.

[0028]

[0029] The step S2 includes the following specific steps:

[0030] Step S21: Initialize the labels of the dataset to be evaluated;

[0031] Set the evaluation value μ of all the data to be evaluated on the measured dimension (color / light / building layout) o ∈

[0032] [0,100], initial μ i =50, Each dimension has a corresponding μ and σ;

[0033] Step S22: Set the probability P of two images being extracted for comparison when evaluating a certain dimension. AB , ensuring the stability of expected value and variance updates as much as possible;

[0034]

[0035]

[0036] Where: β is the variance of each comparison, which is adjusted according to the actual situation and is set by default to c is a measure of the overall uncertainty brought by the two images during this comparison. At the same time, in order to prevent a picture from being evaluated too many times while other pictures have never been evaluated, it is necessary to set an upper limit on the number of times a single user can evaluate a picture. Pictures that reach this limit will no longer be selected by the user.

[0037] Step S23: Each time two images are compared, a comparison prompt word is obtained on one dimension. After answering the comparison prompt word, a basic reward or penalty value R is obtained for each of the two images. For example, a prompt word is obtained: "Which of the following images performs better in terms of color / lighting / architectural layout, etc." This means that the comparison is only conducted on one dimension. There are five options: A is better, B is better, about the same, both are good, and neither is good. The basic reward or penalty values ​​R for the five options for the two images are shown in Table 1:

[0038] Table 1 Basic reward value settings

[0039] Options <![CDATA[Reward and punishment value R A > <![CDATA[Reward and punishment value R B > A is better <![CDATA[R A >0]]> <![CDATA[R B >0]]> B is better <![CDATA[R A <0]]> <![CDATA[R B <0]]> almost <![CDATA[R A =0]]> <![CDATA[R B =0]]> All good <![CDATA[R A >0]]> <![CDATA[R B >0 <!-- 3 -->]]> Neither is good <![CDATA[R A <0]]> <![CDATA[R B <0]]>

[0040] Since two images theoretically have different distributions after multiple rounds of comparison, assuming that the evaluation in the first few rounds is valid, the image with a larger μ should theoretically win. However, unexpected situations always occur in the actual comparison process. Therefore, in addition to the basic reward and penalty values, we should also consider whether the comparison between the two images after several rounds of accumulation meets expectations. The specific expected value and standard deviation updates are shown in Table 2:

[0041] Table 2 Expected value update settings

[0042]

[0043]

[0044] Table 3 Standard deviation update settings

[0045]

[0046] in, To measure the overall uncertainty of the two images during this comparison, ε is used as the threshold for determining the probability of a draw, preferably set to 0.5. The v and w functions in Table 3 are defined as follows:

[0047]

[0048] w(x,ε)=(x,ε)·(v(x,ε)+x-ε)

[0049] Where: φ(·) is the probability density function of the standard normal distribution, Φ(·) is the cumulative distribution function of the standard normal distribution;

[0050] Step S24:

[0051] Let different users repeatedly evaluate the dataset to be tested, iterate the expected value and standard deviation of each image in the current dataset in different dimensions, and finally form the evaluation label of the dataset.

[0052] The step S3 includes the following specific steps:

[0053] Step S31: normalize the data labels annotated in a normal distribution form;

[0054] For image evaluation labels using the normal distribution method, linear scaling is directly used to convert the original range to the current range. The specific formula is as follows:

[0055]

[0056] μ nnw ∈[0,max onw ],μ transfer ∈[0,max transfer ]

[0057] Step S32: normalize the data labels using multi-level annotation;

[0058] For data labels using a discrete grading system, they are mapped to a normal distribution based on the true meaning of the grading. For example, if a data set has the following grading definitions: 1, 2, 3, 4, 5, with 1 being the worst and 5 being the best, the mapping rules are as follows:

[0059]

[0060] The standard deviation is set to the default value Or fine-tune according to the specific distribution of each level of the original dataset: assuming that the evaluation of high-level and low-level images is stable, multiply the high-level images by a scaling factor; assuming that the evaluation of intermediate-level images is fuzzy and unstable, multiply this type of image by a larger scaling factor than the former; the specific decision is based on the size of the migration dataset and the label distribution.

[0061] The present invention also provides a computer system comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the multi-dimensional intelligent evaluation method for architectural space visual comfort based on model knowledge migration.

[0062] The beneficial effects of the present invention are:

[0063] 1) Introducing a multi-dimensional evaluation dataset to improve the measurement dimension of architectural space visual comfort image evaluation;

[0064] By introducing multivariate datasets for model pre-training and knowledge transfer, we can assist in establishing and improving multidimensional measurements of architectural space visual comfort images.

[0065] 2) Achieve model knowledge transfer and alignment of evaluation labels to match the integration of different evaluation rules and standards;

[0066] 3) The teacher model is trained using the migration dataset, and the target student model is guided to learn the newly added target dataset through knowledge distillation, thus achieving model knowledge transfer;

[0067] 4) This paper proposes a label normalization method under different evaluation rule systems, which realizes the integration of multivariate evaluation data sets;

[0068] 5) The present invention uses a multi-gated hybrid expert model to output evaluation values ​​of the building space image to be tested in different dimensions, highlighting more specifically the areas where the visual comfort of the building space to be tested is good and where it is lacking. This is more conducive to designers to implement solution optimization and design, and realize multi-dimensional intelligent evaluation of the visual comfort of the building space. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0070] Figure 1 It is a flowchart of a specific embodiment of the present invention;

[0071] Figure 2 A schematic diagram of a process for pre-training a multi-gated hybrid expert model according to a specific embodiment of the present invention;

[0072] Figure 3 A schematic diagram of the expert scoring process for label division of a newly added dataset according to a specific embodiment of the present invention;

[0073] Figure 4 A schematic diagram of a multi-gated hybrid expert model learning a new data set according to a specific embodiment of the present invention;

[0074] Numbers in the figure:

[0075] S1—Establishment of image multi-dimensional evaluation dataset;

[0076] S2—label division of newly added datasets;

[0077] S3—label normalization of the migrated dataset;

[0078] S4—Pre-training the multi-gated mixture of experts model;

[0079] S5—New dataset for learning multi-gated mixture of experts model. DETAILED DESCRIPTION

[0080] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0081] like Figure 1 As shown, the present invention provides a multi-dimensional intelligent evaluation method for architectural space visual comfort based on model knowledge transfer. By collecting existing open-source image evaluation datasets and newly added data according to annotation rules to form a target dataset, the evaluation labels are standardized and normalized by fitting the normal distribution. A multi-gated hybrid expert model is used, and pre-training and distillation learning are performed through a teacher-student model. Through the transfer of model knowledge and the updating of normal distribution parameters, a multi-dimensional comprehensive evaluation of image color, light and architectural layout is achieved. The method comprises the following steps:

[0082] Step S1: Establish an image multi-dimensional evaluation dataset;

[0083] Collect existing open-source image evaluation datasets, including image evaluation data on color, light, and architectural layout dimensions, to form a migration dataset, such as open-source datasets such as AVA and Place Pulse. Also, take new images within architectural spaces to create a new target dataset for a more comprehensive evaluation of architectural space visual comfort.

[0084] In specific operations, it is necessary to integrate various types of image evaluation datasets, which are not necessarily all image evaluation datasets related to the visual comfort of architectural images, or image datasets related to the color, lighting, etc. of the images themselves or the evaluation of architectural layout; then, based on the evaluation dimensions covered by the collected datasets, such as color, light and architectural layout, the newly added image data is classified and statistical labels are calculated.

[0085] Step S2: The target dataset is divided into labels, such as Figure 3 As shown;

[0086] Generate labels for new datasets by comparing two images together;

[0087] Two images are randomly selected and evaluated on a single dimension, such as color, lighting, or layout. The user is asked to select one of the following options: "A / B is better," "Similar," "Both are good," or "Both are bad." Each comparison assigns a reward or penalty to the image based on the user's choice, which is used to update the expected value and standard deviation of the label. After multiple evaluations, a normal distribution of image labels is established to ensure stability and consistency of the evaluation. The specific steps are as follows:

[0088] Step S21: Initialize the labels of the dataset to be evaluated;

[0089] Set the evaluation value μ of all the data to be evaluated on the measured dimension (color / light / building layout) i ∈[0,100], initial μ i =50, Each dimension has a corresponding μ and σ;

[0090] Step S22: Set the probability P of two images being extracted for comparison when evaluating a certain dimension. AB , ensuring the stability of expected value and variance updates as much as possible;

[0091]

[0092] Where: β is the variance of each comparison, which is adjusted according to the actual situation and is set by default to c is a measure of the overall uncertainty brought by the two images during this comparison. At the same time, in order to prevent a picture from being evaluated too many times while other pictures have never been evaluated, it is necessary to set an upper limit on the number of times a single user can evaluate a picture. Pictures that reach this limit will no longer be selected by the user.

[0093] Step S23: Each time two images are compared, a prompt is given: "Which of the following images performs better in terms of color / lighting / architectural layout, etc." This means that a comparison is only conducted on one dimension. There are five options: A is better, B is better, about the same, both are good, and neither is good. The five options represent the basic reward and penalty values ​​R for the two images, as shown in Table 1:

[0094] Table 1 Basic reward value settings

[0095] Options <![CDATA[Reward and punishment value R A > <![CDATA[Reward and punishment value R B > A is better <![CDATA[R A >0]]> <![CDATA[R B >0]]> B is better <![CDATA[R A <0]]> <![CDATA[R B <0]]> almost <![CDATA[R A =0]]> <![CDATA[R B =0]]> All good <![CDATA[R A >0]]> <![CDATA[R B >0]]> Neither is good <![CDATA[R A <0]]> <![CDATA[R B <0]]>

[0096] Since two images theoretically have different distributions after multiple rounds of comparison, assuming that the evaluation in the first few rounds is valid, the image with a larger μ should theoretically win. However, unexpected situations always occur in the actual comparison process. Therefore, in addition to the basic reward and penalty values, we should also consider whether the comparison between the two images after several rounds of accumulation meets expectations. The specific expected value and standard deviation updates are shown in Table 2:

[0097] Table 2 Expected value update settings

[0098]

[0099]

[0100] Table 3 Standard deviation update settings

[0101]

[0102] in, To measure the overall uncertainty of the two images during this comparison, ε is used as the threshold for determining the probability of a draw, preferably set to 0.5. The v and w functions in Table 3 are defined as follows:

[0103]

[0104] w(x,ε)=(x,ε)·(v(x,ε)+x-ε)

[0105] Where: φ(·) is the probability density function of the standard normal distribution, Φ(·) is the cumulative distribution function of the standard normal distribution;

[0106] Step S24:

[0107] Let different users repeatedly evaluate the dataset to be tested, iterate the expected value and standard deviation of each image in the current dataset in different dimensions, and finally form the evaluation label of the dataset.

[0108] Step S3: Label normalization of the migration dataset;

[0109] Since datasets from different sources have different labels, they need to be unified to the same distribution. For data labels annotated with a normal distribution, they are linearly converted to the range of the current model. For datasets with multi-level ratings, such as a 1 to 5 rating system, they are linearly mapped to a normal distribution to ensure comparability of different labels. The specific steps are as follows:

[0110] Step S31: normalize the data labels annotated in a normal distribution form;

[0111] For image evaluation labels using the normal distribution method, linear scaling is directly used to convert the original range to the current range. The specific formula is as follows:

[0112]

[0113] μ nnw ∈[0,max nnw ],μ transfer ∈[0,max transfer ]

[0114] Step S32: normalize the data labels using multi-level annotation;

[0115] For data labels using a discrete grading system, they are mapped to a normal distribution based on the true meaning of the grading. For example, if a data set has the following grading definitions: 1, 2, 3, 4, 5, with 1 being the worst and 5 being the best, the mapping rules are as follows:

[0116]

[0117] The standard deviation is set to the default value Or fine-tune according to the specific distribution of each level of the original dataset: assuming that the evaluation of high-level and low-level images is stable, multiply the high-level images by a scaling factor; assuming that the evaluation of intermediate-level images is fuzzy and unstable, multiply this type of image by a larger scaling factor than the former; the specific decision is based on the size of the migration dataset and the label distribution.

[0118] Step S4: pre-training of the teacher model of the multi-gated hybrid expert model;

[0119] As attached Figure 2 As shown, the present invention adopts a multi-gated hybrid expert model to fit the μ of the input image in each dimension. i ,σ i The Transformer-based multi-gated hybrid expert model is preliminarily trained using the migration dataset to lay the foundation for the evaluation of the new dataset. The model framework includes an attention mechanism layer, a gating network, an expert layer, and a discriminant layer to ensure the model's multi-dimensional input processing capabilities. The pre-training process is completed by inputting the normalized migration dataset into the model and calculating the weighted KL divergence loss function to optimize the model output;

[0120] The base loss value of the multi-gated hybrid expert model is composed of the KL divergence of the model output and input distribution. Since there are multiple evaluation dimensions, the weighted average is taken as the final base loss value;

[0121] Loss ini =∑ i＝1 w i D KL (In i |Out i ),∑ i＝1 w i =1;

[0122] Step S5: Student model training of multi-gated hybrid expert model, such as Figure 4 As shown;

[0123] A teacher-student model architecture is adopted for further optimization.

[0124] After pre-training the teacher model on the transferred dataset, it guides the student model to learn the new dataset. The new dataset is fed into both the teacher and student models, generating outputs for each evaluation dimension. The KL divergence between the two sets of outputs is calculated to assess the difference between the teacher and student models. The model is then optimized and updated by applying a weighted combination of the distillation loss and the original loss. The details are as follows:

[0125] Use the pre-trained model in step S4 as the teacher model, train a new multi-gated hybrid expert model of the same framework as the student model to learn the newly added data set marked in step S2, input the newly added data set into the teacher model and the student model, and obtain the output μ on different evaluation dimensions. T , and μ S , And calculate the KL divergence D between the two sets of outputs KL (T|S), used to measure the distillation loss of the teacher model and the student model;

[0126]

[0127] This part of the loss value is weighted with the loss value set by the original model to coordinate the training of the model, realize the migration of multivariate data and the multi-dimensional evaluation of visual comfort, and the distillation loss Loss between the input teacher model and the student model output is calculated. dist Multiplying by the scaling factor α, the final overall loss value is:

[0128] Loss total =Loss ini +αLoss dist .

[0129] The present invention also provides a computer system comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the multi-dimensional intelligent evaluation method for architectural space visual comfort based on model knowledge migration.

[0130] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A multi-dimensional intelligent evaluation method for architectural space visual comfort based on model knowledge transfer, comprising the following steps: Step S1: Establishment of image multi-dimensional evaluation dataset; Collect existing open-source image evaluation datasets, including datasets for image color, image lighting, and different dimensions of building layouts, and continuously expand the datasets by supplementing them with on-site architectural space images. Step S2: label division of the newly added data set; Extract two different architectural space images from the dataset to be evaluated at once. Evaluate the selected images on multiple dimensions based on the prompt words and receive corresponding rewards and penalties. After multiple rounds of scoring, the labels for the current dataset are finally formed. By fitting a normal distribution To reflect the evaluation distribution of the current building picture, where μ i Indicates the rating of the current image on a certain evaluation dimension. A larger value indicates a higher rating. Indicates the stability of the current rating. The smaller the value, the more stable the evaluation of the current image. Step S3: Label normalization of the migration dataset; Labels are normalized to a normal distribution Up; μ i Indicates the rating value of the current image on a certain evaluation dimension, corresponding to the original label evaluation value of the migration dataset; Indicates the stability of the current rating; Step S4: pre-training the multi-gated hybrid expert model using the migration dataset; The framework of the multi-gated hybrid expert model includes an attention mechanism layer, a gating network, an expert layer, and a discriminant layer to fit the μ of the input image in each dimension. i , σ i The pre-training process is completed by inputting the normalized migration dataset into the model and calculating the weighted KL divergence loss function to optimize the model output; Step S5: The multi-gated hybrid expert model learns the newly added data set and realizes the multi-dimensional evaluation of visual comfort; The teacher-student model architecture is used for further optimization. After the teacher model is pre-trained using the migration dataset, it guides the student model to learn the new dataset. The new dataset is input into the teacher and student models respectively to obtain the output of each evaluation dimension. By calculating the KL divergence between the two sets of outputs, the difference between the teacher model and the student model is evaluated, and the distillation loss is weightedly combined with the original loss value to optimize and update the model.

2. The multi-dimensional intelligent evaluation method for architectural space visual comfort based on model knowledge transfer according to claim 1 is characterized by: The step S2 includes the following specific steps: Step S21: Initialize the labels of the dataset to be evaluated; Set the evaluation value μ of all the data to be evaluated on the dimension to be tested i ∈[0,100], initial μ i =50, Each dimension has a corresponding μ and σ; Step S22: Set the probability P of two images being extracted for comparison when evaluating a certain dimension. AB , ensuring the stability of expected value and variance updates; Where: β is the variance of each comparison, c is the overall uncertainty brought by the two images during this comparison; Step S23: Each time two images are compared, a comparison prompt word on one dimension is obtained, and after answering the comparison prompt word, a basic reward or punishment value R is obtained for each of the two images; After multiple rounds of comparison, two images have different distributions. Consider whether the comparison between the two images after several rounds of accumulation meets expectations. The specific expected value and standard deviation are updated as shown in Table 2: Table 2 Expected value update settings Table 3 Standard deviation update settings ε is used as the threshold for judging the probability of a draw, and the v and w functions in Table 3 are defined as follows: w(x,ε)=(x,ε)·(v(x,ε)+x-ε) Where: φ(·) is the probability density function of the standard normal distribution, Φ(·) is the cumulative distribution function of the standard normal distribution; Step S24: Let different users repeatedly evaluate the dataset to be tested, iterate the expected value and standard deviation of each image in the current dataset in different dimensions, and finally form the evaluation label of the dataset.

3. The multi-dimensional intelligent evaluation method for architectural space visual comfort based on model knowledge transfer according to claim 1 is characterized by: The step S3 includes the following specific steps: Step S31: normalize the data labels annotated in a normal distribution form; For image evaluation labels using the normal distribution method, linear scaling is directly used to convert the original range to the current range. The specific formula is as follows: m now ∈[0,max now ],m transfer ∈[0,max transfer ] Step S32: normalize the data labels using multi-level annotation; For data labels using a discrete classification system, they are mapped to a normal distribution based on the true meaning of their classification. The mapping rules are as follows:

4. The multi-dimensional intelligent evaluation method for architectural space visual comfort based on model knowledge transfer according to claim 1 is characterized by: The multi-gated hybrid expert model is that the data is first input through an attention mechanism layer, and then respectively input into a gating network and several expert layers, wherein each expert layer is composed of the encoder module of the Transformer, and each expert layer receives the same information from the attention mechanism layer, and then the output parameters are weightedly combined by the weight values ​​output by the gating network; the gating network is composed of several layers of neural networks, and the terminal activation function selects softmax. The weighted reorganized data is then respectively input into each discriminant layer corresponding to the evaluation dimension, and the discriminant layer is also composed of several layers of neural networks. Finally, each discriminant layer outputs the evaluation value μ of the corresponding dimension. i ,σ i ; The base loss value of the multi-gated hybrid expert model is composed of the KL divergence of the model output and input distribution. Since there are multiple evaluation dimensions, the weighted average is taken as the final base loss value; Loss ini =∑ i＝1 w i D KL (In i |Out i ),∑ i＝1 w i =1。 5. The multi-dimensional intelligent evaluation method for architectural space visual comfort based on model knowledge transfer according to claim 1 is characterized by: The new data set is input into the teacher model and the student model to obtain the output μ on different evaluation dimensions. T , and μ S , And calculate the KL divergence D between the two sets of outputs KL (T|S), used to measure the distillation loss of the teacher model and the student model; Distillation loss Loss between the input teacher model and the output of the student model dist Multiplying by the scaling factor α, the final overall loss value is: Loss total =Loss ini +αLoss dist 。 6. A computer system comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the multi-dimensional intelligent evaluation method for architectural space visual comfort based on model knowledge migration as described in any one of claims 1 to 5.

Images