A deep learning-based architectural design system

Abstract

The present invention belongs to the field of architectural design, and discloses an architectural design system based on deep learning, including a data set processing module, a training module, an input module and a generation module; the data set processing module is used to delete duplicate images in the original image set to obtain a screened image set; the training module is used to train a generative adversarial network using the screened image set to obtain a trained generative adversarial network; the input module is used to input architectural design requirements; the generation module is used to convert the design requirements into vector representations, and input the vector representations into the trained generative adversarial network to obtain architectural design images. The present invention can avoid using areas with low recognition as comparison areas, thereby improving the accuracy of screening results, and can avoid using overall information to screen all images, thereby improving the efficiency of the screening process.

Description

Technical Field

[0001] The present invention relates to the field of architectural design, and in particular to an architectural design system based on deep learning. Background Art

[0002] In recent years, the architectural design industry has made significant progress in technology and tools, especially the widespread use of computer-aided design (CAD) software, which has greatly improved the accuracy and efficiency of design. However, although CAD software and other engineering drawing tools have become an indispensable part of architectural design, these tools are mainly focused on executing the designer's clear instructions, making precise drawings and modeling, rather than providing support in design creativity and architectural style innovation.

[0003] At present, various advanced architectural design software, such as Revit, AutoCAD, etc., have extremely high accuracy and reliability in architectural drawing and model building. In addition, some software have begun to try to integrate algorithm-aided design (AAD) functions, using preset algorithms to optimize the structure and functional layout of buildings. However, the main limitations of these software are:

[0004] Creative generation limitations: Most existing technologies rely on parameters and rules preset by designers, lacking autonomy in terms of originality and innovation. Designers need to set all design parameters in advance, and the software executes according to these parameters, which limits the possibility and diversity of design.

[0005] In order to improve the diversity of design, the prior art usually uses a generative adversarial network to generate images that meet the requirements. When training a generative adversarial network, it is usually necessary to first obtain an image set for training and delete duplicate images in the image set. In the prior art, duplicate images are usually screened based only on the overall information of the image or only on the local information of the image. If the screening is based only on local information, when the texture of the image is relatively simple, it is easy to select areas with low recognition for comparison, resulting in inaccurate comparison results; and if all screening is based on overall information, when the number of image sets is too large, it is easy to take a long time to complete the deletion of duplicate images. Summary of the invention

[0006] The purpose of the present invention is to disclose an architectural design system based on deep learning to solve the technical problems raised in the background technology.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] The present invention provides a deep learning-based architectural design system, comprising a data set processing module, a training module, an input module, and a generation module;

[0009] The data set processing module is used to delete duplicate images in the original image set to obtain a screened image set;

[0010] The training module is used to train the generative adversarial network using the screened image set to obtain a trained generative adversarial network;

[0011] The input module is used to input architectural design requirements;

[0012] The generation module is used to convert the design requirements into vector representations and input the vector representations into the trained generative adversarial network to obtain the architectural design image.

[0013] Among them, duplicate images in the original image set are deleted to obtain a screened image set, including:

[0014] Perform the following processing on each image in the original image set:

[0015] S1, for image p in the original image set, calculate the screening coefficient of p;

[0016] S2, obtaining the comparison area based on the screening coefficient;

[0017] S3, screening p based on the comparison area and updating the original image set;

[0018] S4, determining whether there are images in the original image set that have not been screened, if so, proceeding to S5, if not, the remaining images in the original image set constitute the screened image set;

[0019] S5, randomly select an image that has not been screened from the original image set as a new image p, and enter S1.

[0020] Preferably, calculating the screening coefficient of P comprises:

[0021] Convert p to grayscale image gp;

[0022] The screening coefficient was calculated based on gp.

[0023] Preferably, the screening coefficient is calculated based on gp, including:

[0024] The calculation formula of the screening coefficient is:

[0025]

[0026] seldix prepresents the screening coefficient of p, mig and mag represent the minimum and maximum grayscale values ​​in gp respectively, num v Indicates that the grayscale value in gp is equal to the total number of pixels in v, num p Indicates the total number of pixels contained in p, cps indicates the preset contrast value, lts p represents the maximum area of ​​each connected domain in gp, als cp represents the area of ​​gp, w1 and w2 represent the quantity weight and area weight respectively.

[0027] Preferably, obtaining the area for comparison based on the screening coefficient includes:

[0028] Use the following formula to calculate the gp block parameters:

[0029]

[0030] N gp Indicates the block parameters of gp, baseNum indicates the maximum value of the set block parameters, seldix mx Indicates the set screening coefficient comparison value;

[0031] Get the largest positive integer n that satisfies the following inequality;

[0032] 2(n-1)+1≤N gp ≤2n+1

[0033] Divide gp into (2n+1)×(2n+1) regions of equal area;

[0034] Select the comparison area from all areas.

[0035] Preferably, the comparison area is selected from all areas, including:

[0036] Calculate the representative feature value of each region separately;

[0037] The region showing the largest eigenvalue is used as the comparison region.

[0038] Preferably, the calculation formula representing the characteristic value is:

[0039]

[0040] dlval b represents the characteristic value of region b, ub is the set of all pixels in b, gry i is the gray value of i, bsc represents the number of pixels in ub, average represents the gray mean of pixels in ub, lstpx bIt represents the maximum length of the line segment composed of pixels with the same grayscale in b, d1 represents the variance weight of the grayscale value, and d2 represents the length weight.

[0041] Preferably, p is screened based on the comparison area, and the original image set is updated, including:

[0042] Obtain a set CU of coordinates of pixel points in the comparison area;

[0043] Get the image feature tp of the area composed of the coordinates in CU in the image p;

[0044] Filter image p based on tp and update the original image set.

[0045] Preferably, the image p is screened based on tp to update the original image set, including:

[0046] Compute p and any image in the original image set except p as follows:

[0047] For any image t except p in the original image set, if the length of image t is greater than the length of image p, image t is reduced until the length of image t is equal to the length of image p, and image ct1 is obtained;

[0048] Get the image feature tct1 of the area in ct1 consisting of the coordinates in CU;

[0049] Calculate the similarity between tp and tct1. If the similarity is greater than the set similarity threshold, it means that image p and image t are the same. Delete image p or image t from the original image set to update the original image set. If the similarity is less than the set similarity threshold, keep p.

[0050] If the length of image t is less than the length of image p, image t is enlarged until the length of image t is equal to the length of image p, and image ct2 is obtained;

[0051] Get the image feature tct2 of the area in ct2 composed of the coordinates in CU;

[0052] Calculate the similarity between tp and tct2. If the similarity is greater than the set similarity threshold, it means that image p and image t are the same. Delete image p or image t from the original image set to update the original image set. If the similarity is less than the set similarity threshold, retain p.

[0053] Preferably, the architectural design requirements include style type, space requirements and layout requirements.

[0054] Preferably, the generative adversarial network includes a generator and a discriminator.

[0055] Beneficial effects:

[0056] Compared with the existing architectural design method based on generative adversarial networks, the present invention determines the comparison area by calculating the screening coefficient in the preprocessing stage of the image set. On the one hand, it can avoid using areas with low recognition as comparison areas, thereby improving the accuracy of the screening results. On the other hand, it can avoid using overall information to screen all images, thereby improving the efficiency of the screening process, so that the present invention can achieve more efficient deployment when it is actually deployed and used. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0058] Figure 1 A schematic diagram of a deep learning-based architectural design system of the present invention.

[0059] Figure 2 A schematic diagram of a generative adversarial network of the present invention. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present invention.

[0061] like Figure 1 In one embodiment shown, the present invention provides a deep learning-based architectural design system, including a data set processing module, a training module, an input module, and a generation module;

[0062] The data set processing module is used to delete duplicate images in the original image set to obtain a screened image set;

[0063] The training module is used to train the generative adversarial network using the screened image set to obtain a trained generative adversarial network;

[0064] The input module is used to input architectural design requirements;

[0065] The generation module is used to convert the design requirements into vector representations and input the vector representations into the trained generative adversarial network to obtain the architectural design image.

[0066] Among them, duplicate images in the original image set are deleted to obtain a screened image set, including:

[0067] Perform the following processing on each image in the original image set:

[0068] S1, for image p in the original image set, calculate the screening coefficient of p;

[0069] S2, obtaining the comparison area based on the screening coefficient;

[0070] S3, screening p based on the comparison area and updating the original image set;

[0071] S4, determining whether there are images in the original image set that have not been screened, if so, proceeding to S5, if not, the remaining images in the original image set constitute the screened image set;

[0072] S5, randomly select an image that has not been screened from the original image set as a new image p, and enter S1.

[0073] In the preprocessing stage of the image set, the present invention determines the comparison area by calculating the screening coefficient. On the one hand, it can avoid using the area with low recognition as the comparison area, thereby improving the accuracy of the screening result. On the other hand, it can avoid using the overall information to screen all images, thereby improving the efficiency of the screening process, so that the present invention can achieve more efficient deployment when it is actually deployed and used.

[0074] Preferably, the original image set includes multiple images of various architectural styles (such as modernism, Gothic, Baroque, Classicism, etc.).

[0075] Preferably, the screened image set is used to train the generative adversarial network to obtain a trained generative adversarial network, including:

[0076] The first step is to initialize the network weights:

[0077] Before training begins, the parameters of the generator and discriminator are usually initialized randomly.

[0078] In the second step, the generator generates fake data:

[0079] The generator network starts with a random noise vector and generates fake image data.

[0080] The discriminator is evaluated:

[0081] In the third step, the discriminator network receives real images and fake images generated by the generator and tries to distinguish which ones are real and which ones are fake.

[0082] The fourth step is to calculate the loss of the discriminator:

[0083] The loss function of the discriminator usually consists of two parts: the loss of real images and the loss of fake images. The goal of the discriminator is to maximize its ability to distinguish between real and fake images.

[0084] Step 5, back propagation of the discriminator:

[0085] According to the loss function of the discriminator, the network weights of the discriminator are updated through the back-propagation algorithm.

[0086] Step 6: Generator loss calculation:

[0087] The loss function for the generator is to minimize the probability that the discriminator correctly classifies a fake image as real, i.e., fools the discriminator.

[0088] Step 7: Backpropagation of the generator:

[0089] According to the generator's loss function, the generator's network weights are updated with the goal of generating more realistic images.

[0090] Step 8: Alternate training:

[0091] During the training process, the generator and the discriminator are trained alternately. The generator tries to generate more and more realistic images, while the discriminator tries to better distinguish between real and fake images.

[0092] Step 9: Evaluate and adjust:

[0093] During the training process, the quality of the generated images is evaluated regularly and the training parameters or network structure are adjusted as needed.

[0094] Step 10, training iteration:

[0095] Repeat steps 2 to 9 until the quality of the generated image reaches a satisfactory level or the predetermined number of training iterations is reached.

[0096] Step 11: Save and apply the model:

[0097] After training is completed, save the trained generator model to generate new image data

[0098] Preferably, calculating the screening coefficient of P comprises:

[0099] Convert p to grayscale image gp;

[0100] The screening coefficient was calculated based on gp.

[0101] By converting to a grayscale image, the number of color channels can be reduced and the calculation efficiency can be improved.

[0102] Preferably, the screening coefficient is calculated based on gp, including:

[0103] The calculation formula of the screening coefficient is:

[0104]

[0105] seldix p represents the screening coefficient of p, mig and mag represent the minimum and maximum grayscale values ​​in gp respectively, num v Indicates that the grayscale value in gp is equal to the total number of pixels in v, num p Indicates the total number of pixels contained in p, cps indicates the preset contrast value, lts p represents the maximum area of ​​each connected domain in gp, als cp represents the area of ​​gp, w1 and w2 represent the quantity weight and area weight respectively.

[0106] The screening coefficient of the present invention, on the one hand, takes the number distribution of pixel points of various gray values ​​in gp as a reference, and on the other hand, takes the maximum value of the connected domain as a reference. In this way, the screening coefficient can be comprehensively expressed from two different angles to arbitrarily select a region as a comparison region, and the probability of selecting a region with low representativeness is selected. That is, the larger the screening coefficient, the greater the probability of selecting a region with low representativeness. This provides useful parameters for the subsequent calculation of the block parameters of the present invention, which is conducive to more reasonably determining the number of regions and reducing the probability of selecting a region with low representativeness.

[0107] Preferably, the preset contrast value is 7.

[0108] Preferably, the quantity weight is two-thirds and the area weight is one-third.

[0109] Preferably, obtaining the area for comparison based on the screening coefficient includes:

[0110] Use the following formula to calculate the gp block parameters:

[0111]

[0112] N gp Indicates the block parameters of gp, baseNum indicates the maximum value of the set block parameters, seldix mx Indicates the set screening coefficient comparison value;

[0113] Get the largest positive integer n that satisfies the following inequality;

[0114] 2(n-1)+1≤N gp ≤2n+1

[0115] Divide gp into (2n+1)×(2n+1) regions of equal area;

[0116] Select the comparison area from all areas.

[0117] In the process of obtaining the comparison area, the present invention first calculates the block parameters, and then determines the number of blocks based on the block parameters, so as to avoid obtaining too many or too small areas, and improve the probability of using highly representative areas as comparison areas. Specifically, if the screening coefficient is larger, the block parameter is smaller, and the number of obtained areas is smaller, because when the screening coefficient is larger, the probability that the selected comparison area belongs to a low-representative area is greater, at this time, the number of areas of the present invention will be correspondingly smaller, thereby reducing the probability of selecting a low-representative comparison area from the area; if the screening coefficient is smaller, the block parameter is larger, and the number of obtained areas is larger, because when the screening coefficient is smaller, the probability that the selected comparison area belongs to a low-representative area is smaller, at this time, the number of areas of the present invention will be correspondingly larger, thereby reducing the area of ​​the comparison area, which is conducive to improving the efficiency of subsequent feature extraction, thereby improving the efficiency of the image screening process.

[0118] Preferably, the maximum value of the block parameter is set to 10.

[0119] Preferably, the screening coefficient comparison value is set to 8.

[0120] Preferably, the comparison area is selected from all areas, including:

[0121] Calculate the representative feature value of each region separately;

[0122] The region showing the largest eigenvalue is used as the comparison region.

[0123] In the present invention, the larger the eigenvalue is, the more it can represent the entire image.

[0124] Preferably, the calculation formula representing the characteristic value is:

[0125]

[0126] dlval b represents the characteristic value of region b, ub is the set of all pixels in b, gry iis the gray value of i, bsc represents the number of pixels in ub, average represents the gray mean of pixels in ub, lstpx b It represents the maximum length of the line segment composed of pixels with the same grayscale in b, d1 represents the variance weight of the grayscale value, and d2 represents the length weight.

[0127] In addition to considering the variance, the present invention also introduces the maximum length of the line segment composed of pixels with the same grayscale, thereby comprehensively and more accurately representing the representativeness of the region from both the overall and local aspects. Specifically, the larger the variance of the grayscale values ​​of the pixels in the region, the smaller the lstpx b The larger the value of , the larger the eigenvalue. In this way, the probability of using the area with poor continuity of the line segment as the comparison area can be reduced, thereby obtaining a more effective comparison area.

[0128] Preferably, the variance weight of the gray value is three-fifths, and d2 represents the length weight is two-fifths.

[0129] Preferably, the process of confirming a line segment composed of pixels with the same grayscale is as follows: for any pixel on the line segment except the endpoints, there are only two pixels with the same grayscale value as itself in its 8-bit region.

[0130] Preferably, p is screened based on the comparison area, and the original image set is updated, including:

[0131] Obtain a set CU of coordinates of pixel points in the comparison area;

[0132] Get the image feature tp of the area composed of the coordinates in CU in the image p;

[0133] Filter image p based on tp and update the original image set.

[0134] Specifically, image features can be obtained based on the LBP algorithm.

[0135] Preferably, the image p is screened based on tp to update the original image set, including:

[0136] Compute p and any image in the original image set except p as follows:

[0137] For any image t except p in the original image set, if the length of image t is greater than the length of image p, image t is reduced until the length of image t is equal to the length of image p, and image ct1 is obtained;

[0138] Get the image feature tct1 of the area in ct1 consisting of the coordinates in CU;

[0139] Calculate the similarity between tp and tct1. If the similarity is greater than the set similarity threshold, it means that image p and image t are the same. Delete image p or image t from the original image set to update the original image set. If the similarity is less than the set similarity threshold, keep p.

[0140] If the length of image t is less than the length of image p, image t is enlarged until the length of image t is equal to the length of image p, and image ct2 is obtained;

[0141] Get the image feature tct2 of the area in ct2 composed of the coordinates in CU;

[0142] Calculate the similarity between tp and tct2. If the similarity is greater than the set similarity threshold, it means that image p and image t are the same. Delete image p or image t from the original image set to update the original image set. If the similarity is less than the set similarity threshold, retain p.

[0143] During the comparison, the algorithm for obtaining the image features of ct1 or ct2 is the same as the algorithm for obtaining tp from p.

[0144] Preferably, the similarity threshold is set to 0.95.

[0145] Preferably, the architectural design requirements include style type, space requirements and layout requirements.

[0146] Specifically, the space requirements include the size requirements of various types of spaces, such as the size of a bedroom. The layout requirements include the distribution locations of various types of spaces, such as setting the kitchen on the left side of the living room.

[0147] Preferably, if Figure 2 As shown in Figure 1, the generative adversarial network consists of a generator and a discriminator. The generator is responsible for creating novel architectural designs, while the discriminator evaluates the difference between the generated design and the real design. This two-way learning mechanism not only improves the quality of the generated design, but also ensures the diversity and practicality of the design.

[0148] The generator's task is to create novel architectural designs. It usually takes a random noise signal as input, processes it through a series of neural network layers, and outputs a new design. These designs may be completely fictional, but should look real enough to mimic real-world architectural designs to some extent.

[0149] The role of the discriminator is to evaluate the difference between the design solutions produced by the generator and the real design solutions. It learns to distinguish the features between the generated design and the real design samples by analyzing the generated design and the real design samples. During the training process, the discriminator continuously improves its discrimination ability and strives to identify which are generated designs and which are real designs.

[0150] In the bidirectional generative and discriminative learning mechanism, the generator and the discriminator compete with each other and also promote each other. The generator tries to deceive the discriminator and create increasingly difficult to distinguish design solutions. The discriminator strives to improve its ability to identify the generated designs more accurately. Through this dynamic adversarial process, the generator is enhanced in its ability to imitate real designs, while also promoting the innovation and practicality of the designs. The core value of this mechanism lies in its ability to continuously self-optimize, continuously improve the quality and realism of the generated designs, while ensuring the diversity and practicality of the designs, making them more in line with the needs of practical applications.

[0151] Preferably, converting the design requirements into a vector representation includes:

[0152] Calculate the text feature values ​​of design requirements;

[0153] Adaptively select conversion algorithms based on text feature values;

[0154] The conversion is performed into a vector representation based on the selected conversion algorithm according to the design requirements.

[0155] Existing conversion algorithms are generally set in advance, and such conversion algorithms are not well suited to different word distribution characteristics, which can easily lead to inaccurate vector representations. For example, models that are more inclined to global statistics, such as the N-gram model, are more suitable for longer texts because they require enough words for statistics to obtain accurate conversion results.

[0156] Preferably, the calculation formula of the text feature value is as follows:

[0157]

[0158] txtval represents the text feature value, legth represents the text length of the design requirement, bslegth represents the set text length comparison value, numrpt represents the number of words that appear at least twice in the design requirement, numal represents the total number of words in the design requirement, rptu represents the set of words that appear at least twice in the design requirement, dist j It represents the maximum number of characters between two adjacent words j; s1 represents the length weight, s2 represents the repetition weight, and s3 represents the interval weight.

[0159] In addition to considering the length of the text, the text feature value of the present invention also considers the number of repeated words and the number of characters between repeated words. When the length of the text is larger, the number of repeated words is larger, and the average value of the number of characters between repeated words is larger, the model based on global statistics is more likely to be used as the algorithm for obtaining the vector representation. The text feature value can realize the adaptive determination of the conversion algorithm, which is conducive to improving the accuracy of the obtained vector representation.

[0160] Preferably, the text length comparison value is set to 200.

[0161] Preferably, the length weight is one-half, and the repetition weight and spacing weight are both one-quarter.

[0162] Preferably, the conversion algorithm is adaptively selected based on the text feature value, including:

[0163] If the text feature value is greater than the set text feature value threshold, the N-gram model or Doc2Vec model is selected as the conversion algorithm;

[0164] If the text feature value is less than or equal to the set text feature value threshold, Word2Vec, GloVe or FastText is selected as the conversion algorithm.

[0165] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A deep learning-based architectural design system, characterized in that: It includes data set processing module, training module, input module and generation module; The data set processing module is used to delete duplicate images in the original image set to obtain a screened image set; The training module is used to train the generative adversarial network using the screened image set to obtain a trained generative adversarial network; The input module is used to input architectural design requirements; The generation module is used to convert the design requirements into vector representations and input the vector representations into the trained generative adversarial network to obtain the architectural design image. Among them, duplicate images in the original image set are deleted to obtain a screened image set, including: Perform the following processing on each image in the original image set: S1, for image p in the original image set, calculate the screening coefficient of p; S2, obtaining the comparison area based on the screening coefficient; S3, screening p based on the comparison area and updating the original image set; S4, determining whether there are images in the original image set that have not been screened, if so, proceeding to S5, if not, the remaining images in the original image set constitute the screened image set; S5, randomly select an image that has not been screened from the original image set as a new image p, and enter S1.

2. The deep learning-based architectural design system according to claim 1, characterized in that: Calculate the screening coefficient of P, including: Convert p to grayscale image gp; The screening coefficient was calculated based on gp.

3. The deep learning-based architectural design system according to claim 2, characterized in that: Calculate the screening coefficient based on gp, including: The calculation formula of the screening coefficient is: seldix p represents the screening coefficient of p, mig and mag represent the minimum and maximum grayscale values ​​in gp respectively, num v Indicates that the grayscale value in gp is equal to the total number of pixels in v, num p Indicates the total number of pixels contained in p, cps indicates the preset contrast value, lts p represents the maximum area of ​​each connected domain in gp, als cp represents the area of ​​gp, w1 and w2 represent the quantity weight and area weight respectively.

4. The deep learning-based architectural design system according to claim 3, characterized in that: Get comparison areas based on screening factors, including: Use the following formula to calculate the gp block parameters: N gp Indicates the block parameters of gp, baseNum indicates the maximum value of the set block parameters, seldix mx Indicates the set screening coefficient comparison value; Get the largest positive integer n that satisfies the following inequality; 2(n-1)+1≤N gp ≤2n+1 Divide gp into (2n+1)×(2n+1) regions of equal area; Select the comparison area from all areas.

5. The deep learning-based architectural design system according to claim 4, characterized in that: Comparison areas were selected from all areas, including: Calculate the representative feature value of each region separately; The region showing the largest eigenvalue is used as the comparison region.

6. The deep learning-based architectural design system according to claim 5, characterized in that: The calculation formula for the eigenvalue is: dlval b represents the characteristic value of region b, ub is the set of all pixels in b, gry i is the gray value of i, bsc represents the number of pixels in ub, average represents the gray mean of pixels in ub, lstpx b It represents the maximum length of the line segment composed of pixels with the same grayscale in b, d1 represents the variance weight of the grayscale value, and d2 represents the length weight.

7. The deep learning-based architectural design system according to claim 5, characterized in that: Filter p based on the comparison area and update the original image set, including: Obtain a set CU of coordinates of pixel points in the comparison area; Get the image feature tp of the area composed of the coordinates in CU in the image p; Filter image p based on tp and update the original image set.

8. The deep learning-based architectural design system according to claim 7, characterized in that: Based on tp, image p is filtered and the original image set is updated, including: Compute p and any image in the original image set except p as follows: For any image t except p in the original image set, if the length of image t is greater than the length of image p, image t is reduced until the length of image t is equal to the length of image p, and image ct1 is obtained; Get the image feature tct1 of the area in ct1 consisting of the coordinates in CU; Calculate the similarity between tp and tct1. If the similarity is greater than the set similarity threshold, it means that image p and image t are the same. Delete image p or image t from the original image set to update the original image set. If the similarity is less than the set similarity threshold, keep p. If the length of image t is less than the length of image p, image t is enlarged until the length of image t is equal to the length of image p, and image ct2 is obtained; Get the image feature tct2 of the area in ct2 composed of the coordinates in CU; Calculate the similarity between tp and tct2. If the similarity is greater than the set similarity threshold, it means that image p and image t are the same. Delete image p or image t from the original image set to update the original image set. If the similarity is less than the set similarity threshold, retain p.

9. The deep learning-based architectural design system according to claim 1, characterized in that: Architectural design requirements include style type, space requirements and layout requirements.

10. The deep learning-based architectural design system according to claim 1, characterized in that: The generative adversarial network consists of a generator and a discriminator.

Images