Indoor layout automatic generation method based on deep neural network and simulated annealing optimization

Abstract

The invention discloses an indoor layout automatic generation method based on a deep neural network and simulated annealing optimization, and the method comprises the three main steps: firstly, generating an indoor scene initial layout through employing a variational auto-encoder VAE; secondly, performing multi-constraint optimization on the initial layout by adopting a deep learning enhanced simulated annealing algorithm (DL-SA); and finally, converting the optimized layout into a three-dimensional indoor scene. According to the method, the deep neural network model and the simulated annealing optimization method are combined, efficient and attractive indoor layout generation is achieved, and the method is widely applied to the fields of smart home, indoor design and the like.

Description

Technical Field

[0001] The present invention relates to the field of indoor scene layout generation methods, and specifically to an automatic indoor layout generation method based on a deep neural network and simulated annealing optimization. Background Art

[0002] In today's field of interior design, automated layout generation technology is gradually becoming a forefront research topic. Traditional indoor layout design relies on designers' experience and intuition. This method is not only time-consuming and laborious but also difficult to meet the growing personalized and diverse needs. With the rapid development of computer graphics and artificial intelligence technologies, algorithm-based automated layout generation methods have brought new possibilities to interior design.

[0003] Traditional automated layout methods mainly rely on rule engines and heuristic algorithms. These methods generate layouts through predefined rules and strategies. However, this method shows obvious limitations when dealing with complex and unstructured scenarios. The flexibility of the rule engine is insufficient to adapt to changing design requirements, while heuristic algorithms such as genetic algorithms and simulated annealing algorithms often require a large amount of computing time for large-scale layout optimization problems and are difficult to obtain satisfactory results in a short time. In addition, existing deep learning solutions usually ignore the modeling of physical constraints, resulting in layouts where furniture may overlap or the spatial distribution may be unreasonable. Many layout generators are limited to two-dimensional plane design and cannot provide visualization of three-dimensional spatial relationships, restricting the intuitiveness and practicality of the design.

[0004] Facing these challenges, researchers have begun to explore methods that combine deep neural networks with traditional optimization techniques to achieve more intelligent and efficient layout generation. These methods utilize the powerful representation ability of neural networks to capture complex spatial relationships and ensure the rationality and practicality of the generated results by introducing physical constraints. In this way, automated layout generation technology can not only automatically generate reasonable layouts that meet physical constraints but also improve the aesthetics and functionality of the layout through multi-objective optimization. Summary of the Invention

[0005] Aiming at the problems existing in the prior art, the present invention proposes an automatic indoor layout generation method based on a deep neural network and simulated annealing optimization, which integrates a deep neural network and simulated annealing stochastic optimization to improve the automation degree and layout rationality of interior design. This method combines a deep neural network model with a simulated annealing optimization method to achieve efficient indoor layout generation.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows: An automatic indoor layout generation method based on a deep neural network and simulated annealing optimization, comprising the following steps: Step 1: Use a variational autoencoder (VAE). Take random noise sampled from a standard Gaussian distribution as the input of the VAE. Generate indoor scene layout information through the VAE. The indoor scene layout information includes furniture layout information. The furniture layout information includes the initial position, rotation angle, size, and category information of the furniture. Map the furniture layout information to the latent space through the encoder of the VAE, and then generate an initial indoor scene layout containing complete furniture layout information through the decoder of the VAE; Step 2: Use a deep learning enhanced simulated annealing algorithm (DL-SA) to perform multi-constraint optimization on the initial layout obtained in Step 1. The process of the DL-SA algorithm is as follows: a) Extract deep features based on a three-module neural network architecture. The three-module neural network architecture includes: a feature extractor that maps furniture attributes to a high-dimensional feature space; a scene transformer that processes the global spatial relationship between furniture through a multi-head self-attention mechanism; an optimization predictor that generates an optimization guidance signal; b) Under the control of the temperature parameter, generate candidate layouts according to the state generation strategy guided by the deep features, realizing the organic integration of the global exploration ability of simulated annealing and the semantic understanding ability of the neural network; c) Apply a multi-level constraint processing mechanism, including physical constraints, functional relationship constraints, and aesthetic constraints. Among them, physical constraints ensure the basic feasibility of the layout, functional relationship constraints ensure that the layout meets the actual usage requirements, and aesthetic constraints enhance the visual effect and spatial balance of the layout; d) Achieve efficient rotated rectangle collision detection through the separating axis theorem, accurately quantify the overlapping problem between furniture, and guide the collision repair process; e) Adopt a dynamic weight adjustment mechanism and a hierarchical priority strategy to adaptively balance various constraints at different optimization stages, ensuring that important constraints are preferentially satisfied; Thus, an optimized layout is obtained; Step 3: Based on the optimized layout obtained in Step 2, retrieve the corresponding 3D objects from the dataset, select the objects with the closest size in the corresponding category, and then place them according to the layout to construct a 3D indoor scene.

[0007] Furthermore, in Step 1, the VAE uses a multi-layer Transformer network to capture the complex spatial relationship between furniture and optimizes the representation ability of the latent space through an adaptive learning mechanism.

[0008] Furthermore, denote the multi-layer Transformer network in the variational autoencoder (VAE) as SceneEncoder, and introduce a loss function based on multi-scale feature fusion during the training of SceneEncoder.

[0009] Furthermore, the dynamic weight adjustment mechanism in Step 2 is implemented in three stages according to the optimization process. In the initial stage, it focuses on the physical constraint weight. In the middle stage, it enhances the functional relationship constraint weight. In the later stage, it increases the aesthetic constraint weight to achieve a progressive optimization strategy.

[0010] Furthermore, in Step 2, the deep learning enhanced simulated annealing algorithm (DL-SA) combines a strategy based on adversarial learning during training, and simulates potential layout interference factors by introducing an adversarial network.

[0011] Furthermore, the physical constraints in Step 2 specifically include boundary constraints and overlap constraints; the functional relationship constraints specifically include distance constraints; the aesthetic constraints specifically include center constraints, where the boundary constraints ensure that the furniture is within the room range; the distance constraints control the minimum distance between furniture; the overlap constraints avoid furniture penetration by calculating the intersection over union; the center constraints optimize the overall spatial distribution of the furniture.

[0012] Furthermore, the overlap constraints are implemented through a differentiable physics engine, including converting the collision detection process into a differentiable tensor operation, using axis-aligned bounding boxes and the separating axis theorem for collision detection, and optimizing the penetration penalty through backpropagation.

[0013] Compared with the prior art, the advantages of the present invention are as follows: The present invention proposes an indoor layout automatic generation method based on deep neural network and simulated annealing optimization, aiming to overcome the limitations of traditional methods by integrating a deep neural network model with a simulated annealing optimization method. This method generates an initial layout through the combination of a variational autoencoder (VAE) and a Transformer network, and uses a deep learning enhanced simulated annealing algorithm (DL-SA) for multi-constraint optimization, finally generating a three-dimensional indoor scene that meets the actual requirements. This method provides an efficient and flexible solution for indoor design, and can meet the diverse needs of modern design. Through this innovative technical means, designers can generate high-quality indoor layouts more quickly, significantly improving the design efficiency and effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is the algorithm flow chart of the embodiment of the present invention.

[0015] Figure 2 is the neural network architecture diagram based on the variational autoencoder VAE and Transformer.

[0016] Figure 3 Architecture diagram of the Deep Learning Enhanced Simulated Annealing Algorithm (DL-SA).

[0017] Figure 4 Indoor scene layout effect diagram generated by the embodiment of the present invention. Detailed implementation manners

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Step 1: Use the Variational Autoencoder (VAE) to generate an initial layout.

[0020] In this embodiment, a random noise sampled from a standard Gaussian distribution is used as the input of the Variational Autoencoder (VAE). The indoor scene layout information is generated through the Variational Autoencoder (VAE). The indoor scene layout information includes furniture layout information, and the furniture layout information includes the initial position, rotation angle, size, and category information of the furniture. The furniture layout information is mapped to the latent space through the encoder of the Variational Autoencoder (VAE), and then the decoder of the Variational Autoencoder (VAE) generates an initial layout of the indoor scene including the complete furniture layout information.

[0021] Specifically, SceneEncoder is used to construct the latent space mapping. SceneEncoder is an encoder based on a neural network that can convert the input scene information into a latent space representation. By introducing the Graph Attention Network (GAT), this encoder can effectively capture the spatial relationship between furniture and enhance the model's understanding of complex scenes. To improve the discrimination of features, the triplet loss is used for training, making the features of similar furniture closer and the features of different furniture more separated.

[0022] In this embodiment, the Variational Autoencoder (VAE) adopts a multi-layer Transformer network, denoted as SceneEncoder, to capture the complex spatial relationship between furniture. As Figure 2As shown, it is the network structure diagram of the embodiment of the present invention, showing a neural network architecture based on the variational autoencoder (VAE) and Transformer. This network is the core component for realizing the automatic generation of indoor scene layouts. From left to right in the figure are the input feature layer, the latent representation layer, the dense layer and the convolutional layer, the normalization layer, the self-attention mechanism layer, the feature aggregation layer, the scene optimization layer, and the optimized output layer. Among them, the self-attention mechanism layer implements the multi-head attention mechanism, which is used to process the mutual spatial relationships between furniture; the feature aggregation layer and the scene optimization layer jointly implement the multi-constraint optimization function. SceneEncoder optimizes the representation ability of the latent space through an adaptive learning mechanism, making the generated initial layout more reasonable in terms of spatial relationships.

[0023] In this embodiment, a loss function based on multi-scale feature fusion is introduced during the training of SceneEncoder. This loss function enhances the accuracy and diversity of the generated layout by aligning the geometric features of furniture at different scales. Specifically, the loss function aligns the overall layout of furniture at the global scale and the detailed features of furniture at the local scale, thereby ensuring that the generated layout is both globally reasonable and accurate in details.

[0024] Step 2: Use the deep learning enhanced simulated annealing algorithm (DL-SA) to perform multi-constraint optimization on the initial layout obtained in Step 1.

[0025] In this embodiment, the DL-SA algorithm first extracts deep features based on a three-module neural network architecture. This architecture includes: (1) a feature extractor (CondorFeatureExtractor), which uses an improved multi-layer perceptron to map the physical attributes of furniture to a high-dimensional feature space and introduces a type embedding mechanism to assign unique embedding vectors to different types of furniture; (2) a scene transformer (SceneTransformer), which is based on a 4-head self-attention mechanism and 2-layer Transformer encoders to process the global spatial relationships between furniture and can understand the complete layout context; (3) an optimization predictor (OptimizationPredictor), which converts the global scene features into optimization guidance signals to guide the search direction of the simulated annealing algorithm.

[0026] The DL-SA algorithm adopts an exponentially decaying temperature scheduling strategy during the optimization process. The initial temperature is 10.0, and the cooling coefficient is 0.95. In the high-temperature stage, the algorithm tends to perform global exploration and can accept a certain degree of layout deterioration; as the temperature gradually decreases, the algorithm gradually turns to local refinement search and finally converges to a high-quality solution. In the state generation strategy, the position perturbation is associated with the temperature parameter, and the angle adjustment uses an orthogonal angle set {0°, 90°, 180°, 270°}, which not only ensures the reachability of the search space but also improves the regularity of the layout.

[0027] During the optimization process, a comprehensive loss function was designed to balance different optimization objectives. This loss function integrates multiple components: the reconstruction loss is used to measure the difference between the generated layout and the target layout, ensuring that the generated layout is generally consistent with the design objectives; the physical constraint loss ensures that the layout meets basic physical rules; the functional relationship loss guarantees the usage relationships between related furniture; and the aesthetic constraint loss enhances the visual effect and spatial balance of the layout.

[0028] In this embodiment, the constraints considered in the optimization process of the DL-SA algorithm can be divided into three categories: physical constraints, functional relationship constraints, and aesthetic constraints. Physical constraints include boundary constraints and overlap constraints, functional relationship constraints include distance constraints, and aesthetic constraints include center constraints. Specifically, boundary constraints ensure that furniture is within the room range, preventing furniture from exceeding the room boundaries; overlap constraints penalize the overlap of furniture by calculating the intersection over union (IoU) between furniture to avoid furniture penetration; distance constraints ensure a reasonable spacing between functionally related furniture groups by calculating and controlling the minimum distance between furniture; and center constraints encourage furniture to be distributed in the central area of the room, avoiding over-concentration or dispersion of furniture to optimize the overall spatial distribution of furniture.

[0029] A series of physical constraints are applied to ensure that the generated layout conforms to actual physical limitations. The rotation angle constraint ensures a reasonable orientation of furniture by restricting the rotation angle of furniture between 0 and 2π. The size constraint ensures that the size of furniture is appropriate by restricting the size of furniture between preset minimum and maximum values. The floor plane constraint ensures that furniture does not float in the air by fixing the position of furniture on the floor plane.

[0030] In this embodiment, the overlap constraint is implemented through a differentiable physics engine, including transforming the collision detection process into a differentiable tensor operation. The specific implementation uses the Separating Axis Theorem (SAT), which is particularly efficient for collision detection of rotated rectangles. For any two rotated rectangles, the system checks all possible separating axes and calculates the projections of the rectangles on these axes. If there exists an axis such that the projections of the two rectangles on this axis do not overlap, it is determined that there is no collision; otherwise, the system calculates the penetration depth and optimizes the penetration penalty through backpropagation.

[0031] In this embodiment, the DL-SA algorithm adopts a dynamic constraint adjustment mechanism to adaptively adjust the weights of various constraints according to the optimization stage and the degree of constraint violation. At different stages of optimization, the constraint weights are automatically adjusted: the physical constraint weight is the highest in the initial stage, the functional relationship constraint weight increases in the middle stage, and the aesthetic constraint weight increases in the later stage. In addition, the system dynamically adjusts the weights according to the degree of constraint violation, enabling the algorithm to focus on solving the most serious current constraint problem.

[0032] During the training process of the DL-SA algorithm, a strategy based on adversarial learning is incorporated. By introducing an adversarial network to simulate potential layout interference factors, the algorithm's robustness to complex scenarios is enhanced. The adversarial network generates interference layouts, forcing the DL-SA algorithm to maintain a high optimization performance even in the face of adverse conditions. This adversarial training mechanism significantly enhances the algorithm's generalization ability and robustness.

[0033] Step 3: Based on the layout obtained in Step 2, retrieve the corresponding 3D objects from the dataset. Specifically, select the object with the closest size in the corresponding category and then place it according to the layout to construct a 3D indoor scene.

[0034] Specifically, in Step 3, first comprehensively analyze the layout obtained in Step 2 to clarify key information such as the category, size, spatial position, and orientation of each object. Based on this information, locate the set of objects in the corresponding category in the dataset. For each object in the set, carefully calculate the difference between its size and the predicted size in the layout, and select the object with the closest size. Finally, strictly place the selected object in the 3D space according to the position and orientation information given in the layout. Through such a series of operations, gradually construct a complete 3D indoor scene that conforms to the layout plan.

[0035] In this embodiment, the finally generated is a 3D indoor scene model, which can be edited and displayed in various design software to meet the needs of different users.

[0036] Such as Figure 4 shown, the indoor scene layout rendering effect diagram finally generated by the method of the present invention. The figure shows a typical bedroom layout, including furniture such as a bed, bedside table, and wardrobe. The positions of the furniture are reasonable, the spacing is appropriate, and there is no overlap, reflecting the effect of multi-constraint optimization of the present invention. It can be seen from the figure that the spatial relationship between the furniture is clear, meeting the actual use requirements, verifying the practicability and effectiveness of the method of the present invention.

[0037] The preferred embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. The embodiments described in the present invention are only descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. Among the various specific technical features described in the above specific embodiments, they can be combined in any suitable manner without contradiction. As long as such a combination does not violate the idea of the present invention, it should also be regarded as the content disclosed in the present disclosure. To avoid unnecessary repetition, the present invention does not further explain various possible combination methods.

[0038] The present invention is not limited to the specific details in the above embodiments. Without departing from the technical concept of the present invention and within the scope not departing from the design concept of the present invention, various modifications and improvements made by those skilled in the art to the technical solution of the present invention shall fall within the protection scope of the present invention. The technical content claimed by the present invention has been fully recorded in the claims.

Claims

1. An automatic indoor layout generation method based on deep neural network and simulated annealing optimization, characterized in that It includes the following steps: Step 1: Use a variational autoencoder (VAE). Take the random noise sampled from the standard Gaussian distribution as the input of the VAE. Generate indoor scene layout information through the VAE. The indoor scene layout information includes furniture layout information, and the furniture layout information includes the initial position, rotation angle, size, and category information of the furniture. Map the furniture layout information to the latent space through the encoder of the VAE, and then generate the initial indoor scene layout containing the complete furniture layout information through the decoder of the VAE; Step 2: Use a deep learning enhanced simulated annealing algorithm (DL-SA) to perform multi-constraint optimization on the initial layout obtained in Step 1. The process of the DL-SA algorithm is as follows: a) Extract deep features based on a three-module neural network architecture. The three-module neural network architecture includes: a feature extractor that maps furniture attributes to a high-dimensional feature space; a scene transformer that processes the global spatial relationship between furniture through a multi-head self-attention mechanism; an optimization predictor that generates an optimization guidance signal; b) Under the control of the temperature parameter, generate candidate layouts according to the state generation strategy guided by the deep features, realizing the organic integration of the global exploration ability of simulated annealing and the semantic understanding ability of the neural network; c) Apply a multi-level constraint processing mechanism, including physical constraints, functional relationship constraints, and aesthetic constraints. Among them, the physical constraints ensure the basic feasibility of the layout, the functional relationship constraints ensure that the layout meets the actual usage requirements, and the aesthetic constraints improve the visual effect and spatial balance of the layout; d) Implement efficient rotating rectangle collision detection through the separating axis theorem, accurately quantify the overlapping problem between furniture, and guide the collision repair process; e) Adopt a dynamic weight adjustment mechanism and a hierarchical priority strategy to adaptively balance various constraints at different optimization stages, ensuring that important constraints are preferentially satisfied; Thus, the optimized layout is obtained; Step 3: Based on the optimized layout obtained in Step 2, retrieve the corresponding 3D objects from the dataset, select the objects with the closest size in the corresponding category, and then place them according to the layout to construct a 3D indoor scene.

2. The automatic indoor layout generation method based on deep neural network and simulated annealing optimization according to claim 1, wherein In Step 1, the variational autoencoder (VAE) uses a multi-layer Transformer network to capture the complex spatial relationship between furniture and optimizes the representation ability of the latent space through an adaptive learning mechanism.

3. An automatic indoor layout generation method based on deep neural network and simulated annealing optimization according to claim 2, characterized in that, Denote the multi-layer Transformer network in the VAE as SceneEncoder. SceneEncoder introduces a loss function based on multi-scale feature fusion during training.

4. The automatic indoor layout generation method based on deep neural network and simulated annealing optimization according to claim 1, wherein The dynamic weight adjustment mechanism in Step 2 is implemented in three stages according to the optimization process. In the initial stage, it focuses on the physical constraint weight. In the middle stage, it increases the functional relationship constraint weight. In the later stage, it increases the aesthetic constraint weight to achieve a progressive optimization strategy.

5. The automatic indoor layout generation method based on deep neural network and simulated annealing optimization according to claim 1, characterized in that In Step 2, the deep learning enhanced simulated annealing algorithm (DL-SA) combines a strategy based on adversarial learning during training and simulates potential layout interference factors by introducing an adversarial network.

6. A method for automatically generating an indoor layout based on a deep neural network and simulated annealing optimization according to any one of claims 1-5, characterized in that The physical constraints in Step 2 specifically include boundary constraints and overlap constraints; the functional relationship constraints specifically include distance constraints; the aesthetic constraints specifically include center constraints, where the boundary constraints ensure that the furniture is within the room; the distance constraints control the minimum distance between furniture; the overlap constraints avoid furniture penetration by calculating the intersection over union; and the center constraints optimize the overall spatial distribution of the furniture.

7. The automatic indoor layout generation method based on deep neural network and simulated annealing optimization according to claim 6, wherein The overlap constraints are implemented through a differentiable physics engine, including converting the collision detection process into differentiable tensor operations, using axis-aligned bounding boxes and the separating axis theorem for collision detection, and optimizing the penetration penalty through backpropagation.

Images