Indoor space intelligent design method based on improved Resnet neural network

Abstract

In today’s rapidly developing era, intelligent indoor design can not only improve the efficiency and comfort of space use, but also bring new experiences to life and work. However, intelligent design of indoor spaces currently lacks the ability to extract features, and the quality and diversity of generated images are limited. To improve the effectiveness of intelligent indoor space design, in terms of recognition model, the MobileNetV3 model is adopted and a simple parameter free attention mechanism is taken to improve it. In the segmentation model, the study adopts a positional segmentation object model and improves its residual network. In the intelligent design of indoor spaces, an intelligent image generation model based on generative adversarial networks is designed, which includes a three-level generator. The recognition accuracy of the recognition model was 95.98%, which outperformed comparison models. Themaximum accuracy of the segmentation model was 98.36%. The minimum time required to design an intelligent image generation model was 13.55 s, and the rationality, aesthetics, and innovation scores were all better than the comparison models. The recognition, segmentation, and intelligent image generation models can provide support for intelligent design of indoor spaces. The contribution of the research is to construct an integrated indoor space intelligent design method of ‘recognition segmentation generation’, which improves the effectiveness of indoor space intelligent design.

Introduction

In the architecture and design, interior space design has always been a core element. People’s demand for functionality and experience in indoor spaces continues to rise. However, the limitations of traditional design models on efficiency and personalization are gradually becoming apparent^1,2. Driven by artificial intelligence technology, especially the deep learning in image recognition and generation, intelligent interior space design has gradually become a research hotspot³. This design method based on artificial intelligence can not only improve the efficiency and comfort of space use, but also bring new experiences to people’s lives and work. Common indoor space intelligent designs include Convolutional Neural Networks (CNN), Generative Adversarial Networks (GANs), genetic algorithms, and neural network algorithms^4,5. Meanwhile, many scholars have conducted research on intelligent design of indoor spaces and related content.

Chen S et al. proposed a parameterized design method based on tree structure interactive evolutionary algorithm for office furniture and its spatial layout design. Meanwhile, the study also designed an improved flexible neural tree model. The parameterized design strategy had significant advantages. The flexible neural tree algorithm based on similarity evaluation had no outliers in the selected dataset, with efficient stability⁶. Zhang Y et al. designed an interior space design recognition system to optimize the artistic quality and user experience of interior design. The system used long short-term memory networks and CNN to complete spatial layout feature recognition tasks, and optimized the CenterNet algorithm. The method took less than 5 min and had an average accuracy of 91.0%⁷. Mostafavi F et al. proposed an image-based deep learning prediction model pix2pix to evaluate the overall sunlight, energy, and ventilation performance. The average structural similarity index of the predicted illuminance map was 0.86⁸. Ribino P et al. transformed the furniture layout task into a multi-objective Markov decision process for optimizing indoor furniture layout, and implemented it through reinforcement learning. Meanwhile, the study clarified the reward functions for thermal, acoustic, and visual comfort aspects. The results showed that the proposed method could determine the optimal furniture layout and optimize indoor furniture layout design⁹.

However, in the current field of intelligent indoor space design, most research only focuses on a single link, such as individual scene recognition or image generation, lacking integrated design of the entire process of ‘scene recognition instance segmentation image generation’, and some methods have weak feature extraction capabilities and limited image quality and diversity, making it difficult to meet the efficiency and accuracy requirements of actual design scenarios.To achieve intelligent design of indoor space, an indoor scene recognition model based on improved MobileNetV3 and a segmentation model based on improved Segmenting Objects by Locations version 2 (SOLOv2) are designed. The Residual Network 50 (ResNet50) in SOLOv2 is replaced and improved. Finally, an intelligent image generation model relying on improved GAN is designed. The study aims to construct an efficient, precise, and intelligent method for interior space design, promoting the interior space design towards intelligence. The innovations of this study are: (1) Filling the gap in full-process integration by constructing an integrated "identification-segmentation-generation" design methodology: The research deeply combines material acquisition (identification and segmentation) of indoor scenes with the generation of design images, forming a complete intelligent design pipeline that spans from material extraction in the early design phase to scheme generation in the later stage, meeting the demand for full-process efficiency in practical design scenarios. (2) Addressing the weak feature extraction capability of certain methods by introducing the Simple Parameter-Free Attention Mechanism (SimAM) and a novel histogram equalization algorithm. SimAM enhances key feature representation and reduces interference from irrelevant information without increasing model parameters or computational costs, while the new histogram equalization algorithm optimizes the model’s adaptability to varying brightness conditions. Together, these improvements enhance the accuracy of feature extraction, providing high-quality data support for subsequent segmentation and generation stages. (3) Optimizing three core model categories—MobileNetV3, SOLOv2, and GAN. The improved MobileNetV3 enhances recognition accuracy and efficiency for indoor scenes, the optimized SOLOv2 improves instance segmentation accuracy, and the enhanced GAN (employing a three-tier generator structure) boosts the quality and diversity of design image generation. These optimizations effectively address the shortcomings of traditional models in recognition accuracy, segmentation performance, and image generation quality, meeting the precision requirements of practical design scenarios.

Methods and materials

To achieve intelligent design of indoor spaces, indoor scene recognition models and indoor scene instance segmentation models that can provide accurate material and data support have been designed. Subsequently, an intelligent generation model for indoor space design images based on a three-level generator is developed.

Construction of indoor scene recognition model based on improved mobilenetv3 model

To achieve intelligent design of indoor space, the research first constructs an indoor scene recognition model using improved MobileNetV3 and an indoor scene instance segmentation model relying on improved SOLOv2 to provide accurate materials and data support. Subsequently, an intelligent generation method for interior space design based on improved GAN is designed. In the indoor scene recognition model, the MobileNetV3 model is selected as the basis and improved from three perspectives to enhance the performance. MobileNetV3 is a lightweight deep learning network that combines techniques such as depthwise separable convolution, Squeeze-and-Excitation (SE) module, and Hard Swish activation function to significantly reduce computational complexity and model size while maintaining high accuracy. It has extensive applications in image recognition, real-time video processing, and object detection^10,11. Figure 1 displays the MobileNetV3.

Fig. 1.

The network architecture of MobileNetV3.

From Fig. 1, MobileNetV3 mainly includes input image, Fully Connected Layer (FCL), convolutional layer, inverse residual module, batch normalization, adaptive average pooling, Hard Swish activation function, and output. In the network architecture of MobileNetV3, the input image first goes through a series of convolutional layers and inverse residual modules. These modules utilize depthwise separable convolution techniques to significantly reduce computational complexity while maintaining the effectiveness of feature extraction. Meanwhile, the inverse residual module further optimizes the feature extraction process through three steps: extended convolution, deep convolution, and linear projection. The inverse residual module also embeds the SE module, which can generate channel attention weights through Global Average Pooling (GAP) and FCL to dynamically adjust the importance of feature channels and enhance the feature expression ability. In addition, the Hard Swish activation function introduces adaptivity in nonlinear transformations, improving the ability to fit complex features. After multiple layers of feature extraction, the feature map is dimensionality reduced through an adaptive average pooling layer, and the spatial dimension is compressed into a global feature vector. Finally, the global feature vectors are classified and processed through a FCL to output the final recognition result. To enhance the anti-interference ability of the MobileNetV3 model, three improvements are performed.

The first improvement is to take SimAM to replace the original SE module to enhance the feature extraction capability. SimAM is a lightweight and parameter free attention mechanism that has advantages such as parameter free design, strong applicability, and ease of integration and use^12,13. In addition, compared to the SE module, SimAM can avoid the interference of irrelevant information in indoor scenes on model recognition, improve the accuracy of scene feature extraction, and do not increase the number of model parameters and computational costs. At the same time, its parameter free design reduces the training parameter size of the model and reduces the risk of overfitting. At the same time, the regularization term of its energy function can further constrain the attention weight generation process, improve the generalization ability of the model in indoor scenes with different brightness and layout, and ensure the stability of feature extraction^14,15.SimAM is more suitable for lightweight MobileNetV3. Figure 2 presents the SimAM.

Fig. 2.

The structure of SimAM.

From Fig. 2, SimAM includes input feature maps, 3D attention weights, weight fusion, and output feature maps. The input feature map contains the quantity of channels, height, and width. After being processed by the SimAM module, these feature maps generate a three-dimensional weight matrix. The rectangular prisms of different colors in Fig. 2 represent different three-dimensional weights. Subsequently, the three-dimensional weights are multiplied element by element with the original input feature map to achieve weight fusion, thereby emphasizing important features and suppressing unimportant features, and enhancing the attention mechanism of the model. Finally, the feature map after weight adjustment is used as the output of the SimAM module. SimAM infers the three-dimensional attention weights of feature maps through an energy function. This energy function is shown in Eq. (1)¹⁶.

1

In Eq. (1), represents the target neuron. and are linear transformations of weights and biases, respectively. represents other neurons. and are linear transformations of and , respectively. represents the variable. is a constant. is the quantity of neurons on a certain channel. is the index of neurons. signifies the output value of the target pixel or target neuron after linear transformation. After minimizing and adding regularization terms, the final is shown in Eq. (2).

2

In Eq. (2), represents the regularization coefficient. In addition, the fast closed solution is shown in Eq. (3)¹⁷.

3

In Eq. (3), and represent the mean and variance of all neurons excluding . The rapid closure solution is shown in Eq. (4)¹⁸.

4

The minimum energy is shown in Eq. (5).

5

In Eq. (5), and represent the average and variance of all neurons in a certain channel, respectively. The second improvement is to take ELU activation function to replace the original RuLU in the stacked residual module. The advantage of the ELU activation function is that it not only provides nonlinear activation, but also reduces neuronal death by outputting negative exponential decay, and helps gradient flow and network convergence, achieving zero centralization of output¹⁹. In addition, compared to the Hard Swish activation function in the original architecture of MobileNetV3 and other classic activation functions commonly used in the field of deep learning, such as ReLU and Leaky ReLU, ELU can provide nonlinear activation, reduce neuron death by outputting negative exponential decay, assist gradient flow and network convergence, achieve output zero centralization, and adapt to complex brightness features in indoor scenes to reduce feature loss in dark areas, reduce the risk of model training oscillations in small sample data, and optimize computational efficiency in conjunction with MobileNetV3 lightweight architecture^20,21. Therefore, ELU was chosen for the study.The ELU activation function is shown in Eq. (6)²².

6

In Eq. (6), represents the input. is a hyper-parameter. Figure 3 presents the specific structure of the improved inverse residual module.

Fig. 3.

The improved inverse residual module.

From Fig. 3, the improved inverse residual module is mainly composed of SimAM, ELU activation function, deep convolution, dilated convolution, and Dropout. The third improvement of the MobileNetV3 model takes a new histogram equalization algorithm to optimize the adaptability to various brightness conditions. The process of the histogram equalization algorithm mainly consists of five steps, where the first step is to input the image. The second step is to obtain the pixel values of different components of red, green, and blue. The third step is to obtain brightness. The fourth step is to determine whether the image is overall dark/bright. If it meets the criteria, the global histogram equalization algorithm is used to process and output the image. Otherwise, the image enters the fifth step. Whether the image is locally dark/bright is determined. If it is, the contrast suppression adaptive histogram equalization algorithm is used to process and output the image. Otherwise, the image is directly output.

Design of solov2 instance segmentation model based on improved Resnet

This study has designed an indoor scene recognition model using improved MobileNetV3. To provide accurate material and data support for the intelligent generation of final indoor space design images, an optimized SOLOv2 instance segmentation model based on improved Resnet is proposed. Due to the complex and diverse indoor environment, the study performs indoor scene instance segmentation to obtain a more specific dataset. To construct the instance segmentation model, SOLOv2 is selected as the basis for the study, and improvements are made to its backbone network and neck network. SOLOv2 is an advanced instance segmentation algorithm with simplicity, intuitiveness, fast speed, high accuracy, no need for masked annotations, and the ability to handle different numbers of instances. It has a wide range of applications in fields such as autonomous driving, video surveillance, and robot vision that require fast and accurate image segmentation^23–25. SOLOv2 mainly includes structures such as backbone network, neck network, head network, loss function, and post-processing module. The backbone network is the foundation of the SOLOv2 to extract features from input images and lay the foundation for subsequent instance segmentation tasks. The neck network needs to further process and fuse feature maps to optimize the feature expression ability. Therefore, the feature and extraction fusion process of SOLOv2 (backbone network + neck network) is shown in Fig. 4.

Fig. 4.

The feature extraction and fusion process of SOLOv2.

From Fig. 4, the process includes the input layer, ResNet50 layer, extracted feature map, convolution layer, addition operation, up-sampling, and fused feature map. ResNet50 extracts features from input images, while up-sampling and feature fusion are accomplished through feature pyramids. ResNet50 can learn deep features of images by stacking multiple residual blocks and capture target information at different scales, with strong robustness^26,27. However, ResNet50 has a high computational cost. Meanwhile, the performance of the backbone network directly affects the model performance. Therefore, the study selects the ResNet with Next (ResNeXt) model to replace the original ResNet50. ResNeXt is an improvement of ResNet, with advantages such as modular design, efficient feature aggregation, strong flexibility, and high computational efficiency^28,29. To enhance the feature extraction capability, ResNeXt is improved by introducing a Spatial Attention Module (SAM) to optimize the segmentation accuracy. An example of the SAM is presented in Fig. 5.

Fig. 5.

The example of SAM.

From Fig. 5, the SAM mainly includes input feature maps, max pooling, average pooling, concatenation, multiplication, convolution + sigmoid activation function, feature maps of different sizes, and final output. Meanwhile, H, W, and C signify the height, width, and number of channels of the feature map. According to the SAM, key regions in the image can be identified and their feature representations can be enhanced, thereby improving the segmentation accuracy for the target. Meanwhile, the SAM can also suppress unimportant features and reduce their interference with model decision-making³⁰. Therefore, a SAM is added at the end of each ResNeXt layer. To optimize the accuracy of feature pyramid in feature fusion, two improvements are made to it. Improvement one is to replace the nearest neighbor algorithm used in up-sampling with bilinear interpolation method. The principle of bilinear interpolation method is to estimate the value of the target point by weighted averaging four nearest neighbors on a two-dimensional plane, which has high accuracy and computational efficiency.

The second improvement is to introduce a Channel Attention Module (CAM) into the feature pyramid to optimize the feature fusion performance. The CAM mainly consists of GAP, global maximum pooling, and activation functions, which can help the model automatically select the most useful features for the current task. It can also enhance important features and suppress unimportant features by adjusting the channel weights, and improve the performance. The CAM adopts the sigmoid activation function, which is expressed as Eq. (7)^31,32.

7

In Eq. (7), represents the input variable of the sigmoid function. Therefore, the improved feature extraction and fusion process is shown in Fig. 6.

Fig. 6.

The process of improved feature extraction and fusion.

From Fig. 6, the improved feature extraction fusion process includes an input layer, ResNeXt layer, SAM, convolutional layer, up-sampling, addition operation, CAM, and the fused feature map. The improved feature extraction and fusion module is used to segment indoor instances such as sofas, tables, and TV cabinets, and serves as a dataset and material library for subsequent intelligent image generation methods for interior space design.

Design of intelligent image generation method for interior space design based on improved gan

This study designs an indoor scene recognition model relying on improved MobileNetV3 and an indoor scene instance segmentation model using improved SOLOv2 to provide accurate material and data support for the intelligent generation of indoor space design images. To construct an intelligent generation method for indoor space design images, a GAN model is adopted and improved through a three-level generator structure. Each level generator is designed. GAN is a deep learning model with a generator and a discriminator, which can generate high-quality images, achieve data augmentation, and improve model robustness. It has extensive applications in image generation, creative design, and medical image analysis, especially in image generation^33,34. Therefore, the main structure of the intelligent generation method for indoor space design images is shown in Fig. 7.

Fig. 7.

The main structure of intelligent image generation method for interior space design.

From Fig. 7, the intelligent generation method for indoor space design images mainly consists of three-level generators. The first layer generator mainly includes input image data , skip connections, gated convolution, encoder-decoder structure, and roughly generated indoor space design image . The second layer generator mainly consists of input image, gated convolution, residual module, and locally generated image . Meanwhile, the input image of the second layer generator is shown in Eq. (8).

8

In Eq. (8), represents the mask. represents the operation of multiplying by the original product. The third layer generator mainly includes input image , skip connections, encoder-decoder structure, gated convolution, attention module, and globally generated image. Meanwhile, the input image expression of the third layer generator is similar to that of the second layer. Based on the three-level generator structure, the research model can optimize image generation effects from different perspectives to avoid the limitations of a single network structure. In the discriminator of GAN, it mainly involves , convolutional layers, and FCLs. In addition, different generators also correspond to various loss functions. Therefore, when generating rough images, the effective pixel loss is shown in Eq. (9).

9

In Eq. (9), represents the original real image. represents the number of non-zero elements in . The occlusion area loss is shown in Eq. (10).

10

In Eq. (10), represents the number of zero elements in . The expression of global loss is shown in Eq. (11).

11

In Eq. (11), represents the distribution of the generated image. represents the metric function. The overall loss expression of the rough image generation process is shown in Eq. (12).

12

In Eq. (12), and both represent the proportion of loss weights. Meanwhile, the overall loss of the local image generation process is expressed as Eq. (13).

13

In Eq. (13), , , and are all loss weights. and respectively represent the effective pixel loss and occlusion area loss in the local image generation process. , and represent total variation loss, content loss, and style loss, respectively. The overall loss and in the global image generation process are similar. Therefore, the overall loss expression of the model is shown in Eq. (14).

14

In Eq. (14), represents the discriminator loss. The intelligent image generation method based on a three-level generator can generate different interior space design images, promoting the intelligent development of interior space design.

Results

To verify the research model, the experimental environment, experimental dataset, and comparison methods are selected for the study. Meanwhile, the study also sets the parameters of the model. In addition, in terms of comparative indicators, the study selects factors such as time consumption, memory usage, accuracy, and F1 score.

Performance verification of indoor scene recognition model

A complete testing plan was developed for the performance validation of indoor scene recognition models, instance segmentation models, and intelligent image generation models, which includes four stages: data preprocessing, testing environment deployment, indicator testing, result validation, and statistical analysis. Firstly, in the data preprocessing stage, the study conducted data filtering and format standardization, converting all images to JPEG format uniformly and adjusting the resolution according to different model input requirements. It also involved label normalization, data standardization, and scientific partitioning of the dataset. In the deployment of the testing environment, the operating system used in the study is Windows 10, the Central Processing Unit (CPU) is Intel Core i5-12600 K, the dynamic acceleration frequency is 4.9 GHz, the third level cache is 20 MB, and the maximum memory bandwidth is 76.8 GB/s. In terms of indicator testing, for the indoor scene recognition model, the study used Recall and Accuracy indicators, while for the indoor scene instance segmentation model, the study selected Accuracy and F1score indicators. For the intelligent image generation model, the study selected indicators such as total time consumption, peak signal-to-noise ratio, and structural similarity index. In the results validation and statistical analysis, independent sample t-test was conducted to verify the differences in indicators between the improved model and the comparative model. If p < 0.05, the difference was considered statistically significant and the improvement effect was effective.

The reason for choosing CPU for neural network training in the study is: firstly, the improved MobileNetV3, SOLOv2, and GAN core models all adopt lightweight architectures, which enable the models to meet training requirements on Intel Core i5-12600 K CPU without relying on Graphics Processing Unit (GPU). Secondly, the actual terminal application scenarios for intelligent design of indoor spaces are mostly office computers commonly used by designers. These devices use CPUs as the main computing core and are mostly not equipped with high-performance independent GPUs. Training with CPUs can simulate the running status of models in actual deployment scenarios, avoiding efficiency degradation caused by mismatched training and deployment environments.The datasets used in the experiment are SUN397 dataset and ADE20K dataset³⁵. The SUN397 dataset contains 397 scene categories and rich images, covering various indoor and outdoor scenes. The ADE20K dataset contains 20,210 images and covers scenes such as living rooms, bedrooms, and kitchens. Due to the fact that the publicly available datasets are mainly from abroad, and there are certain differences between foreign and domestic interior designs, the study also collects some indoor scene data. Finally, a total of 3,000 images are collected for the study, with 5 types of scenes and 600 images for each scene. In the division of the training and testing sets, a 7:3 ratio is used in the study. To mitigate the risk of overfitting caused by small datasets, various strategies were adopted during the training process to prevent overfitting, including data augmentation techniques such as random rotation, flipping, and color perturbation of images, as well as Dropout random deactivation and Early Stopping mechanisms. In addition, the SimAM parameter free attention mechanism introduced in the improvement of MobileNetV3 can constrain the attention weight generation process through the regularization term of the energy function, enhance the model’s generalization ability in indoor scenes with different brightness and layout, and reduce overfitting caused by a single data distribution. The ResNeXt backbone network used to improve the SOLOv2 segmentation model, with its modular design and efficient feature aggregation capability, can extract general features more stably under limited data, avoiding the model’s excessive dependence on local features of a small number of samples. At the same time, the added SAM module can focus on key region features, suppress redundant information interference, and further reduce the model’s overfitting to noisy samples. Improve the three-level generator structure of GAN generation model design, optimize the image generation process in stages, gradually learn the feature mapping rules from rough to fine under small data volume, and reduce the generation feature offset and overfitting caused by insufficient data. These targeted optimizations at the model architecture level, combined with data augmentation, Dropout random deactivation, and Early Stopping mechanisms, effectively alleviate the risk of overfitting on small datasets. Therefore, research can achieve stable training and good performance without relying on large-scale datasets.In terms of comparison methods, the study selects MobileNetV3 model, ResNet50 model, Densely connected convolutional Network 121 (DenseNet121), combination model 1 combining MobileNetV3, data augmentation, and transfer learning, and combination model 2 combining ResNet50, data augmentation, and knowledge distillation. Meanwhile, the model has 100 iterations and a learning rate of 0.001. The ablation experiment based on the improved MobileNetV3 indoor scene recognition model is presented in Table 1.

Table 1.

Ablation experiment based on improved MobileNetV3 indoor scene recognition model.

MobileNetV3	SimAM	ELU	New histogram equalization algorithm	Index
				Recall	Accuracy	Floating point operations	Parameter quantity
√	√	√	√	93.91%	94.18%	210MMac	2.14 M
√	√	√	×	90.55%	92.64%	217MMac	2.66 M
√	√	×	×	88.37%	89.85%	220MMac	2.97 M
√	×	×	×	78.32%	80.07%	229MMac	5.95 M

From Table 1, in the ablation experiment results of the indoor scene recognition model, the MobileNetV3 + SimAM + ELU + new histogram equalization model had better performance in recall, accuracy, floating point operations, and parameter quantity, with an accuracy value of 94.18%. Meanwhile, the accuracy values of the MobileNetV3 + SimAM + ELU, MobileNetV3 + SimAM, and MobileNetV3 were 92.64%, 89.85%, and 80.07%, respectively, which were 1.54%, 4.33%, and 14.11% lower than 94.18%. This indicates that the SimAM module, ELU activation function, and new histogram equalization algorithm introduced in the study can all improve the performance of the MobileNetV3 model and enhance its accuracy and speed in indoor scene recognition. In addition, statistical verification has shown that there is a p < 0.05 difference in performance between different module combinations, indicating that these differences are statistically significant.The recognition accuracy and time consumption are presented in Fig. 8.

Fig. 8.

Recognition accuracy and time consumption of different models.

From Fig. 8a, in terms of recognition accuracy comparison, the improved MobileNetV3 model performed better, followed by combination model 2, combination model 1, DenseNet121, and ResNet50 models, with MobileNetV3 model at the bottom. In terms of specific values, the maximum recognition accuracy of these six models was 95.98%, 93.03%, 92.56%, 90.33%, 89.17%, and 80.87%, respectively. In Fig. 8b, the maximum value of the improved MobileNetV3 was 32 s and the minimum value was 18 s. The maximum recognition time of MobileNetV3, ResNet50, DenseNet121, combination model 1, and combination model 2 were 89 s, 76 s, 62 s, 47 s, and 42 s, respectively, which were 57 s, 44 s, 30 s, 15 s, and 10 s longer than 32 s. The differences in recognition accuracy and time consumption between different models also showed p < 0.05, indicating that these differences are also statistically significant.In summary, the MobileNetV3 model has higher recognition accuracy and lower recognition time. This may be due to the introduced SimAM module in the study, which can enhance the expression ability of key features and optimize the recognition accuracy. Meanwhile, the parameter free design of SimAM module also avoids additional computational overhead, which helps to reduce the running time. The CPU utilization and memory usage are presented in Fig. 9.

Fig. 9.

CPU utilization and memory usage among different models.

In Fig. 9a, when comparing the CPU utilization, the maximum values of the improved MobileNetV3 model and five comparison models were 10.66%, 27.59%, 25.11%, 23.87%, 19.23%, and 15.60%, respectively. Compared with the improved MobileNetV3 model, the values of the five comparison models were reduced by 16.93%, 14.45%, 13.21%, 8.57%, and 4.94%, respectively. From Fig. 9b, the improved MobileNetV3 model designed in the study performed better on memory usage, with a maximum value of 14.98% and a minimum value of 10.73%. Meanwhile, the MobileNetV3 model performed at the bottom, with a maximum and minimum memory usage of 30.07% and 25.63%, respectively. In addition, the maximum memory usage of ResNet50, DenseNet121, combination model 1, and combination model 2 was 28.05%, 24.66%, 20.80%, and 17.54%, respectively.The differences in CPU utilization and memory usage between different models are all statistically significant with p < 0.05, indicating that these differences are also statistically significant. Overall, the improved MobileNetV3 model has lower CPU utilization and memory usage, resulting in better performance. This may be due to the introduced lightweight SimAM module in the study, which improves model recognition accuracy while reducing computational resource consumption.

Performance verification of indoor scene instance segmentation model

To verify the performance of the indoor scene instance segmentation model, the study also uses Windows 10 operating system and Intel Core i5-12600 K processor. On the dataset, the study also takes the UN397 dataset and ADE20K dataset. Meanwhile, the study also collects data on some common objects in domestic life through supplementary filming. Subsequently, the study performs data augmentation on these data, such as random cropping, horizontal flipping, etc. Finally, a total of 3,000 images are used in the study. In terms of comparison models, the study selects SOLOv2 model, DeepLabV3 + model, Pyramid Scene Parsing Network (PSPNet), combination model 3 combining DeepLabV3 + and U-shaped network, and combination model 4 combining PSPNet and You Only Look Once version 8 (YOLOv8). In addition, the initial learning rate of the model is 0.002, the number of iterations is 50, and the momentum parameter is 1.0. In terms of evaluation indicators, the study selects mean Intersection over Union (mIoU), mean Average Precision (mAP), accuracy, and F1 score for comparison. The mIoU and mAP0.5 among different models are presented in Fig. 10.

Fig. 10.

Comparison of mIoU and mAP0.5 among different models.

In Fig. 10, mAP0.5 represents the mAP value of the model when IoU = 0.5. From Fig. 10a, the maximum value of the improved SOLOv2 was 0.985, the minimum was 0.952, and the average was 0.979. Meanwhile, the maximum mIoU values of SOLOv2, DeepLabV3 + , PSPNet, combination model 3 and combination model 4 were 0.782, 0.803, 0.817, 0.895, and 0.921, respectively, which were 0.203, 0.182, 0.168, 0.090, and 0.064 less than 0.985. From Fig. 10b, in the comparison of mAP0.5, the improved SOLOv2 instance segmentation model performed better, with a maximum of 0.991 and an average of 0.986. Meanwhile, the DeepLabV3 + model performed at the bottom, with an average mAP0.5 value of 0.747. In addition, the average mAP0.5 values of SOLOv2, PSPNet, combination model 3, and combination model 4 were 0.763, 0.796, 0.805, and 0.862, respectively. The differences in mIoU and mAP^0.5 between different models also showed p < 0.05, indicating that these differences are also statistically significant.In summary, the improved SOLOv2 instance segmentation model has significantly better mIoU and mAP0.5 than comparison models. This may be due to the improved ResNeXt as the backbone network and the introduced SAM, which has powerful feature extraction and enhancement capabilities, resulting in a positive change in the segmentation accuracy of the model for the target, thereby affecting the segmentation effect, and improving mIoU and mAP0.5. The accuracy and F1 score of different models are shown in Table 2.

Table 2.

Comparison of accuracy and F1 scores of different models.

Method	Accuracy/%					F1 score
	Number of experiments					Number of experiments
	1	2	3	4	5	1	2	3	4	5
SOLOv2	84.05	84.38	84.63	84.15	85.52	0.823	0.853	0.830	0.870	0.860
DeepLabV3 +	86.89	88.80	86.61	88.25	85.34	0.867	0.894	0.888	0.857	0.896
PSPNet	89.39	89.78	88.68	88.66	90.30	0.899	0.898	0.900	0.890	0.910
Combination model 3	92.60	92.07	92.25	92.70	92.03	0.932	0.929	0.937	0.938	0.930
Combination model 4	92.28	92.45	92.21	92.95	92.27	0.939	0.947	0.936	0.938	0.945
Improved SOLOv2	95.35	96.27	96.28	98.36	96.50	0.956	0.955	0.979	0.978	0.962

From Table 2, in terms of accuracy comparison, the average values of the improved SOLOv2 and the five comparison models were 96.55%, 84.55%, 87.18%, 89.36%, 92.33%, and 92.43%, respectively. The average accuracy of the five comparative models was 12.00%, 9.37%, 7.19%, 4.22%, and 4.12% lower than 96.55%, respectively. In addition, in the comparison of F1 scores, the average value of the improved SOLOv2 exceeded that of comparison models, with a value of 0.966. The differences in accuracy and F1 scores among different models also showed p < 0.05, indicating that these differences are also statistically significant.Overall, the improved SOLOv2 has higher accuracy and F1 score. This may be due to the introduced improved feature pyramid fusion method in the study, whose precise feature fusion characteristics enable the model to better preserve and highlight key information when dealing with complex indoor scenes, further improving the classification and segmentation performance of the model, and resulting in significantly higher accuracy and F1 score than the comparison model. The time consumption and CPU utilization are presented in Fig. 11.

Fig. 11.

Time consumption and CPU utilization of different models.

From Fig. 11a, the improved SOLOv2 instance segmentation model designed for research had a more obvious advantage on time consumption, with a maximum of 158.70 s and a minimum of 112.65 s. At this time, the maximum time consumption of SOLOv2, DeepLabV3 + , PSPNet, and combination models 3 and 4 was 239.87 s, 210.19 s, 194.33 s, 170.22 s, and 165.45 s, respectively, all greater than 158.70 s. In Fig. 11b, for CPU utilization comparison, the maximum of the improved SOLOv2 and the five comparison models was 12.27%, 28.42%, 25.03%, 22.19%, 18.75%, and 17.68%, respectively. The differences in time consumption and CPU utilization among different models are also statistically significant with p < 0.05, indicating that these differences are also statistically significant.Overall, the improved SOLOv2 instance segmentation model has lower time consumption and CPU utilization, and better performance.

Performance verification of intelligent image generation method for interior space design

To verify the performance of the intelligent generation method for indoor space design images, the same experimental environment is used in the study. The material library is the processed data of indoor scene recognition models and instance segmentation models. In terms of comparison methods, the study selects GAN, Deep Convolutional GAN (DCGAN), Cycle-Consistent GAN (CycleGAN), combination model 5 combining GAN and U-shaped network, and combination model 6 combining context encoder and pixel recurrent neural network. The study invites 100 professionals to evaluate the rationality, aesthetics, and innovation of the model in interior space design. The comparison of time consumption and memory usage among different models is shown in Table 3.

Table 3.

Comparison of time consumption and memory usage rates among different models.

Method	Time consuming/s					Memory usage rate/%
	Number of experiments					Number of experiments
	1	2	3	4	5	1	2	3	4	5
GAN	72.80	73.62	68.91	70.88	71.01	33.88	33.93	32.55	30.87	35.88
DCGAN	36.34	36.78	39.89	38.19	38.30	30.08	30.40	30.13	32.11	30.03
CycleGAN	30.01	32.56	33.53	34.65	33.09	26.58	25.26	25.56	27.41	27.91
Combination model 5	26.64	28.03	28.86	29.78	25.89	20.25	21.74	19.91	18.80	18.27
Combination model 6	23.62	25.93	27.86	27.19	23.28	17.84	17.30	16.79	14.18	16.57
Improved GAN	13.55	19.06	16.38	14.90	14.17	10.76	9.66	10.17	9.10	11.59

From Table 3, the time consumption of the improved GAN indoor space design intelligent image generation method was below that of the comparison method, with an average time of 15.61 s. The average time consumption of the five comparison methods was 71.44 s, 37.90 s, 32.77 s, 27.84 s, and 25.58 s, respectively. In terms of memory usage, the improved GAN intelligent image generation method for indoor space design still had more obvious advantages, with an average value of 10.26%. Overall, the improved GAN has lower time and memory usage, and better performance. This may be because the improved GAN adopts a three-level generator structure, and its phased optimization feature enables the model to process and transmit information more efficiently when generating indoor space design images, reducing unnecessary computational redundancy and significantly reducing the time and memory usage required for image generation. In addition, there were significant differences in time consumption and memory usage between different module combinations, with p < 0.05, indicating that these differences were also statistically significant.

In order to further verify the performance of the research design image intelligent generation method, the current optimal method was selected for comparison, including the advanced model 1 combining Stable Diffusion and control network, the advanced model 2 combining Transformer and Diffusion, and the Transformer based time series GAN (advanced model 3)^36–38. The comparative method used in the study is an advanced method from 2023 to 2025, covering the field of intelligent image generation categories, which can comprehensively evaluate the performance of the research model. The comparison of Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index (SSIM) for different models is shown in Table 4.

Table 4.

Comparison of PSNR and SSIM among different models.

Method	PSNR					SSIM					References
	Number of experiments					Number of experiments
	1	2	3	4	5	1	2	3	4	5
Advanced model 1	25.123	25.098	25.156	25.076	25.112	0.821	0.815	0.825	0.812	0.82	Kaleta et al. 36
Advanced model 2	26.345	26.298	26.387	26.276	26.321	0.853	0.848	0.857	0.845	0.852	Liet al 37
Advanced model 3	24.876	24.821	24.901	24.843	24.867	0.812	0.808	0.818	0.805	0.811	Wanget al 38
Improved GAN	35.678	35.623	35.712	35.654	35.690	0.967	0.963	0.97	0.965	0.968	\

From Table 4, it can be seen that in terms of PSNR indicators, the ranges of Advanced Model 1, Advanced Model 2, and Advanced Model 3 are [25.076, 25.156], [26.276, 26.387], and [24.821, 24.901], while the PSNR values of Improved GAN are between [35.623, 35.712], significantly higher than the other three advanced models, reflecting better image generation quality and obvious advantages in pixel level image fidelity. On the SSIM metric, the SSIM value of Improved GAN is closer to 1, indicating that the images generated by Improved GAN are closer to real images in terms of structural similarity, and have better generation effects in overall structure, texture, and other aspects of the images. In addition, there were significant differences in PSNR and SSIM between different models with p < 0.05, indicating that these differences were also statistically significant. The comparison of the rationality, aesthetics, and innovation scores of different models in indoor space design is shown in Fig. 12.

Fig. 12.

Comparison of rationality, aesthetics, and innovation scores of different models in indoor space design.

From Fig. 12a, in terms of rationality, the maximum value of the improved GAN intelligent generation method was 9.853, and the minimum value was 8.522. Meanwhile, the maximum rationality scores of the five comparison methods were 7.661, 8.015, 8.369, 8.924, and 9.136, all of which were less than 9.853. In Fig. 12b, the improved GAN performed better on aesthetics, with a maximum value of 9.637. The bottom performing GAN had a maximum value of 6.134. From Fig. 12c, for innovation, the maximum of the improved GAN and the five comparison methods was 9.760, 5.217, 6.608, 7.139, 7.682, and 8.244, respectively. The differences in rationality, aesthetics, and innovation scores among different models also have p < 0.05, indicating that these differences are also statistically significant.In summary, the improved GAN method has better performance and can generate more reasonable, beautiful, and innovative interior space design drawings, promoting intelligent design of interior spaces.

Conclusion

Aiming at the intelligent design problem of indoor space, an indoor scene recognition model relying on the improved MobileNetV3 and an instance segmentation model based on the improved SOLOv2 model were designed. Meanwhile, an intelligent design image generation model based on the improved GAN was designed. The results show that the improved MobileNetV3 model has significantly better recognition accuracy and time consumption than the compared model. This may be due to the SimAM module, which can enhance the expression ability of key features and thus optimize the recognition accuracy. Meanwhile, the parameter free design of SimAM module also avoids additional computational overhead, which helps to reduce the running time. The performance of the segmentation model is also better than that of the comparison model.This may be due to the improved ResNeXt as the backbone network and the SAM. The powerful feature extraction and enhancement capabilities of this module have improved the segmentation accuracy for the target. The designed image intelligent generation model can generate higher quality images. The research model has good performance. The contribution of the research results to the field of interior design is to build an integrated intelligent design technology system of "scene recognition instance segmentation image generation", shorten the design cycle, reduce the hardware investment cost of small and medium-sized design enterprises, and solve the pain points of "low efficiency, high cost, and difficult balance between personalization and standardization" in traditional design through precise material extraction and personalized solution generation, providing the industry with efficient and universal intelligent design tools. However, the improved SOLOv2 instance segmentation model still needs further optimization in terms of training efficiency and resource consumption when dealing with large-scale datasets. Future research can optimize it through distributed training, knowledge distillation, and pruning techniques. And it can also refer to the channel pruning method based on category aware competition ratio optimization designed by scholars such as Hu W³⁹.

Author contributions

J.Y.W. processed the numerical attribute linear programming of communication big data, and the mutual information feature quantity of communication big data numerical attribute was extracted by the cloud extended distributed feature fitting method. J.Y.W. Combined with fuzzy C-means clustering and linear regression analysis, the statistical analysis of big data numerical attribute feature information was carried out, and the associated attribute sample set of communication big data numerical attribute cloud grid distribution was constructed. J.Y.W. did the experiments, recorded data, and created manuscripts. All authors read and approved the final manuscript.

Funding

This research received no external funding.

Data availability

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.