Implicit neural representation generation method based on attention mechanism and super network

Abstract

The invention discloses an implicit neural representation generation method based on an attention mechanism and a super network. The method comprises the following steps: firstly, converting a parameter matrix and a hidden state of a target INR into a unified sequence representation through a row-level Token mechanism; secondly, constructing a super network based on Transform, respectively simulating a forward propagation process of a target network and a back propagation process based on a chain rule by using a cross attention mechanism, and explicitly capturing an element-level dependency relationship in a Jacobian matrix through attention weight so as to accurately estimate a parameter update gradient; and finally, in combination with an adaptive parameter initialization strategy and a Token initialization network, the distribution stability of the generated parameters is ensured. According to the method, large-scale 2D images such as CelebA-HQ and ImageNet and complex 3D radiation field data can be efficiently processed, and the convergence speed, the reconstruction precision and the generalization ability of generated parameters are remarkably improved while high expandability is kept.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, deep learning, and computer graphics, and in particular to a method for generating implicit neural representations (INRs) parameters using hypernetworks, applicable to tasks such as image reconstruction, 3D shape generation, and neural radiation field (NeRF) view synthesis. Background Technology

[0002] Implicit neural representations (INRs), as an emerging paradigm of signal representation, use neural networks (typically multilayer perceptrons, MLPs) to represent continuous input coordinates (such as...). or INRs are directly mapped to output signal values ​​(such as RGB color, density, or symbol distance). Compared with traditional discrete representations (such as pixel grids, voxels, and point clouds), INRs have advantages such as resolution independence, compact storage, and strong topology adaptability, and have been widely used in fields such as image reconstruction, 3D shape modeling, and neural radiation fields (NeRF).

[0003] To enable the INR to adapt to different data instances (such as different faces or different 3D objects), researchers introduced the technique of hypernetworks. A hypernetwork is a type of meta-neural network that receives conditional input about the target instance (such as image features or task ID) and outputs weight parameters of the target INR.

[0004] However, existing INR generation methods based on hypernetworks have two significant technical bottlenecks: 1. Lack of Inter-Layer Dependency: Many existing methods (such as HyperDiffusion) tend to generate the weights of each layer of the target network independently, or simply connect them through fully connected layers. This approach ignores the essential characteristic of the target neural network as a composite function, namely, the close gradient flow dependencies between layers. Failure to simulate this dependency makes it difficult for the generated parameters to work together, thereby reducing the quality of high-frequency details in the reconstruction.

[0005] 2. Limited Scalability: To simulate inter-layer dependencies, some works (such as HyPoGen) attempt to use MLPs to predict the Jacobian matrix. However, the dimension of the Jacobian matrix increases quadratically with the network width, making it difficult for such methods to scale to wide networks or high-dimensional signals. Other Transformer-based methods (such as LDMI), while improving scalability, mainly operate in the latent space and fail to explicitly model the optimal gradient structure of the target network, resulting in insufficient performance on complex 3D tasks.

[0006] Therefore, there is an urgent need for an implicit neural representation generation method that can both explicitly capture the gradient flow dependency between layers of the target network and has high efficiency and scalability. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for generating implicit neural representations based on attention mechanisms and hypernetworks. This invention solves the scalability problem by combining a Transformer architecture with a row-level tokenization mechanism, and resolves the inter-layer dependency problem by explicitly simulating the chain rule of backpropagation through a cross-attention mechanism.

[0008] This invention discloses an implicit neural representation generation method based on attention mechanisms and hypernetworks. Its core process begins with constructing a training dataset containing 2D images or 3D multi-view data and extracting task embedding vectors. Subsequently, a hypernetwork based on a Transformer architecture is constructed, and row-level tokenization is used to unify the parameter matrix and hidden states of the target implicit neural representation (INR) into a sequence representation. In the parameter generation stage, the hypernetwork explicitly simulates the forward activation propagation of the target network through a cross-attention mechanism and uses a modality-consistent attention query strategy to approximate the multiplication of the Jacobian matrix and the chain rule matrix, thereby accurately capturing inter-layer gradient dependencies in the backpropagation simulation. Simultaneously, an adaptive parameter initialization strategy is combined to dynamically adjust the variance distribution of the output weights to adapt to the initialization requirements of the target network (such as SIREN). Finally, the generated weights are loaded into the target INR to construct a neural renderer, which renders the input coordinates to obtain the predicted image. The hypernetwork parameters are optimized end-to-end by minimizing the reconstruction loss function between the rendered result and the real data.

[0009] The specific technical solution adopted in this invention is as follows: This invention provides a method for generating implicit neural representations based on attention mechanisms and hypernetworks, comprising the following steps: S1: Obtain the dataset of the signal to be reconstructed, which includes 2D images, 3D shape-occupied grids or neural radiation field data, and extract the task embedding vector from the signal to be reconstructed. S2: Construct a Transformer-based super network, perform row-level tokenization on the parameter matrix and hidden states of the target INR network, and convert them into a sequence format with a unified dimension; S3: Construct a Token initialization network to generate the initial parameter Token sequence of the target implicit neural representation INR network through task embedding vectors; S4: Execute the iterative optimization loop of the supernetwork, which includes two phases: forward propagation simulation and back propagation simulation; S5: In the forward propagation simulation phase, the cross-attention mechanism is used to simulate the forward calculation process of each layer of the target INR, and to establish the dependency relationship between the hidden state of the current layer and the hidden state of the previous layer and the parameters of the current layer. S6: In the backpropagation simulation stage, the gradient backpropagation process of the target INR is simulated by using the cross-attention mechanism. The Jacobian matrix is ​​approximated by the attention weight and the chain rule matrix multiplication is used to estimate the update gradient of the parameters of each layer of the target INR. S7: Iteratively update the parameter Token of the target INR based on the estimated update gradient, and adjust the variance distribution of the output parameters using an adaptive parameter initialization strategy; S8: Decode the optimized parameter Token into a weight matrix of the target INR, use the target INR to fit and render the signal to be reconstructed, and train the supernetwork by minimizing the reconstruction loss function.

[0010] Furthermore, the row-level tokenization specifically includes: For the target INR Layer parameter matrix each of its rows Treated as an independent unit, it is mapped to a sequence of parameter tokens through a multilayer perceptron encoder. ; For the target INR Hidden state vector of layer Each activated neuron Treated as an independent unit, mapped as a sequence of state tokens. ; The parameter token sequence and the state token sequence are processed in a unified dimension within the Transformer hypernetwork.

[0011] Furthermore, the token initialization network is a lightweight Transformer structure, for each layer of the target INR. Initialize a set of learnable parameter query vectors. task embedding vector As keys and values, layer-specific initial parameters (Tokens) are generated through a layer-independent cross-attention module. .

[0012] Furthermore, the forward propagation simulation phase utilizes a cross-attention mechanism to simulate forward computation of the target neuron. The specific calculation formula is as follows: ; Among them, the parameter Token of the current layer As query Q, the hidden state Token of the previous layer. Simultaneously serving as key K and value V, the matrix multiplication operation of the linear layer is simulated through attention-weighted aggregation.

[0013] Furthermore, the backpropagation simulation stage utilizes a cross-attention mechanism to approximate the Jacobian matrix, specifically adhering to the modality consistency principle: Calculate the Jacobian matrix of the parameters with respect to activation. : Layer activation state For querying, parameters For keys and values, i.e. ; Calculate the Jacobian matrix of inter-layer activation : with parameter Token For querying, activation status For keys and values, that is: ; Gradient aggregation using the chain rule is performed through cross-attention: the gradient back from the upstream is used as the query, and the Jacobian matrix mentioned above is used as the key and value to calculate the parameter gradient update amount of the current layer. Its formula is: ; ; Update the parameters using the calculated values For the parameter token Update: ,in This is a learnable step size parameter.

[0014] Furthermore, the adaptive parameter initialization strategy specifically includes: Calculate the target variance of the target INR at each level. For the target network of the SIREN architecture, it is set as ,in The number of input neurons; Calculate the actual statistical variance of the output weights of the supernetwork. ; Calculate the dynamic scaling factor ; Apply the dynamic scaling factor to the weights of the hypernetwork output: .

[0015] Furthermore, the number of iterations in the optimization loop The training iterations are set to 3; the hypernetwork is trained using the Adam optimizer with a learning rate set to [value missing]. And it does not use a learning rate decay strategy.

[0016] Furthermore, the reconstruction loss function is determined based on the type of signal to be reconstructed: For 2D image reconstruction tasks, the loss function To predict the L1 distance between the pixel value and the actual pixel value KL divergence loss Discriminator loss The combination of is represented as ; For the novel perspective synthesis task of Neural Radiation Field (NeRF), the loss function Including weight space loss and image spatial loss , represented as ,in This is the L1 distance between the predicted weights and the pre-trained NeRF weights.

[0017] This invention is not only embodied in the above-described method, but can also be embodied in the following forms: Electronic device: includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the method described in any of the above embodiments.

[0018] Computer-readable storage medium: storing computer instructions that, when executed by a processor, implement the methods described in any of the above embodiments.

[0019] The method provided by this invention has the following beneficial effects: 1. Superior rendering effect: By explicitly introducing the chain rule of backpropagation into the supernetwork for modeling, the generated INR parameters are more in line with the optimized gradient flow of the neural network, thus capturing high-frequency details of images or 3D scenes more accurately. The PSNR and SSIM indicators are superior to existing technologies.

[0020] 2. No additional training required (Inference-only): Thanks to large-scale pre-training, when applied to new unseen samples, this invention only needs a single forward inference to generate the corresponding INR weights, without the need for time-consuming overfitting training (Per-scene optimization) for each new sample.

[0021] 3. Supports infinite scaling: Based on the characteristics of implicit neural representation, the generated images or 3D models are continuous function representations, supporting sampling and rendering at arbitrary resolutions, and have the characteristic of "infinite resolution", which solves the resolution limitation problem of traditional discrete representations (such as voxels and pixels).

[0022] 4. High scalability: By adopting row-level tokenization and Transformer architecture, the high computational cost of fully connected Jacobian matrices is avoided, enabling the method to effectively handle large-scale datasets (such as ImageNet) and complex network structures. Attached Figure Description

[0023] Figure 1 This is a simplified flowchart of an implicit neural representation generation method based on attention mechanism and hypernetwork according to the present invention. Figure 2 The qualitative results of the implementation cases on the CelebHQ, ImageNet, ERA5, and ShapNet Chairs datasets are shown in the figure. Figure 3 This is a qualitative result graph of the implementation case on the NeRF dataset; Figure 4 This is a flowchart of one training cycle of the present invention. Detailed Implementation

[0024] See Figure 1 and Figure 4 This invention proposes a hypernetwork framework that models the parameter generation of Implicit Neural Representation (INR) as an optimization process. By explicitly simulating the forward propagation of the target network and the backpropagation based on the chain rule, this method can generate high-performance INR parameters. The technical solution of this invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Step 1: Dataset Preparation and Task Embedding Extraction This embodiment validates the effectiveness of the method on datasets with multiple modalities, including 2D image reconstruction, 3D shape reconstruction, and 3D neural radiation field (NeRF) synthesis.

[0026] 1.1 2D Image Data Processing: The CelebA-HQ dataset contains 30,000 high-resolution face images (28,000 for training and 2,000 for testing). Images are preprocessed and downsampled to 64x64 resolution. A pre-trained ResNet-18 encoder is used to extract the global feature vectors of the images for task embedding, denoted as . The dimension is set to 128.

[0027] ImageNet dataset: A subset of ILSVRC 2012, containing 1,281,167 training images. Image resolution was adjusted to 256x256. A pre-trained VQ-VAE model was used to extract latent features, and the feature maps were flattened into sequences as task conditions. The dimension is set to 256.

[0028] ERA5 climate data: Temperature field data with a resolution of 46x90, processed in the same way as 2D images.

[0029] 1.2 3D Data Processing: ShapeNet Chairs: Contains 6778 3D chair models, converted into 32x32x32 occupancy grids.

[0030] The pre-trained NeRF dataset contains 103,000 objects, and renders images from 36 different viewpoints for each object.

[0031] Input settings: In each training iteration, for a certain object, 6 rendered images from 36 images from different perspectives and their corresponding camera pose parameters (intrinsic and extrinsic parameter matrices) are randomly selected as input conditions.

[0032] Supervision settings: Six additional unseen viewpoints are selected as ground truth values ​​to calculate the image reconstruction loss synthesized from the new viewpoints.

[0033] Task embedding: Features of the input multi-view images are extracted using the DINO (ViT-S / 8) encoder and aggregated to obtain a task embedding vector with a dimension of 768. .

[0034] Step 2: Row-wise tokenization of the target INR To enable the Transformer architecture to efficiently process the parameters of the target INR (such as SIREN or MLP), this embodiment constructs a Transformer-based hypernetwork, denoted as […]. ,in This represents the learnable weight parameters of the hypernetwork itself. The goal of the hypernetwork is to generate the parameters of the target INR based on the input conditions. This embodiment employs a row-level tokenization strategy.

[0035] 2.1 Parameter Tokenization: Assuming the target INR's first The layer weight matrix is Each of its rows This is treated as an independent unit. Through a shared linear projection layer (MLP Encoder), this unit is mapped to a parameter Token. ,in , The hidden layer dimension of the hypernetwork (its value is...) (The dimensions are the same). That is, the entire first... The parameters of the layer are represented as sequences. .

[0036] 2.2 State Tokenization: Correspondingly, the hidden state vector output by this layer Each element is mapped to a state token. That is, the hidden state vector of the i-th layer is represented as: This row-level segmentation method preserves the neuron-level connection structure while avoiding the high cost of calculating the fully connected Jacobian matrix. This ensures that the token sequence length is only linearly related to the network width, greatly improving the scalability of the method.

[0037] Step 3: Token Initialization and Network Construction To accelerate convergence and provide a good starting point for optimization, this invention does not use random noise initialization, but instead constructs a token initialization network. .

[0038] Structural Design: The network consists of 4 Transformer layers with a hidden layer dimension of d.

[0039] Generation process: For each layer of the target network Define a set of learnable query vectors. Embedded by task As key and value, As a query, the initial parameter Token is generated through the Cross-Attention module: This step directly predicts the initial estimate of the target INR parameter. Compared to optimizing from scratch, this significantly reduces the number of iterations required.

[0040] Step 4: Iterative generation of a gradient-based loop (Hypernetwork Optimization Loop) The inference process of the hypernetwork is designed as an iterative loop simulating the optimization process of a neural network. In this embodiment, the number of loops is set to... .

[0041] In the In this iteration, the following sub-steps are executed sequentially: 4.1 Forward Propagation Simulation: Utilizing cross attention Simulate target neuron The weighted summation and activation process. For the target network's... Layer: Using the parameter Token of the current layer As a query, the activation status token of the previous layer. As keys and values.

[0042] The calculation formula is: This step is performed implicitly using Attention weights. Linear operations were performed to obtain the simulated activation state of the current layer.

[0043] 4.2 Backpropagation Simulation and Jacobi Approximation: The core principle of approximating the Jacobian matrix for gradient descent using the cross-attention mechanism is to follow the "modal consistency" principle.

[0044] Calculate the parameter gradient dependency (the Jacobian matrix of parameters on activation). : Layer input state For querying, parameters For keys and values.

[0045] Calculate the inter-layer gradient dependency (the Jacobian matrix of inter-layer activation). : with parameters For querying, the layer inputs the status. For keys and values.

[0046] Gradient aggregation: converting the gradient signal of the output layer into a single gradient. As the initial query, it is passed forward layer by layer through the Attention module mentioned above, and the update gradient of the parameters of each layer is calculated by combining the chain rule. .

[0047] 4.3 Parameter Update: Simulate gradient descent update rule, using the calculated gradient Update parameter Token: in This is a learnable step size parameter. After the T loops have completed, the inverse tokenization process will be executed. Reconvert to parameters For detailed procedures, please refer to step 5.

[0048] Step 5: Adaptive Parameter Initialization Strategy In the final token sequence Before decoding into a weight matrix, the mismatch between the supernetwork output variance and the target INR (especially for SIREN networks) initialization requirements must be addressed.

[0049] 5.1 Calculate the objective variance: The activation function is determined based on the type of activation function used in the target network. For INR(SIREN) using a sinusoidal activation function, the... Target variance of layer Set as: in This represents the number of input neurons in that layer.

[0050] 5.2 Apply adaptive scaling: The actual variance of the parameter Token output by the statistical supernetwork .

[0051] Calculate the dynamic scaling factor : Multiply the decoded weights by this factor: .

[0052] This strategy ensures that the generated INR weight distribution is in a linear region where network parameters can be easily optimized, effectively avoiding gradient vanishing or saturation problems in the early stages of training.

[0053] Step 6: Rendering and Loss Function Optimization The generated weights Load it into the target network architecture (e.g., 5-layer SIREN, 256 hidden layers) to build a neural renderer.

[0054] 6.1 Rendering Execution: For 2D tasks: target network Learning from pixel coordinates Mapping to color values. During rendering, the input is a normalized pixel coordinate grid, calculated using the following formula: in Normalized pixel coordinates This is the predicted pixel color value.

[0055] For 3D NeRF tasks: target network Receive 3D spatial position and observation direction Output volume density and color During rendering, the volumetric rendering formula is used along the light rays. Integrate points: in, The pixel color corresponding to the light. The number of sampling points. The distance between adjacent sampling points. Transmittance represents the rate at which light reaches the 6th digit of the sphere. The probability that a sample point is not occluded is calculated using the following formula: This formula simulates the cumulative color and opacity of light as it passes through the corresponding 3D field.

[0056] 6.2 Loss Calculation: Different combinations of loss functions are used to train the hypernetwork end-to-end, depending on the task type.

[0057] Image loss With KL divergence loss Discriminator loss Combinations: Image spatial loss ( ): Calculates the L1 distance between the rendered image and the ground truth image.

[0058] KL divergence loss This loss is a regularization constraint that measures the difference between the latent feature distribution and the standard normal distribution to ensure that the generated implicit representation parameters have good continuity and smoothness in the latent space.

[0059] Discriminator loss This loss introduces an adversarial training mechanism, which, by distinguishing the differences between generated and real images, prompts the supernetwork to generate reconstructions with higher frequency texture details and stronger perceptual realism. Its coefficient.

[0060] Weight space loss ( ): For NeRF tasks only, in order to introduce 3D geometric priors, calculate the L1 distance between the generated weights and the weights of the pre-trained Teacher NeRF model.

[0061] Total loss function: defined as .

[0062] 6.3 Backpropagation: Use the Adam optimizer to set The learning rate is End-to-end training was performed on an NVIDIA H800 GPU, updating all Transformer parameters and Query vectors in the supernetwork based on the total loss function.

[0063] Step 7: Experimental Verification and Result Analysis To verify the effectiveness and advancement of this invention in generating implicit neural representations, comparative experiments were conducted on multiple datasets with current mainstream methods (including HyPoGen, LDMI, VaMoH, etc.). Peak signal-to-noise ratio (PSNR, unit: dB) was uniformly used as the evaluation criterion; a higher PSNR value indicates better reconstruction quality.

[0064] The table below shows in detail the PSNR comparison results between the method of this invention and various baseline methods on different datasets: Table 1: Comparison of PSNR (dB) performance between the present invention and existing methods on different datasets Results analysis: 1. 2D Image Reconstruction: On both the CelebA-HQ and ImageNet datasets, the method of this invention achieved the highest PSNR values. Particularly on CelebA-HQ, this method improved by 11.5 dB compared to HyPoGen and by 2.9 dB compared to LDMI, demonstrating the importance of explicitly modeling interlayer gradient dependencies for capturing high-frequency details. Visualization results are shown in the appendix. Figure 2 The upper (ImageNet) section.

[0065] 2. Scientific data reconstruction: On ERA5 climate data, this method achieved an extremely high PSNR of 49.3dB, which is close to the level of lossless reconstruction and is superior to all baseline methods.

[0066] 3. 3D Tasks: In the ShapeNet chair reconstruction task, the PSNR of the rendered view generated by this method reaches 23.8dB, which is better than LDMI and HyPoGen. In the complex NeRF novel viewpoint synthesis task, this method achieves 27.8dB under joint supervision of weights and images, significantly better than LDMI's 26.5dB and HyPoGen's 22.0dB, proving that this method can generate 3D implicit representations with higher geometric consistency and rendering quality. See the appendix for visualization results. Figure 2 The following section (ShapeNet Chairs) is attached. Figure 3 (NeRF) section.

[0067] This invention is not only embodied in the above-described method, but can also be embodied in the following forms: Electronic device: includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the method described in any of the above embodiments.

[0068] Computer-readable storage medium: storing computer instructions that, when executed by a processor, implement the methods described in any of the above embodiments.

[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0070] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for generating implicit neural representations based on attention mechanisms and hypernetworks, characterized in that, Includes the following steps: S1: Obtain the dataset of the signal to be reconstructed, which includes 2D images, 3D shape-occupied grids or neural radiation field data, and extract the task embedding vector from the signal to be reconstructed. S2: Construct a Transformer-based super network, perform row-level tokenization on the parameter matrix and hidden states of the target INR network, and convert them into a sequence format with a unified dimension; S3: Construct a Token initialization network to generate the initial parameter Token sequence of the target implicit neural representation INR network through task embedding vectors; S4: Execute the iterative optimization loop of the supernetwork, which includes two phases: forward propagation simulation and back propagation simulation; S5: In the forward propagation simulation phase, the cross-attention mechanism is used to simulate the forward calculation process of each layer of the target INR, and to establish the dependency relationship between the hidden state of the current layer and the hidden state of the previous layer and the parameters of the current layer. S6: In the backpropagation simulation stage, the gradient backpropagation process of the target INR is simulated by using the cross-attention mechanism. The Jacobian matrix is ​​approximated by the attention weight and the chain rule matrix multiplication is used to estimate the update gradient of the parameters of each layer of the target INR. S7: Iteratively update the parameter Token of the target INR based on the estimated update gradient, and adjust the variance distribution of the output parameters using an adaptive parameter initialization strategy; S8: Decode the optimized parameter Token into a weight matrix of the target INR, use the target INR to fit and render the signal to be reconstructed, and train the supernetwork by minimizing the reconstruction loss function.

2. The method according to claim 1, characterized in that, The row-level tokenization specifically includes: For the target INR Layer parameter matrix each of its rows Treated as an independent unit, it is mapped to a sequence of parameter tokens through a multilayer perceptron encoder. ; For the target INR Hidden state vector of layer Each activated neuron Treated as an independent unit, mapped as a sequence of state tokens. ; The parameter token sequence and the state token sequence are processed in a unified dimension within the Transformer hypernetwork.

3. The method according to claim 1, characterized in that, The token initialization network is a lightweight Transformer structure, for each layer of the target INR. Initialize a set of learnable parameter query vectors. task embedding vector As keys and values, layer-specific initial parameters (Tokens) are generated through a layer-independent cross-attention module. .

4. The method according to claim 1, characterized in that, The forward propagation simulation phase utilizes a cross-attention mechanism to simulate forward computation of the target neuron. The specific calculation formula is as follows: ； Among them, the parameter Token of the current layer As query Q, the hidden state Token of the previous layer. Simultaneously serving as key K and value V, the matrix multiplication operation of the linear layer is simulated through attention-weighted aggregation.

5. The method according to claim 1, characterized in that, The backpropagation simulation phase utilizes a cross-attention mechanism to approximate the Jacobian matrix, specifically adhering to the modality consistency principle: Calculate the Jacobian matrix of the parameters with respect to activation. : Layer activation state For querying, parameters For keys and values, i.e. ； Calculate the Jacobian matrix of inter-layer activation : with parameter Token For querying, activation status For keys and values, that is: ； Gradient aggregation using the chain rule is performed through cross-attention: the gradient back from the upstream is used as the query, and the Jacobian matrix mentioned above is used as the key and value to calculate the parameter gradient update amount of the current layer. Its formula is: ； ； Update the parameters using the calculated values For the parameter token Update: ,in This is a learnable step size parameter.

6. The method according to claim 1, characterized in that, The adaptive parameter initialization strategy specifically includes: Calculate the target variance of the target INR at each level. For the target network of the SIREN architecture, it is set as ,in The number of input neurons; Calculate the actual statistical variance of the output weights of the supernetwork. ; Calculate the dynamic scaling factor ; Apply the dynamic scaling factor to the weights of the hypernetwork output: .

7. The method according to claim 1, characterized in that, The number of iterations for optimization The training iterations are set to 3; the hypernetwork is trained using the Adam optimizer with a learning rate set to [value missing]. And it does not use a learning rate decay strategy.

8. The method according to claim 1, characterized in that, The reconstruction loss function is determined based on the type of signal to be reconstructed: For 2D image reconstruction tasks, the loss function To predict the L1 distance between the pixel value and the actual pixel value KL divergence loss Discriminator loss The combination of is represented as ; For the novel perspective synthesis task of Neural Radiation Field (NeRF), the loss function Including weight space loss and image spatial loss , represented as ,in This is the L1 distance between the predicted weights and the pre-trained NeRF weights.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-8.

10. A storage medium comprising computer-executable instructions that, when executed by a computer processor, implement the method as claimed in any one of claims 1-8.

Images