Intelligent design of display space layout based on two-stage deep learning network

Abstract

In an age of big data and information overload, recommendation systems have evolved rapidly. Throughout the traditional design of interior spaces, the specialised nature of the work and the high rate of human involvement has led to high costs. With the continuous development of artificial intelligence technology, it provides a favourable environment for reducing the development cost of the system. This study proposes a two-stage modelling scheme based on deep learning networks for the intelligent design of display space layouts, divided into two parts: matching and layout, which greatly improves design efficiency. The research results show that through comparison tests, its prediction accuracy reaches more than 80%, which can well meet the matching requirements of household products. The training number of Epochs is between 15 and 30, its training curve tends to saturate and the best accuracy can reach 100%, while the running time of the hybrid algorithm proposed in this study is only 20.716 s, which is significantly better compared with other algorithms. The proposed hybrid algorithm has a running time of only 20.716 s, which is significantly better than other algorithms. The approach innovatively combines deep learning technology with computer-aided design (CAD), enabling designers to automatically generate display space layouts with good visibility and usability based on complex design constraints. This study presents an innovative application of the research methodology by combining quantitative and qualitative methods to analyse the data. The application of both methods provides a more comprehensive understanding of the problem under study and provides insight into the key factors that influence the results. The findings of this study can provide useful insights for policy makers and practitioners.

Keywords

Recommendation systems artificial intelligence deep learning network display space layouts matching requirements

1. Introduction

Layout design, in layman’s terms, is the correct arrangement and placement of objects. In modern engineering and practical life, there are many problems to be solved, such as container transport in freight yards and the planning and design of building structures [1], and layout design has a wide range of applications and great commercial value, so the ubiquitous layout design problem makes it universal and of far-reaching practical significance [2]. As most layout design problems are closely linked to space, such as the tailoring of one-dimensional space and the design of two-dimensional interior layout solutions, problems concerning the three-dimensional arrangement of cockpits and the arrangement of components in a limited space are also included [3]. In general, layout design involves a process that includes the selection of an appropriate layout form, the arrangement of objects in a specific space, and the optimization of the design to improve the efficiency of the layout. In addition, layout design also includes the selection of specific layout patterns, the determination of the interrelationships between objects, and the adjustment of objects to meet the requirements of the layout. In order for users to experience interior design more intuitively, as well as to meet their individual requirements for interior spaces [4], the design of interior spaces is heavy and repetitive for designers, resulting in inefficient design, and for general users, this non-professional design approach often leads to unsatisfactory results, limiting the web-based interior design industry to develop in a data-driven, efficient and personalised direction. Based on this, the study aims to propose a research method for intelligent design of display space layout based on a two-stage deep learning network, which has great application in furniture layout design, computer animation, military simulation and circuit layout.

Layout design has been widely used in many areas, such as container transport in freight yards, the planning and design of building structures, the tailoring of one-dimensional space, the design of two-dimensional interior layout solutions, and the three-dimensional arrangement of cockpits. However, due to the manual design approach, the design efficiency of interior spaces is low, and the results are often unsatisfactory. Therefore, this study aims to propose a research method for intelligent design of display space layout based on a two-stage deep learning network, which has great application in furniture layout design, computer animation, military simulation and circuit layout, aiming to improve the design efficiency and provide users with more personalized interior design experiences.

2. Related work

The literature review for this study includes various algorithms related to the design of intelligent display space layouts. These include traditional interior space design algorithms such as Interior Design Optimisation (IDO) and Space Filling Algorithms (SFA), as well as more modern algorithms such as Deep Learning Network based Modelling Schemes (DLNMS) and Hybrid Algorithms (HA) to name a few. In today’s information overloaded network era, recommendation systems are developing rapidly, which are used to obtain users’ interests by analysing their behavioural characteristics and item features, so as to provide them with appropriate recommendation services. This paper focuses on a two-stage deep learning network-based intelligent design method for display space layout, and now presents an exposition of the current status of domestic and international research on several mainstream algorithms. Li et al. [5] proposed a mutual correlation-based (FBACC) method to reduce the number of iterations by using generating adversarial images and then ignoring errors in smoothed regions in the case of continuous misclassification of second-stage labels. Lins et al. [6] allocated to by-products was reduced by 51.69% using a variety of methods that combined interrelated regions and sectors. As a result, the Sustainable Development Goals (SDG) were achieved. The elimination of physical waste and production losses can be achieved from a (re) layout project combined with CP by optimising areas, sectors, flows and processes. Song et al. [7] proposes a method for evaluating layout results based on user feedback. For rooms with non-rectangular floors, our algorithm can also handle this case using shape normalisation techniques. The experimental results show that the algorithm proposed by the institute is efficient and can meet the practical response requirements for online furniture layout. Zhang et al. [8] have designed an automated sleep grading system using deep neural networks. The study established a stable and reliable model with scoring criteria similar to those of human experts. The selection of the best sleep grading metric is an issue of interest in future research. Zhang et al. [9] proposed an online optimization method for constant force grinding controller parameters based on deep reinforcement learning Rainbow, and the results showed a 26.54% and 78.39% reduction in roughness for its empirically adjusted parameters compared to those with poorer sampling grinding performance, validating the effectiveness and practicality of the proposed method.

Hung et al. [10] used a two-stage approach based on numerical simulation and deep learning correction of field observation data from the Fukai wind farm, which allows fast querying and display of massive historical and daily forecast data on a web interface to predict wind energy production and schedule wind turbine shutdowns and maintenance sequences by using a hybrid SQL and HBase database system. Han et al. [11] provided an end-to-end semantic segmentation system for urban scenes, in which the point cloud deep learning network consists of three main components: (1) an effective sampling strategy under the point cloud space, (2) a point-like feature extraction module for efficient encoding of local features using spatial aggregation, and (3) a loss function to handle various types of non-equilibrium problems to improve the overall performance. To verify the correctness of the point cloud deep learning network, the study used two sets of profiles to validate it, and the results showed that the IoU reached an average of 70.8% in the Toronto 3D and Shanghai MLS profiles, and 73.9% in the Shanghai MLS profile. Liu et al. [12] proposed a new topology design method based on experimental and numerical analysis, and experimentally demonstrated that the improved CNN hidden layer topology outperformed the benchmark method in the classification task outperformed the benchmark method by 30% compared to the benchmark method. Aliakbari et al. [13] considered improving the performance of a new C4 robot as a coordinate measuring system (CMM) to achieve a quality-improved workspace without any physical changes to the structure, and the selection of high-quality workspace areas for intelligent operations helps to obtain accurate results in industrial applications (especially measurement tasks). Liu and Pu [14] designed an intelligent medical disease diagnosis and classification product based on theoretical studies of machine learning under big data in order to promote the application of machine learning in smart manufacturing, especially in medical products and smart hospitals, and came to the concluded that the medical diagnosis classification effectiveness of both methods remained above and below 90%, while the study proved that the corresponding accuracy could only be improved when fuzzy matching was applied to more positive examples.

Through domestic and foreign scholars research can be seen, at present for the deep learning network and display space layout intelligent design of more research, but the combination of the two research is less, therefore, the research is mainly based on two-stage deep learning network on display space layout intelligent design research, so that its research method can be fast and accurate to complete the intelligent design work.

3. Establishment of a display space layout method based on a two-stage deep learning network

3.1 Determination of the recommended model for display space matching

To address the lack of data, low efficiency and low labour costs of current intelligent design technologies, the study divided them into two categories: ’space matching’ and ’space layout’, in order to enable lay people to better complete interior design. Based on this, the study has designed different recommendation modes, and triggered different recommendation modes in different situations in a hybrid mode of conversion, thus providing users with a choice of homes and a layout of scenarios. In order to reduce the heavy workload of designers and general users in terms of home matching, and based on the large amount of design solution materials posted by professional designers on the interior design platform, a new combined recommendation model was built by combining a collaborative project-based filtering recommendation algorithm and a recommendation algorithm based on the content characteristics of the home project. In addition, the study also designed and developed a 3D virtual reality interactive system to enable users to quickly and accurately understand the space, allowing them to directly interact with the design elements, generating design solutions and increasing the efficiency of design. The system has been integrated with the layout recommendation model, which can automatically generate layout solutions, and can also generate 3D images with high realism and accuracy [15]. In conclusion, this study has made some progress in intelligent design technology, improving the efficiency of interior design and providing users with more choices in design and layout.

Based on the actual situation of the home project, an offline set of matchability tables for the home model was generated. After the user has made personalised customisation of the furnishable house, the system will recommend the most representative $N$ (TopN) furniture models based on the overall style of the house and the offline matchability based on the type of furnishable house, providing a useful reference for the matching design of interior spaces. In the recommendation algorithm, based on the calculation of similarity, the study proposes a new matching recommendation model which uses a further improvement of Content-Base (CB), a content-based recommendation algorithm that uses Euclidean distance to measure the similarity of images, as shown in Eqs (1) and (2).

$\displaystyle d(x,y)=\sqrt{\sum{(x_{i}-y_{i})^{2}}}$ (1)

$\displaystyle\textit{sim}(x,y)=\frac{1}{1+d(x,y)}$ (2)

In Eqs (1) and (2), $d(x,y)$ is the distance between $x$ , $y$ , and the Euclidean distance is a relatively small absolute distance, which is similar to $\textit{sim}(x,y)$ the greater the similarity, because the amount of data between vectors may have different degrees of difference, so the data need to be normalized first [16], and then use the Euclidean distance to measure the data, in measuring low-dimensional, vector size of the larger data, the use of Euclidean distance works better, and the Euclidean distance is shown in Fig. 1 as the metric for measuring image similarity in this study.

Figure 1.

Euclidean distance.

The rapid development of the Internet has given rise to the advent of the era of information explosion, and a large amount of information data has emerged, making the problem of information overload even more serious. In order to solve the problem of information overload, recommendation algorithms obtain effective recommendation data based on the analysis of user behavioural characteristics and item attributes, so as to provide better recommendation services to users. The advantages of collaborative filtering (CF) algorithms are that they do not require rigorous modelling of the subject, they do not require the subject’s attributes to have machine-understandable characteristics, and the recommendation mechanism is not dependent on the domain in which the subject is involved and can be commonly used in general [17]. Overall, the CF algorithm is at its core based on historical data and is therefore prone to cold-start problems and its recommendation results are closely related to historical data. In addition, the recommendation accuracy issues that arise when using sparse matrices cannot be ignored as more and more data is available for the study population. In contrast to the CF algorithm, the content-based (CB) recommendation algorithm uses the similarity between feature vectors to provide recommendations to users after constructing a feature vector of the target based on the intrinsic characteristics of the object under test. The study of the item-based CF algorithm establishes a matching matrix that can reflect the degree of matching between items, and due to the complexity of the matching data and the limitations of the data set, making a large number of zero values in the matching matrix, if the recommendation pattern is too small, then the effectiveness of the recommendation will be reduced, the cold start problem in the CF algorithm is also an urgent need to solve, and the CB recommendation algorithm can just fill the gap of this research [18]. The design of 3D models was investigated and a recommendation algorithm based on engineering content was proposed, which has a complex and inefficient computational process. In order to make the CB recommendation algorithm faster, the study proposed a convolutional neural network-based feature extraction algorithm for home engineering images, which was further compressed by PCA. The underlying structure of the neural network is composed of artificial neurons, and the results of the neurons are shown in Fig. 2.

Figure 2.

Single neuron structure.

A neuron is a basic computational unit and is the most important component of an artificial neural network model. It consists of an input, an output and an activation function. The input is the signal that the neuron receives from other neurons, usually determined by a series of weights and biases. The input is the neuron’s input, which is weighted and summed over the input signal and then fed into the activation function. The activation function is a function used to convert the weighted sum into an output. It works by letting the input signal pass a threshold, and when the input signal is greater than the threshold, the activation function converts the input signal into an output. The output is the signal that the neuron transmits to other neurons. It is the input signal that has been processed by the activation function, which is the output of the neuron. The arrows represent the direction of signal propagation from the input to the activation function, and the activation function to the output, indicating that the input signal has been processed by the activation function to obtain the output. The relationship between the output $x_{j}^{\prime}$ and the N inputs $x_{1},x_{2},\ldots,x_{N}$ of the jth neuron structure given in Fig. 2 can be expressed as shown in Eq. (3).

$\displaystyle X_{j}=\sum\limits_{i=1}^{N}{w_{i}\cdot x_{i}+S_{j}}$ (3)

In Eq. (3), $S_{j}$ is the feedback signal. The feature function $f$ has many optional functions that can be selected according to its input and output characteristics. The overall structure of the neural network is shown in Fig. 3.

Figure 3.

Neural network basic structure.

As shown in Fig. 3, it is a simple forward neural network consisting of an input layer, Hidden layer and Output layer, which organises the individual neurons together, but forward transmission alone is not sufficient to build a neural network model; in addition to forward operation between the input and output layers, inverse transmission from the output layer to the input layer must be used, and inverse transmission is a method that can incorporate reverse transfer network output errors, correct network parameters and train a supervised method [19]. The specific algorithm steps are shown in Eqs (4) to (9).

$\displaystyle z^{l}=w^{l-1}a^{l-1}+b^{l-1}$ (4) $\displaystyle a^{l}=f(z^{l})$ (5) $\displaystyle\delta^{L}=\left|{y-a^{L}}\right|*\sigma^{\prime}(z^{L})$ (6)

As shown in Eqs (4) to (6), $x$ represents the input matrix, $l$ is the number of layers of the current network layer, $z$ is the linear sum, $a$ is the value obtained by non-linear transformation through the activation function, $b$ is the bias of the current layer, $w$ is the weight, Eq. (4) is the linear operation of each layer, and Eq. (5) uses the activation function to enhance the non-linear representation of the network. The partial derivatives are then calculated, and the equations are shown in Eqs (7) to (8).

$\displaystyle\frac{\partial f}{\partial b_{j}^{l}}=\delta_{j}^{l}$ (7) $\displaystyle\frac{\partial f}{\partial w_{kj}^{l}}=a_{k}^{l-1}*\delta_{j}^{l}$ (8)

The parameters $w$ , $b$ are updated according to the direction of the partial differential and are calculated as shown in Eqs (9) and (10).

$\displaystyle w_{kj}^{l}\to(w_{kj}^{l})^{\prime}=w_{kj}^{l}-\eta\frac{\partial f% }{\partial w_{kj}^{l}}=w_{kj}^{l}-\eta*a_{k}^{l-1}*\delta_{j}^{l}$ (9) $\displaystyle b_{j}^{l}\to(b_{j}^{l})^{\prime}=b_{j}^{l}-\eta\frac{\partial f}% {\partial b_{j}^{l}}=b_{j}^{l}-\eta\delta_{j}^{l}$ (10)

As shown in Eqs (9) and (10), where the parameters $w$ , $b$ are updated in the partial differential direction and $\eta$ is the learning rate, the above five steps are repeated until an appropriate network model is obtained. In a fully-connected network, the nodes of each layer participate in any of the nodes in the current layer, and when the input to the network is image data, the presence of a high-dimensional pixel matrix makes the network model more complex to operate. Whereas in a convolutional neural network the nodes in the current layer are only connected to local regions in the previous layer, Fig. 4 illustrates briefly how CNNs work, i.e. alternating connections between convolutional and pooling layers, typically with one or more fully connected layers added at the end, in which case the structure of the convolutional neural network is designed to extract useful feature information from the image, making it more resilient, and by pooling the filtering, so that features are formed throughout the image in the entire connected layer [20, 21].

Figure 4.

How convolutional neural networks work.

The working process of the convolutional neural network is shown in Fig. 4. Based on this, the study proposes an image-based matching recommendation model that takes the subtle differences between images, selects the pixel information that matches the size of the network input layer by capturing and reading the image, and unifies the image into a 224*224*3 size image. The convolutional layer uses convolutional operations to extract the most basic image features from an array of image pixels, and as the number of image layers increases, the structure becomes more complex. The convolution operation can be used not only for the original pixel data, but also for the extraction of features that are already fundamental to the convolution operation and can be performed with the convolution method [22]. The graph used in the study is a 224*224*3 matrix, so the convolution is computed as a three-dimensional convolution, which is similar to a one-dimensional convolution. The research divides the data set into a training set and a test set, and uses the training set to design a matching recommendation model, and identifies the recommendation model based on the recommendation results. The hybrid recommendation model proposed in the study will recommend the TopN most representative furniture items, and if there is a matching item in the TopN recommendations, then the result accuracy of the recommendation is up to standard, which is calculated as shown in Eq. (11).

$\displaystyle P=\frac{N_{right}}{N_{all}}*100\%$ (11)

Figure 5.

Offline construction flow chart with recommendation model.

In Eq. (11), $N_{\textit{right}}$ is the number of items with high recommended accuracy, $N_{\textit{all}}$ is the total number of items in the test set, and visual Geometry Group Network (VGG) is a convolutional neural network model proposed by Oxford University in 2014, which has promising applications in object recognition and image classification. VGG has been rapidly developed for its practicality and simplicity, and has been widely used in image recognition. It consists of 13 convolutional layers, 5 pooling layers and 3 fully connected layers. The purpose of extracting image features is mainly to find similar items by performing similarity analysis between similar items in the same type of house, matching zero-valued items and thus estimating them, making the sparse matching pattern more intensive and increasing the recommendation accuracy. Using Principal Component Analysis (PCA) algorithm, the above-mentioned feature vector is dimensionalised, thus maintaining better feature values and thus greatly speeding up the operation. PCA is a common dimensionality reduction algorithm for a $n$ dimensional feature vector, and to dimensionalise it to $k$ dimensional, the $k$ primitives must be selected and mapped to a new The process is shown in Eqs (12) to (14).

$\displaystyle A=[X_{1}\ X_{2}\ X_{3}\ldots X_{i}\ldots\ X_{1000}]^{T}$ (12) $\displaystyle A_{\textit{cov}}x=\textit{ax}$ (13) $\displaystyle P=[x_{1}\ x_{2}\ x_{3}\ldots x_{k}]^{T}$ (14)

In Eqs (12) to (14), $X_{i}$ is a 4096-dimensional vector that is also the eigenvector of the $i$ image, $\alpha$ is the eigenvalue, and $x$ is an eigenvector associated with that eigenvalue and normalized to it.

By normalising this matrix, PX is the eigenvector array that is reduced to $k$ dimension. The operational flow of the recommendation algorithm is shown in Fig. 5.

3.2 Determination of the layout model of the display space collocation

Long Short Term Memory networks (LSTM) are a special type of recurrent neural network that can efficiently handle time intervals and can effectively delay more sequential information. Figure 6 shows the internal construction of the LSTM network. The LSTM utilises three gates at each moment to control the current state of the unit.

The forgetting matrix is calculated by the method of Eq. (15) with a joint matrix consisting of the input at the previous time $h_{t-1}$ and the input at this time $x_{t}$ which is represented by $w_{f}$ for the weighting matrix and $b_{f}$ for the bias and this output forgetting matrix $f_{t}$ reflecting the selection of the cell status matrix at the previous time.

$\displaystyle f_{t}=\sigma\left({W_{f}*[h_{t-1},x_{t}]+b_{f}}\right)$ (15)

Input gates are used to determine the information to be added to the cell state, as shown in Eqs (16) and (17), with the same input form as the forgetting gate, where the Sigmoid layer determines the information needed to update the original cell state $i_{t}$ , and the tanh layer generates the update content $\tilde{C}_{t}$ , multiplied by the two corresponding positions to form the cell state update matrix.

$\displaystyle i_{t}=\sigma\left({W_{i}*[h_{t-1},x_{t}]+b_{i}}\right)$ (16) $\displaystyle\tilde{C}_{t}=\tanh\left({W_{c}*[h_{t-1},x_{t}]+b_{c}}\right)$ (17)

$h_{t}$ The output signal depends on the state of the cell, in Eq. (18) the sigmoid layer described determines the information to be output, and the final output is obtained by combining Eq. (19) with the desired output cell state $C_{t}$ .

$\displaystyle o_{t}=\sigma\left({W_{0}*[h_{t-1},x_{t}]+b_{0}}\right)$ (18) $\displaystyle h_{t}=o_{t}*\tanh(C_{t})$ (19)

The results of the two types of division methods proposed in the study are closely related to both the first half and the second half of the sequence, so the two-way LSTM algorithm model is used for layout recommendation. The accuracy of the study is calculated as shown in Eq. (20).

$\displaystyle P_{\textit{accuracy}}=\frac{N_{t}}{N}*100\%$ (20)

In Eq. (20), $N$ is the total number of test data, and $N_{t}$ is the complete sequence of predicted results. For the actual modeling, a set of LSTM superparameters must be decided first, and the LSTM superparameter settings described in the study are shown in Table 1.

Table 1

LSTM network training parameters

Network parameters	LSTM network	Network parameters	LSTM network
Enter the vector dimension	1	The number of network layers	5
The output vector dimension	1	Learning rate	0.005
The number of hidden neuron nodes	200	The maximum number of iterations	200

Figure 6.

LSTM network internal structure diagram.

It should be added here that the data set used in the design network model, including the long segmentation after precise segmentation, is less accurate in terms of its layout results if the long strip in the unit household type to be laid out cannot be precisely divided at present. In the layout function section, the layout problem is transformed into a segmentation problem of segmented sections and household segments, the layout information in the scene is digitised, and the cross features between individual units are extracted using a word embedding algorithm, and the feature matrices of the two network models are dimensionalised using a bi-directional LSTM method, and the corresponding model parameters are established by two different LSTM methods.

Figure 7.

Layout recommendation model construction flow chart.

4. Analysis of the application effect of research on intelligent design of display space layout based on two-stage deep learning network

According to the above algorithm procedure, the PCA dimensionality reduction pcaNum, the number of repeated updates of the matching recommendation model iterNum, and the number of similar terms used to fill the zero values in the sparse matrix simNum, all have an impact on the final matching recommendation model. the recommendation matching model will recommend the top TopN home items with the largest matchability, the larger the TopN, the higher the accuracy of the recommendation The VGG-16 convolutional neural network has a strong sparsity and its dimension is 4096. In order to further improve the accuracy and efficiency of the algorithm, the PCA algorithm was used to reduce the dimension of the feature matrix of the image and divide it into four parts, namely simNum $=$ 4, pcaNum $=$ 4096, 512, 256, 128, 64 to perform a comparative test in which pcaNum $=$ 4096 means that it has not been dimensionalised by the PCA algorithm, the process of which is shown in Fig. 8 in detail.

Figure 8.

Comparative experiment diagram of pcaNum.

As shown in Fig. 8, Figures (a) and (b) are bar charts of the recommendation accuracy of the family matching recommendation models constructed by a single CF algorithm at TopN $=$ 10 and TopN $=$ 30, and Figures (c) and (d) are improved family matching recommendation models at TopN $=$ 10 and TopN $=$ 30. By comparing the recommendation accuracy of (a) to (d) for pcaNum $=$ 4096 and other dimensions, it can be seen that the performance of each model has been improved by using the PCA algorithm for dimensionality reduction, especially the recommendation accuracy of pcaNum $=$ 128 has been improved, in addition to comparing the recommendation accuracy of (a) and (c) for the case of pcaNum $=$ 128 with 66% and 70% and with pcaNum $=$ 128 the prediction accuracy was 73.50% and 88%, indicating a significant improvement in the accuracy of the model.

Table 2

Time test

TopN	$T$ (300 strips/s)	TopN	$T$ (300 strips/s)
1	0.3368	11	0.4715
2	0.3491	12	0.3146
3	0.4614	13	0.3098
4	0.3641	14	0.2918
5	0.3678	15	0.2996
6	0.3564	16	0.2861
7	0.3013	17	0.3098
8	0.3189	18	0.3547
9	0.3073	19	0.3199
10	0.3011	20	0.3413

Figure 9.

Comparative experiment diagram of iterNum.

When simNum $=$ 4, we chose iterNum $=$ 0, 1, 2 and 3 as comparison tests. With iterNum of 0 for the number of iterations, it means that the recommendation model was not added to the construction model of the CF recommendation algorithm based on content features, and as seen in Fig. 9, the recommendation accuracy was improved after repeated updates and iterations. In the case of the number of repetitions greater than 2, the accuracy curve of the recommendation is basically the same, because the iterative update of the matching recommendation model is carried out on the basis of repetition, and each iteration of the update from the second iteration is based on the previous update, therefore, the role played by the improvement of the accuracy of the recommendation model will gradually diminish in the process of subsequent iterations, so comprehensive above, this study chose iterNum $=$ 2 for subsequent controlled trials.

The matching recommendation pattern created by the Institute is offline, which only needs to be matched in offline mode, and then TopN recommendations are made. When new designs or materials are added to the database, the matching recommendation pattern is updated for online recommendations, Table 2 shows the data for 200 recommendations, for example, at TopN $=$ 5, 200 data were obtained in 0.3678 seconds recommendations. The length of the recommendation time varies slightly due to different data lengths in the test tuples and variations in the experimental apparatus, but overall the model is able to provide recommendations for home matching in real time.

Figure 10.

Layout network model training set accuracy curve and loss value curve.

The study provides a detailed analysis of the selection of parameters in the experiment through an example of the adjustment of the parameters of the planar layout network model, and the accuracy and loss value curves of the training set of the layout network model shown in Fig. 10(a) and (b) are obtained through a comparative experiment of OUTPUT_DIM, where the learning rate is 0.1, the number of hidden layers HIDDEN_SIZE is 128 and the BATCH_SIZE is 256. With OUTPUT_DIM values of 256 to 512, the training times of Epochs are between 15 and 30, and their training curves tend to saturate with optimal accuracy up to 100%. In order to reduce the complexity of the model and speed up the training, we choose an OUTPUT_DIM of 256.

Table 3

Running time comparison

Layout algorithm	Layout runtime (s/1000 share)
The layout algorithm of this paper	20.716
Distance Field Layout Algorithm	528.640
Placement Field Layout Algorithm	101.398

In Table 3, the layout algorithm studied is compared with other typesetting algorithms. Here, 1000 samples to be laid out are randomly selected and the overall performance of the algorithm is evaluated based on the layout time of each sample. In addition, in Table 3, the running times of the distance field layout algorithm and the placement field layout algorithm are estimated based on the authors’ experimental results, while the time used for the study is the sum of the segmentation and layout times, and by comparing them, we can see that the solution has better real-time processing capability. The RMSE and $R^{2}$ of the model’s prediction results are shown in Table 4.

Table 4

RMSE and $R^{2}$ for model on datasets

TopN	Algorithm	Training		Testing
		RMSE	$R^{2}$	RMSE	$R^{2}$
1	Distance field layout	30.85	0.8545	32.11	0.8351
	The paper	3.56	0.9982	3.18	0.9915
2	Distance field layout	31.49	0.8214	30.98	0.8211
	The paper	3.18	0.9451	3.09	0.9569
3	Distance field layout	29.65	0.7534	33.89	0.7896
	The paper	3.45	0.9932	3.09	0.9973

As can be clearly seen from Table 4, the RMSE and R2 values of the algorithm proposed by the study for TopN $=$ 1, for example, are 3.56 as well as 0.9982, indicating that the model has a good prediction effect, and the same is true for the test set, where the RMSE and R2 values are 3.18 and 0.9915, respectively, indicating a significant model effect, indicating that the algorithm proposed by the study for the matching recommendation algorithm of the home model and the layout The study of the scheme recommendation algorithm can further optimise the accuracy of the algorithm.

5. Conclusion

In recent years, people’s requirements for the quality of living environment have become higher and higher, and the intelligent interior space design industry has also emerged. In order to improve the efficiency of intelligent design, this paper takes the interior space design software platform as the background and studies the matching recommendation algorithm and layout scheme recommendation algorithm for 3D household models. feature extraction of images, a new content-based recommendation method is constructed, and after data processing of scenes, a partition network model and a layout network model are established using bidirectional LSTM to pre-segment the household area and obtain layout results. The results showed that the prediction accuracy was 73.50% and 88% at pcaNum $=$ 128, indicating a significant improvement in the accuracy of the model. The study selected iterNum $=$ 2 for a subsequent controlled trial, and the study showed that 200 recommendations were made, for example, at TopN $=$ 5, and 200 recommendations were made in 0.3678 seconds Overall, the model was able to provide recommendations for family matching in real time, with a learning rate of 0.1, a number of hidden layers HIDDEN_SIZE of 128, a batch_size of 256, and a training count of Epochs between 15 and 30 at OUTPUT_DIM values of 256–512, with a training curve that saturated and an optimal accuracy can reach 100%, indicating that the research on the intelligent design of display space layout based on two-stage deep learning network proposed in this study has good practical significance.

Research on matching recommendation algorithms for home models and layout plan recommendation algorithms can further optimise the accuracy of the algorithms so that they can recommend the user’s favourite home models and layout plans more accurately. More advanced machine learning techniques, such as deep learning and reinforcement learning, can be used to improve the recommendation of home models and layout solutions. It is also possible to improve the accuracy of the models by continuously increasing the training samples to better meet users’ needs; and to introduce more intelligent technologies such as natural language processing, computer vision, etc. to improve the functionality of the intelligent interior design software platform so as to better meet users’ needs.

Footnotes

Funding

The research is supported by the Jiangsu Province University Philosophy Social Science Fund Project entitled “Research on the protection mechanism and display and dissemination strategy of the intangible cultural heritage of the Grand Canal (Jiangsu Section) under the background of cultural ecology” (Project No.2022SJYB1842).

References

Yan

Wang

Jiang

Cai

Zhang

Liu

Cai

. Classification of hyperspectral images using a propagation filter and convolutional neural network. Remote Sens Lett. 2022; 13(5): 429-440.

Cooper

Krishnan

Bhat

. Deep learning and the future of the model for end-stage liver disease-sodium score. Liver Transplant. 2022; 28(7): 1128-1130.

Xiang

Wang

Pan

Chen

Zhou

Zhang

. Accelerated topology optimization design of 3D structures based on deep learning. Struct Multidiscip O. 2022; 65(3): 99.

Yang

Song

Jiang

Ling

. Automatic interior layout with user-specified furniture. Comput Graph-UK. 2021; 94: 124-131.

Cheng

Yang

Lei

Shen

. A fast two-stage black-box deep learning network attacking method based on cross-correlation. CMC-Comput Mater Con. 2020; 64(1): 623-635.

Lins

Kiperstok

Cunha

RDA

Rapôso

ALQRES

Merino

EAD

Cesar

. (Re) layout as a strategy for implementing cleaner production: Proposal for a furniture industry company. Sustainability. 2021; 13(23): 13109-13109.

Song

Zheng

Jia

Gao

. Web3D-based automatic furniture layout system using recursive case-based reasoning and floor field. Multimed Tools Appl. 2019; 78(4): 5051-5079.

Zhang

Wang

, et al. Automated multi-model deep neural network for sleep stage scoring with unfiltered clinical data. Sleep Breath. 2020; 24(2): 581-590.

Zhang

Yuan

Zou

. Online optimization method of controller parameters for robot constant force grinding based on deep reinforcement learning rainbow. J Intell Robot Syst. 2022; 105(4): 85.

10.

Hung

Chang

Tsai

Kuo

Liau

Lin

, et al. Development of high-precision wind, wave and current forecast system for offshore wind energy industry in Taiwan: a two- stage method of numerical simulation and AI correction. J Chin Inst Eng. 2021; 44(6): 532-543.

11.

Han

Dong

Yang

. A point-based deep learning network for semantic segmentation of MLS point clouds. ISPRS J Photogramm. 2021; 175: 199-214.

12.

Liu

Shao

Jiang

Deng

. CNN-based hidden-layer topological structure design and optimization methods for image classification. Neural Process Lett. 2022; 54(4): 2831-2842.

13.

Aliakbari

Mahboubkhah

Sadaghian

Barari

Akhbari

. Computer integrated work-space quality improvement of the C4 parallel robot CMM based on kinematic error model for using in intelligent measuring. Int J Comput Integr Manuf. 2022; 35(4-5): 444-461.

14.

Liu

. Analysis and research on intelligent manufacturing medical product design and intelligent hospital system dynamics based on machine learning under big data. Enterp Inf Syst. 2022; 16(2): 193-207.

15.

Zhang

Yang

Fan

. Semisupervised image classification by mutual learning of multiple self-supervised models. Int J Intell Syst. 2022; 37(5): 3117-3141.

16.

Hong

Wang

Chen

Yang

Chen

. Rapid assessment of gasoline quality by near-infrared (NIR) deep learning model combined with fractional derivative pretreatment. Anal Lett. 2022; 55(11): 1745-1756.

17.

Liao

Zhang

Huang

. Intelligent structural design of shear wall residence using physics-enhanced generative adversarial networks. Earthq Eng Struct D. 2022; 51(7): 1657-1676.

18.

Umair

Iqbal

Khan

Kadry

. A deep learning based method for network application classification in software-defined IoT. Int J Uncertain Fuzz. 2022; 30(3): 463-477.

19.

Devi

TAM

Darwin

. Hyper spectral fruit image classification for deep learning approaches and neural network techniques. Int J Uncertain Fuzz. 2022; 30(3): 357-383.

20.

Jia

Wang

. Network intrusion detection algorithm based on deep neural network. IET Inform Secur. 2019; 13(1): 48-53.

21.

Selvakumar

Muneeswaran

. Firefly algorithm based feature selection for network intrusion detection. Comput Secur. 2019; 81: 148-155.

22.

Noferesti

Jalili

. Inline high-bandwidth network analysis using a robust stream clustering algorithm. IET Inform Secur. 2019; 13(5): 486-495.