Wavelet-based frequency domain perception and enhanced WE-GPMConv-StyleGAN lightweight image generation method

Abstract

The invention belongs to the field of computer vision, and relates to an efficient image synthesis technology of a generative adversarial network (GAN). Aiming at the problems of insufficient high-frequency details, limited geometric denaturation and the like and low calculation efficiency in high-resolution image synthesis of the existing generation method, the invention provides a frequency domain sensing generation method based on wavelet driving WE-GPMConv-StyleGAN. According to the model, discrete wavelet transform is embedded in a StyleGAN3 architecture, a traditional convolution operation is decomposed into a two-stage process of wavelet domain feature extraction and cross-band fusion, and calculation complexity is reduced through grouping separable convolution; designing a low-frequency enhancement module to expand the global structure modeling capability, and expanding a receptive field by using secondary wavelet decomposition; a high-frequency enhancement module is introduced to enhance texture details through a residual structure and cross-channel interaction; in combination with a staged training strategy, low-frequency global features are optimized preferentially in the early stage, and local details are refined by combining a high-frequency module in the later stage. According to the invention, through a wavelet domain multiband decoupling and fusion mechanism, the high-frequency generation quality and geometric denaturation are significantly improved, and a frequency domain driving solution with both efficiency and precision is provided for high-resolution image generation.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, specifically to the field of image generation, and more particularly to a WE-GPMConv-StyleGAN lightweight image generation method based on wavelet-based frequency domain perception and enhancement. Background Art

[0002] In recent years, generative adversarial networks (GANs) have made breakthroughs in image synthesis. The StyleGAN family of models, in particular, significantly improves the realism and diversity of high-resolution images through a progressive generative architecture and frequency-domain feature mapping. However, as resolutions increase to 4K, existing generative methods face three core challenges: insufficient generation of high-frequency details, limited geometric variability, and excessive computational resource consumption, severely hindering their practical application.

[0003] When it comes to generating high-frequency details, the local receptive field and fixed weight sharing mechanism of traditional convolutional neural networks cause high-frequency information (such as hair and skin texture) to gradually attenuate during layer-by-layer transmission. Although StyleGAN3 enhances frequency-domain perception through Fourier feature mapping, its pixel-domain-based generation approach still struggles to avoid high-frequency aliasing, particularly under geometric transformations such as rotation and translation, which can lead to blurred edges and texture distortion. Furthermore, the sensitivity of deep networks to high-frequency noise further exacerbates the instability of detail generation. Existing research has shown that content generation in the wavelet domain can produce higher-quality images and present more realistic high-frequency content, with significant advantages in detail representation and high-frequency information recovery. The wavelet transform, with its multi-resolution analysis capabilities, can decompose images into low-frequency subbands (LL) and high-frequency subbands (LH, HL, HH), providing a natural framework for frequency division optimization.

[0004] From the perspective of computational efficiency, high-resolution generation is usually accompanied by an exponential increase in the number of parameters and computational complexity. For example, StyleGAN3 requires more than 70M parameters at high resolution, resulting in training time and video memory usage far exceeding the hardware's carrying capacity. Although lightweight techniques (such as model pruning and knowledge distillation) can compress the number of parameters, they often come at the expense of generation quality. Existing solutions (such as MobileStyleGAN) attempt to reduce the amount of computation through grouped convolution, but the lack of channel interaction caused by the grouping operation severely weakens the feature expression capability, especially in the frequency domain information fusion scenario, where performance is significantly degraded.

[0005] WE-GPMConv-StyleGAN is a new image generation model based on StyleGAN3. The new model was evaluated on the FFHQ and MetFaces datasets at the same resolution. The model significantly reduces the computational cost of the model without reducing the quality of the generated images or the model's equivariance. Summary of the Invention

[0006] The purpose of the present invention is to provide an image generation model with low computational cost, higher model variability and more stable training to solve some problems existing in the prior art.

[0007] To achieve the above objectives, the present invention adopts the image generation technology model of WE-GPMConv-StyleGAN, and the scheme includes the following steps: Step S1: Prepare datasets of different resolutions. Downscale the 1024×1024 resolution dataset to obtain a 256×256 resolution dataset. The two datasets are used to train the model and verify the applicability of the model. Step S2: Latent vector mapping and initial feature generation: the input latent vector is converted into a style vector through the mapping network, and the initial feature map is generated through Fourier feature mapping; Step S3: The feature map is input into the wavelet-based GP-modulation convolution module, and modulated and demodulated with the previously obtained style vector to achieve the hierarchical control of different layers unique to StyleGAN; Step S4: Progressive training, starting from a low resolution (e.g., 4×4) and gradually increasing the resolution to improve training stability. The feature maps output by each layer are identified by the discriminator, and the loss function of the discriminator is improved using quantifiable conditions to constrain the image generation of the generator; Step S5: The training adopts a phased frequency domain enhancement strategy, designs a low-frequency enhancement module and a high-frequency enhancement module, and optimizes the low frequency first in the early stage of training, optimizes the high and low frequencies jointly in the middle stage of training, and finally optimizes the high frequency details; Step S6: Finally, evaluate the trained image generation model by generating 50K images and using the entire training set as the reference distribution to calculate the FID score. This also evaluates the equivariance of the model and the computational cost of the training process.

[0008] In the above scheme, the step S3 of constructing the GP-modulation convolution module is based on traditional modulation convolution. The traditional modulation convolution module modulates and demodulates the weight of ordinary 3×3 convolution. The GP-modulation convolution module is based on the weight modulation and demodulation of grouped separation convolution, wherein the grouped separation convolution decomposes the ordinary 3×3 convolution into a cascade structure of grouped convolution and point-by-point convolution. The wavelet-based GP-modulation convolution module is a design for the second layer of point-by-point convolution of the modulation convolution. The feature map after grouped modulation convolution is decomposed into four frequency bands (LL, LH, HL, HH) through discrete wavelet transform (DWT), and feature extraction and integration are performed on each sub-band using point-by-point convolution. Finally, the inverse wavelet transform (IWT) reintegrates the multi-scale frequency information and restores it to a high-resolution image feature map.

[0009] In the above scheme, the number of layers of the generated network is reduced from 14 to 7 in the progressive training of step S4, while retaining the resolution growth strategy and non-uniform channel strategy of the original model. Quantifiable conditions include directional optimality, separability, and injectability. Now it is guaranteed that the discriminator can find the most obvious difference direction, compress it to a one-dimensional measurable space, and ensure that the difference direction is consistent after projection, and the feature extraction function is guaranteed to be an injective function. The specific operation is to normalize the linear layer to force the weight vector to be projected onto the unit hypersphere, making the distribution difference metric more stable and alleviating the training instability of the traditional discriminator caused by overfitting of local features. Finally, the optimization goal of the entire network becomes to add the directional optimal term formula to the loss of the original discriminator: The loss function of the original StyleGAN3 discriminator loss generator remains unchanged, <, > represents the inner product operation, h is the feature extraction function, which maps the input x to the high-dimensional feature space. ω is the optimal direction selected in the feature space, λ is the hyperparameter that balances the adversarial loss and the direction optimality (usually set to 1), μ0 is the real data distribution, μ θ is the generator distribution.

[0010] In the above-mentioned solution, the frequency domain enhancement module of the network generated in step S5 divides the four subbands after DWT decomposition into the LL low-frequency component and the three high-frequency components LH, HL, and HH, respectively, enhancing them. The low-frequency enhancement module uses secondary wavelet decomposition to expand the receptive field and optimize global structural consistency. The high-frequency enhancement module suppresses noise interference through residual structure and cross-channel interaction, improving the quality of high-frequency details.

[0011] In the above scheme, the WE-GPMConv-StyleGAN model training experiment was conducted on a Linux platform, using Python version 3.8 and PyTorch version 1.11.0. The batch size was set to 2, 4, or 8, respectively, and the number of groups for group convolution was set to 8.

[0012] The computational cost of the most commonly used metrics for evaluating the quality of image generation models in step S6, such as FID and EQ-T and EQ-R, includes training time, graphics memory usage, and model parameter count. The training time is based on the training time per 10,000 iterations on a single graphics card with a batch size of 8.

[0013] In summary, compared with some existing target detection algorithms, the present invention significantly improves the ability to generate high-frequency details through wavelet domain frequency band decomposition and dynamic fusion mechanism, effectively solving the blurring and distortion problems of traditional generation models on fine structures such as textures and edges; combined with lightweight convolution design and frequency division optimization strategy, it greatly reduces computing resource consumption and achieves a balance between efficiency and quality of high-resolution image synthesis. The model significantly enhances the output stability under geometric transformation through collaborative training of low-frequency global perception and high-frequency residual enhancement, avoiding deformation and artifacts caused by operations such as translation and rotation. In addition, the phased optimization strategy prioritizes the integrity of low-frequency structures and refines high-frequency details in the later stage to improve the overall coordination and authenticity of the generated content. While reducing dependence on hardware resources, this solution provides an efficient and reliable frequency domain driven solution for high-fidelity image generation, which is suitable for industrial scenarios such as virtual reality and digital media that have strict requirements on real-time performance and visual quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a training flowchart of the WE-GPMConv-StyleGAN image generation model in the present invention.

[0015] Figure 2 This is the network architecture diagram of the WE-GPMConv-StyleGAN generator in this invention.

[0016] Figure 3 Schematic diagram of the improved process of the WE-GPMConv-StyleGAN discriminator in this invention.

[0017] Figure 4 This is the overall design diagram of the wavelet-based frequency domain enhancement module in the present invention.

[0018] Figure 5 Schematic diagram of the LFE module in the present invention.

[0019] Figure 6 Schematic diagram of the HFE module in the present invention. Specific implementation plan

[0020] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the exemplary embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0021] In this embodiment, combined with Figure 1 , provides a video target detection algorithm based on an improved CenterNet backbone network, the method comprising the following steps: Step S1: Prepare a dataset. The present invention collects open source dataset images including a 1024×1024 resolution MetFaces dataset and a 256×256 resolution MetFaces and FFHQ dataset, which are used to train the model and verify the applicability of the model respectively. Step S2: Latent vector mapping and initial feature generation: the input latent vector is converted into a style vector through the mapping network, and the initial feature map is generated through Fourier feature mapping. Step S2: Latent vector mapping and initial feature generation: the input latent vector is converted into a style vector through the mapping network, and the initial feature map is generated through Fourier feature mapping. Step S3: Input the feature map into the wavelet-based GP-modulation convolution module (such as Figure 2 Dark gray box), and performs modulation and demodulation operations with the previously obtained style vector to achieve the hierarchical control of different layers unique to StyleGAN. Step S4: Progressive training, starting from low resolution (such as 4×4) and gradually increasing the resolution to improve training stability. The feature map output by each layer is identified by the discriminator, such as Figure 4 The loss function of the discriminator is improved with quantifiable conditions to constrain the image generation of the generator. Step S5: The training adopts a phased frequency domain enhancement strategy to design a low-frequency enhancement module (such as Figure 5 ) and high frequency enhancement modules (such as Figure 6 ), and optimize low frequencies first in the early stage of training, optimize high and low frequencies jointly in the middle stage of training, and finally optimize high frequencies for details. Step S6: Finally, evaluate the trained image generation model by generating 50K images and using the entire training set as the reference distribution to calculate the FID score. This also evaluates the equivariance of the model and the computational cost of the training process.

[0022] Furthermore, in this embodiment, the step S3 of constructing the GP-modulation convolution module is based on traditional modulation convolution. The traditional modulation convolution module modulates and demodulates the weights of the ordinary 3×3 convolution. The GP-modulation convolution module is based on the weight modulation and demodulation of the grouped separable convolution, where the grouped separable convolution decomposes the ordinary 3×3 convolution into a cascade structure of grouped convolution and point-by-point convolution. This modulation is achieved by the following formula: w′=s·w gw w″=s′·w pw Where: w gw and w pwCorresponding to the original weights of 3x3 grouped convolution and 1x1 point-by-point convolution, w′ and w″ are their corresponding modulation weights, and s and s′ are the scales of the input feature map. The demodulation coefficient is the same as StyleGAN2 and can be calculated as: Where: i, j and k enumerate the input / output feature maps and spatial footprint of the convolution respectively. Substitute w′ and w″ into the above formula to obtain the corresponding demodulation coefficients demod′ and demod″. Finally, the demodulated output feature map x out , the formula is as follows: x′=w′·demod′·x x out =w″·demod″·x′

[0023] The wavelet-based GP-modulation convolution module is a design for the second layer of point-by-point convolution of the modulation convolution. The feature map after group modulation convolution is transformed by discrete wavelet transform (DWT), such as Figure 4 The image is decomposed into four frequency bands (LL, LH, HL, HH). Point-by-point convolution is used to extract and integrate features for each subband. Finally, the inverse wavelet transform (IWT) is used to reintegrate the multi-scale frequency information and restore the high-resolution image feature map.

[0024] In the above scheme, the step S4 progressive training is as follows Figure 2 As shown in the figure, the number of layers of the generative network is reduced from 14 to 7, while retaining the resolution growth strategy and non-uniform channel strategy of the original model. The quantifiable conditions include directional optimality, separability, and injectability. Now it is guaranteed that the discriminator can find the most obvious difference direction and compress it into a one-dimensional measurable space while ensuring that the difference direction is consistent after projection. The feature extraction function is guaranteed to be an injective function. Based on the above theoretical basis, the optimization of the discriminator structure can be divided into the following: Figure 3 As shown: First, the discriminator is decomposed into the inner product of the output of the feature extraction network h(x) and ω, that is, f(x) = <ω,h(x)>. Then, the last layer of the discriminator uses a normalized linear layer to force the weight vector ω to lie on the unit hypersphere (S D-1 ), ensuring the normalization of the projection direction. Finally, the optimization goal of the entire network becomes to add the direction optimal term to the loss of the original discriminator: The loss function of the original StyleGAN3 discriminator loss generator remains unchanged, <, > represents the inner product operation, h is the feature extraction function, which maps the input x to the high-dimensional feature space. ω is the optimal direction selected in the feature space, λ is the hyperparameter that balances the adversarial loss and the direction optimality (usually set to 1), μ0 is the real data distribution, μ θ is the generator distribution.

[0025] Furthermore, in this embodiment, the frequency domain enhancement module of the network generated in step S5 is to divide the four sub-bands after DWT decomposition into LL low-frequency components and LH, HL, HH three high-frequency components to enhance them respectively. Among them, the LFE module expands the receptive field through secondary wavelet decomposition and optimizes the global structural consistency. Figure 5 As shown, the LFE process decomposes the low-frequency component X from the feature map 1 LL performs a secondary decomposition to generate a coarser low-frequency X 2 LL and corresponding high frequency component X 2 LH, X 2 HL and X 2 HH, then in X 2 A 1×1 convolution operation is performed on the LL subband, and the same convolution operation is performed independently on all high-frequency subbands to further extract low-frequency details. The features are then reconstructed step by step through the inverse wavelet transform, and the secondary processed X 2 LL features and level X 1 The LL convolution results are added element by element to form the final covering original Figure 4 ×4 region. This process preserves the original high-frequency information through inter-layer residual connections, allowing the 1×1 convolution to expand the receptive field in the low-frequency domain after dimensionality reduction, while the number of parameters only increases linearly with the decomposition level. Through two-level wavelet decomposition, LFE gradually expands the receptive field in low-frequency subbands, while using 3×3 convolutions to efficiently extract multi-scale features at each subband level, significantly improving the global modeling capabilities of CNNs while avoiding the over-parameterization problem of traditional large convolution kernels.

[0026] According to the multi-level grading characteristics of Haar wavelet, the low-frequency components can be further decomposed and fused, and finally the feature map with the same resolution as the original image is output. In order to increase the computing resources too much, the present invention uses a two-level wavelet decomposition. The multi-level decomposition formula is as follows: Where: Z (i) represents the fusion result of the i-th level, Y (i) LL is the low-frequency convolution output, Y (i) LH and others are high-frequency convolution outputs.

[0027] The high-frequency enhancement module suppresses noise interference through residual structure and cross-channel interaction, improving the quality of high-frequency details. Figure 6As shown in Figure 1, this module concatenates the horizontal, vertical, and diagonal high-frequency subbands along the channel dimension to form a joint representation of multi-directional edge features. It also introduces a lightweight residual path consisting of two 1×1 convolutions and ReLU activations. The core of the residual structure lies in the skip connection, which allows the original high-frequency information to be transmitted losslessly to the output layer, avoiding the over-smoothing of details by deep nonlinear transformations. The two-stage 1×1 convolution dynamically fuses high-frequency content from different directions through cross-channel weight learning, suppressing random noise while amplifying effective high-frequency information.

[0028] In StyleGAN3, the relationship between high- and low-frequency signal processing and model equivariance is the core of the architecture design, which directly affects the stability and quality of the generated image during geometric transformations (such as translation and rotation). In image generation based on the StyleGAN architecture, high-frequency information in the early stages of training is not that important. At the same time, the role and impact of high- and low-frequency enhancement in different layers of the generative network are also different. For this reason, the model WE-GPMConv-StyleGAN of the present invention adopts a high- and low-frequency separation training strategy, that is, a low-frequency enhancement module (LFE) is added to the first layer. * ) for initial training, and then add a high-frequency enhancement module (HFE) to the last layer in the later stage. * )Joint optimization.

[0029] Furthermore, in the above scheme, the most commonly used evaluation index for evaluating the quality of the image generation model in step S6 is And evaluation model variability indicators: EQ-R=PSNR (W·A·V·C) (I r ,K r ) I t (w,x,p,c)=T x [G(z0;w)][p,c] I r (w,a,p,c)=R a [G(z0;w)][p,c] K t (w,x,p,c)=G(t x [z0];w)[p,c]) Where: w~W are random intermediate potential codes generated by the mapping network, x=(x0,x1)~X 2 is a random translation offset, p enumerates the pixel positions in the mutually valid region V, c~C are color channels, and z0 represents the Fourier feature of the input. Generate a reference image I t After generating an image for the original input, directly translate the output image. Generate the perturbation image K t is to generate an image for the translated input, Kr The random rotation angle a is taken from A=U(0°, 360°), and the operator r a Corresponding to the input Fourier features relative to the canvas [0,1] 2 Continuous rotation of the center. a High-quality rotation operations corresponding to Lanczos interpolation, Represents a pseudo-rotation operator, including frequency-domain filtering, to eliminate aliasing artifacts after rotation. Model computational overhead includes training time, graphics memory usage, and model parameter count. Training time is based on training time per 10,000 iterations on a single graphics card with a batch size of 8.

[0030] This paper uses a Linux platform to conduct training experiments on the WE-GPMConv-StyleGAN model, using Python version 3.8, PyTorch version 2.1.0+cu12, CUDA version 12.1, and CuDNN version 8.9.2. Training is performed on an NVIDIA GeForce RTX 3090 GPU (24GB). The batch size for training images with 256×256 resolution is set to 8, and the batch size for experiments on a 1024×1024 dataset is set to 2.

[0031] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will appreciate that the foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, and the like made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A WE-GPMConv-StyleGAN lightweight image generation method based on wavelet frequency domain perception and enhancement, the scheme includes the following steps: Step S1: Prepare datasets of different resolutions. Downscale the 1024×1024 resolution dataset to obtain a 256×256 resolution dataset. The two datasets are used to train the model and verify the applicability of the model. Step S2: Latent vector mapping and initial feature generation: the input latent vector is converted into a style vector through the mapping network, and the initial feature map is generated through Fourier feature mapping; Step S3: The feature map is input into the wavelet-based GP-modulation convolution module, and modulated and demodulated with the previously obtained style vector to achieve the hierarchical control of different layers unique to StyleGAN; Step S4: progressive training, starting from a low resolution (e.g., 4×4) and gradually increasing the resolution to improve training stability; The feature map output by each layer is identified by the discriminator, and the loss function of the discriminator is improved with quantifiable conditions to constrain the image generation of the generator; Step S5: The training adopts a phased frequency domain enhancement strategy, designs a low-frequency enhancement module and a high-frequency enhancement module, and optimizes the low frequency first in the early stage of training, optimizes the high and low frequencies jointly in the middle stage of training, and finally optimizes the high frequency details; Step S6: Finally, evaluate the trained image generation model by generating 50K images and using the entire training set as the reference distribution to calculate the FID score. This also evaluates the equivariance of the model and the computational cost of the training process.

2. In the above scheme, the step S3 constructs the GP-modulation convolution module based on traditional modulation convolution. The traditional modulation convolution module modulates and demodulates the weight of ordinary 3×3 convolution. The GP-modulation convolution module is based on the weight modulation and demodulation of grouped separation convolution, wherein the grouped separation convolution decomposes the ordinary 3×3 convolution into a cascade structure of grouped convolution and point-by-point convolution. The wavelet-based GP-modulation convolution module is a design for the second layer of point-by-point convolution of the modulation convolution. The feature map after grouped modulation convolution is decomposed into four frequency bands (LL, LH, HL, HH) by discrete wavelet transform (DWT), and feature extraction and integration are performed on each sub-band by point-by-point convolution. Finally, the inverse wavelet transform (IWT) reintegrates the multi-scale frequency information and restores it to a high-resolution image feature map.

3. In the above scheme, in the progressive training of step S4, the number of layers of the generative network is reduced from 14 to 7, while retaining the resolution growth strategy and non-uniform channel strategy of the original model; the quantifiable conditions include directional optimality, separability and injectability; now it is guaranteed that the discriminator can find the most obvious difference direction and compress it into a one-dimensional measurable space while ensuring that the difference direction after projection is consistent, and the feature extraction function is guaranteed to be an injective function; based on the above theoretical basis, the optimization of the discriminator structure can be divided into: first, the discriminator is decomposed into the inner product of the output of the feature extraction network h(x) and ω, that is, f(x) = <ω, h(x)>; then the last layer of the discriminator uses a normalized linear layer to force the weight vector ω to be located on the unit hypersphere (S D-1 ), ensuring the normalization of the projection direction; finally, the optimization goal of the entire network becomes to add the direction optimal term formula to the loss of the original discriminator: The loss function of the original StyleGAN3 discriminator loss generator remains unchanged, <, > represents the inner product operation, h is the feature extraction function, which maps the input x to the high-dimensional feature space; ω is the optimal direction selected in the feature space, λ is the hyperparameter (usually set to 1) to balance the adversarial loss and direction optimality, μ0 is the real data distribution, μ θ is the generator distribution.

4. In this embodiment, the frequency domain enhancement module of the network generated in step S5 is to divide the four sub-bands after DWT decomposition into LL low-frequency components and LH, HL, HH three high-frequency components respectively to enhance; wherein the LFE module expands the receptive field through the secondary wavelet decomposition and optimizes the global structural consistency; the LFE process is to decompose the low-frequency components X 1 LL performs a secondary decomposition to generate a coarser low-frequency X 2 LL and corresponding high frequency component X 2 LH, X 2 HL and X 2 HH, then in X 2 Perform 1×1 convolution operation on the LL subband, and perform the same convolution operation independently on all high-frequency subbands; further extract low-frequency details; then reconstruct features step by step through inverse wavelet transform, and convert the secondary processed X 2 LL features and level X 1 The LL convolution results are added element by element to finally form enhanced features covering the 4×4 area of the original image; this process retains the original high-frequency information through the residual connection between the layers, so that the 1×1 convolution can achieve receptive field expansion in the low-frequency domain after dimensionality reduction, and the number of parameters only increases linearly with the decomposition level; through the secondary wavelet decomposition, LFE gradually expands the receptive field on the low-frequency sub-band, and uses 3×3 convolution to efficiently extract multi-scale features on each sub-band, significantly improving the global modeling ability of CNN while avoiding the over-parameterization problem of traditional large convolution kernels; the high-frequency enhancement module uses the residual structure to The structure and cross-channel interaction suppress noise interference and improve the quality of high-frequency details; this module splices the three high-frequency sub-bands of horizontal, vertical and diagonal along the channel dimension to form a joint representation of multi-directional edge features, and introduces a lightweight residual path consisting of two 1×1 convolutions and ReLU activations; the core of the residual structure is that the jump connection allows the original high-frequency information to be lostless to the output layer, avoiding excessive smoothing of details by deep nonlinear transformations; the two-level 1×1 convolution dynamically fuses high-frequency content in different directions through cross-channel weight learning, amplifying effective high-frequency information while suppressing random noise.

5. In the above scheme, the relationship between high- and low-frequency signal processing and model equivariance is the core of the architecture design, which directly affects the stability and quality of the generated image during geometric transformations (such as translation and rotation). In image generation based on the StyleGAN architecture, high-frequency information in the early stage of training is non-essential information. At the same time, the role and impact of high- and low-frequency enhancement in different layers of the generative network are also different. Therefore, the method WE-GPMConv-StyleGAN of the present invention adopts a high- and low-frequency separation training strategy, that is, a low-frequency enhancement module (LFE) is added to the first layer. * ) for initial training, and then add a high-frequency enhancement module (HFE) to the last layer in the later stage. * )Joint optimization.

6. In the above scheme, the Linux platform was used to conduct training experiments of the WE-GPMConv-StyleGAN model, with Python version 3.8, PyTorch version 2.1.0+cu12, CUDA version 12.1, and CuDNN version 8.9.2; NVIDIA GeForce RTX 3090 GPU (24GB) was used for training, and the batch size for training 256×256 resolution images was set to 8. The experimental batch size on the 1024×1024 dataset was set to 2.

7. The computational costs of the most commonly used evaluation indicators for evaluating the quality of the image generation model in step S6, such as FID and the EQ-T and EQ-R evaluation models, include training time, video memory consumption, and model parameter count; the training time is based on a single graphics card with a batch size of 8 and training time per 10,000 iterations.

Images