A spatial and texture collaborative image generation method for small sample torch combustion state

Abstract

The application provides a small sample torch combustion state-oriented space and texture collaborative image generation method, comprising: obtaining and processing an image set to obtain basic sample data; performing noise correction and latent space standardization to obtain a latent representation map; constructing a position-aware spatial transformation network to generate a spatial feature map with explicit position information and geometric constraints; designing a center mask deformable convolution to obtain a structure-adaptive total feature map; designing a generator side, a discriminator side and a loss constraint to construct a lightweight adversarial network; training the lightweight adversarial network in stages using training data to obtain a generation model; generating a torch combustion state map and performing image generation quality evaluation. The application can realize high-quality generation of complex texture structures of flames and smoke in the torch combustion state image under small sample conditions, and obtain a more visually realistic torch combustion state image.

Description

Technical Field

[0001] This invention relates to the field of image generation technology, and in particular to a spatial and texture co-generation method for small-sample torch combustion state images. Background Technology

[0002] In modern industrial systems, industries such as petrochemicals, metallurgy, and coking emit large amounts of volatile organic compounds (VOCs) during production. If these harmful substances are released directly without effective treatment, they will seriously threaten the atmospheric environment and human health. While elevated flares are a primary method for VOCs treatment, incomplete combustion can still release pollutants and greenhouse gases, affecting air quality, public health, and even exacerbating climate change. Therefore, timely monitoring and control of flare combustion status is crucial for reducing pollutant emissions and protecting the environment. However, abnormal flare combustion states are strictly controlled by enterprises, and relevant data can only be obtained under very few abnormal emissions or special operating conditions, with safety and environmental risks involved in the collection process. Secondly, due to the high degree of automation in petrochemical production, flare combustion states are relatively stable under normal operating conditions. Diverse combustion state samples can only be obtained during short-term concentrated emissions such as unit start-up, shutdown, maintenance, or sudden accidents. This makes the actual usable collection time extremely limited, further exacerbating sample scarcity.

[0003] For real-time monitoring of torch combustion, deep learning methods can achieve automatic identification of combustion status and anomaly detection. However, these methods typically require a large number of high-quality samples for model training. But the availability of torch combustion status images is very limited, resulting in a severe shortage of usable training samples, which restricts the generalization ability and practical application effectiveness of torch combustion status detection models. To address the problem of scarce training samples, sample generation techniques are introduced to enrich the dataset. Current methods for few-shot image generation include regularization-based optimization, data augmentation, and architectural innovations, all aimed at reducing reliance on large datasets. Among these, Generative Adversarial Networks (GANs), especially the lightweight FastGAN, have shown great potential in torch combustion image generation tasks due to their simplified structure and enhanced training stability. FastGAN, as a lightweight model optimized for few-shot generation tasks, achieves cross-resolution feature fusion through skip connection activation modules and significantly improves training efficiency by combining a streamlined generator and discriminator architecture. This model can generate highly diverse images using limited data. However, when this model was applied to the task of generating images of torches in combustion, checkerboard artifacts were frequently observed in the generated images when flames and smoke occupied most of the image area. Furthermore, the generated images exhibited significant structural anomalies, such as the sudden, irregular appearance of smoke and an abnormal number of torch heads. These defects severely impacted the realism and usability of the generated images. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a spatial and texture co-generation method for small-sample torch combustion state images. This method can achieve high-quality generation of complex texture structures of flames and smoke in torch combustion state images under small-sample conditions, effectively improving the spatial consistency and detail restoration capability of the generated images, and obtaining more visually realistic torch combustion state images.

[0005] To achieve the above objectives, the present invention provides the following solution: a method for generating spatial and texture co-images of small-sample torch combustion states, comprising: The combustion video of the torch test base was acquired, and image sets of different combustion states were collected. The image sets were preprocessed to obtain basic sample data. Based on the aforementioned basic sample data, noise correction and latent space standardization are performed to obtain the latent representation map; A location-aware spatial transformation network is constructed, and the latent representation map is used to drive the generation of spatial feature maps with explicit location information and geometric constraints; Design a center mask deformable convolution, and use the center mask deformable convolution to process the upsampling stage of the spatial feature map to obtain the overall structural adaptability feature map; By combining the spatial feature map and the overall structural adaptability feature map, generator-side, discriminator-side, and loss constraints are designed to construct a lightweight adversarial network. By combining the basic sample data and the latent representation map, training data is obtained. The lightweight adversarial network is trained in stages using the training data, and the model parameters are output to obtain the generative model. The potential representation map is sampled using the generative model to generate a torch combustion state map. Then, FID, KID, and LPIPS are used to quantitatively evaluate the torch combustion state map to obtain the image generation quality.

[0006] Optionally, acquire combustion videos from the torch test site, collect image sets of different combustion states, preprocess the image sets to obtain basic sample data, including: The combustion video of the torch test base was acquired, and a set of 80 images under different combustion states was collected. The image set was converted to tensor, pixel normalized and standardized, and mean and variance were unified to obtain a standard image. For the standard image, a first layer of data augmentation is performed, including rotation, scaling, and color perturbation. Then, a differential augmentation strategy is used to apply color transformation and translation to the real image and the generated image to complete the second layer of data augmentation and obtain the basic sample data.

[0007] Optionally, noise correction and latent space standardization are performed based on the aforementioned basic sample data to obtain a latent representation map, including: Based on the aforementioned basic sample data, a variational inference mechanism is introduced, and the original noise is input into the variational encoder to obtain the potential distribution parameters including the mean and variance. The latent distribution parameters are reparameterized to obtain normalized noise, which is then input into a generator to obtain a latent representation map. The variational encoder maintains consistency between the posterior and prior a prior by minimizing the KL divergence, the formula for minimizing the KL divergence being: ; in, Let be the prior distribution of the latent variables. This is the posterior distribution obtained by the variational encoder based on the input samples.

[0008] Optionally, a location-aware spatial transformation network is constructed, which utilizes the latent representation map to drive the generation of spatial feature maps with explicit location information and geometric constraints, including: By combining a localization network, parameter regression, a sampler, and a position-aware encoder, a position-aware spatial transformation network is obtained, and the position-aware spatial transformation network is inserted into the low-resolution feature layer of the generator. For the potential representation map, the localization network adaptively learns and regresses spatial transformation parameters. Based on the spatial transformation parameters, the sampler performs spatial resampling on the original feature map, and integrates a learnable spatial location-aware coding matrix between the input features and the sampler. The spatial location-aware encoding matrix is ​​initialized with a Gaussian distribution. The Gaussian distribution is then added point-by-point to the original convolutional features to obtain explicit spatial-aware features. Then, using parametric regression and a sampler, affine transformations, translations, rotations, and scaling are applied to the explicit spatial-aware features to obtain a spatial feature map with explicit location information and geometric constraints.

[0009] Optionally, a center-mask deformable convolution is designed to process the upsampling stage of the spatial feature map, resulting in a total structural adaptability feature map, including: Based on the aforementioned spatial features, standard convolution is used to extract basic representation features, and offset and mask prediction branches are introduced to obtain deformable convolution offset and dynamic mask; wherein, the channel dimension of the offset is 2K², the offset channels are normalized and constrained by the Tanh activation function, and the initial mask is K² channels. A Gaussian distribution-based central region guided coding is introduced to obtain Gaussian coding. The Gaussian coding is then smoothly decayed from the center to the edge to highlight the central structural changes and suppress edge noise. Finally, the Gaussian coding is multiplied point by point with the initial mask to obtain the final center mask. The deformable convolution offset is used to adaptively deform the input spatial feature map, and the deformation intensity is adjusted using the dynamic mask to obtain a deformable convolution feature map. Then, the deformable convolution feature map and the ordinary feature map output by standard convolution are weighted and fused to generate a structurally adaptive total feature map.

[0010] Optionally, by combining the spatial feature map and the total structural adaptability feature map, generator-side, discriminator-side, and loss constraints are designed to construct a lightweight adversarial network, including: On the generator side, a position-aware spatial transformation network is introduced into the low-resolution feature layer, and skip connections are used to stabilize the spatial structure of the low-resolution feature layer. Furthermore, a center-masked deformable convolution is introduced during the upsampling from 128×128 to 256×256 resolution to complete the generator-side design. Specifically: The expression for calculating the generator offset regularity is: ; in, The offset learned by the deformable convolutional layer; The expression for calculating the generator optimization objective function is: ; in, For hyperparameters; Based on the discriminator side, hinge loss and perceptual loss are introduced to perform multi-scale discrimination, learn global and local features to distinguish complex textures, and complete the design of the discriminator side. Combining the generator side and the discriminator side, a lightweight adversarial network is obtained; where: The objective function for discriminator optimization is expressed as follows: ; in, For the hinge loss of the real sample, To generate sample hinge loss, In order to perceive loss, This is a hyperparameter.

[0011] Optionally, training data is obtained by combining the basic sample data and the latent representation map. The lightweight adversarial network is then trained in stages using the training data to output model parameters and obtain a generative model, including: By combining the basic sample data and the latent representation map, training data is obtained. The center mask deformable convolution is frozen, and the training data is used to perform feature extraction and adversarial training using standard convolution to complete the first stage of training. The deformable convolution with the center mask is gradually unfrozen to capture non-rigid structures and detailed changes in flames and smoke, completing the second stage of training. During the first and second stages of training, offset regularization terms and Adam optimizers are introduced for constraints to output model parameters, resulting in a generated model.

[0012] This invention discloses the following technical effects by providing a spatial and texture co-generation method for small-sample torch combustion states: 1. Small sample friendly: Through latent space standardization (variational inference) and structural prior injection (PA-STN, central Gaussian mask), the impact of insufficient data on generalization is significantly mitigated.

[0013] 2. Structure-Texture Collaboration: The complementary collaboration of spatial structure priors and deformable texture modeling preserves the macroscopic configuration while restoring microscopic details.

[0014] 3. Training stability: Improved hinge loss, multi-scale perception, two-stage training and offset regularization jointly enhance adversarial stability and deformation controllability.

[0015] 4. Lightweight and efficient: Modular and low-cost injection (position encoding is point-by-point addition, and deformable convolution with center mask is only used in critical upsampling segments), improving quality under controllable computation.

[0016] 5. Empirical advantages: Under conditions of extremely small sample size (80 images), FID, KID, and LPIPS are comprehensively superior to InfoGAN, WGAN-GP, and FastGAN.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention; Figure 2 This is a structural diagram of the PA-STN module provided in an embodiment of the present invention; Figure 3 A schematic diagram of the learnable position-aware coding principle provided in an embodiment of the present invention; Figure 4 This is a structural diagram of a CMDC module provided in an embodiment of the present invention; Figure 5 This is a diagram illustrating the overall structure of the MDCGAN generator provided in an embodiment of the present invention. Figure 6 Comparison of generated image results of the comparison model provided in the embodiments of the present invention; Figure 7 A comparison image between the generated image and the real image provided in the embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] like Figure 1 As shown, this invention provides a spatial and texture co-generation method for small-sample torch combustion states, comprising: Step 1: Obtain combustion videos from the torch test site, collect image sets of different combustion states, and preprocess the image sets to obtain basic sample data; Step 1 includes: The combustion video of the torch test base was acquired, and a set of 80 images under different combustion states was collected. The image set was converted to tensor, pixel normalized to the [0,1] interval and standardized to the [-1,1] interval, and the mean and variance were unified to obtain a standard image.

[0023] For the standard image, a first layer of data augmentation is performed, including rotation, scaling, and color perturbation. Then, a differential augmentation strategy is used to apply color transformation and translation to the real image and the generated image to complete the second layer of data augmentation and obtain the basic sample data.

[0024] Step 2: Based on the aforementioned basic sample data, perform noise correction and latent space standardization to obtain the latent representation map; Step 2 includes: Based on the aforementioned basic sample data, a variational inference mechanism is introduced. The original noise is input into the variational encoder to obtain the mean value. With variance The latent distribution parameters.

[0025] The latent distribution parameters are reparameterized to obtain normalized noise, which is then input into a generator to obtain a latent representation map. The variational encoder maintains consistency between the posterior and prior a prior by minimizing the KL divergence, the formula for minimizing the KL divergence being: ; in, Let be the prior distribution of the latent variables. The posterior distribution is obtained by the variational encoder based on the input samples. KL divergence optimization ensures that the latent variable distribution output by the encoder is consistent with the prior distribution, thus providing the generator with a structured and standardized input representation. Noise, after variational inference correction and standardization, exhibits stronger data relevance and structure, effectively suppressing meaningless noise interference. The standardized latent space makes the generator's response to the input smoother and more controllable, contributing to the generation of high-quality, coherent flame and smoke images.

[0026] Step 3: Construct a location-aware spatial transformation network, and use the latent representation map to drive the generation of spatial feature maps with explicit location information and geometric constraints; Step 3 includes: By combining a localization network, parameter regression, a sampler, and a position-aware encoder, a position-aware spatial transformation network is obtained, and the position-aware spatial transformation network (such as...) Figure 2 The PA-STN module shown is inserted into the low-resolution feature layer (8×8 and 16×16 stages) of the generator.

[0027] For the latent representation map, the localization network adaptively learns and regresses spatial transformation parameters. Based on the spatial transformation parameters, the sampler performs spatial resampling on the original feature map, and integrates a learnable spatial location-aware coding matrix P between the input features and the sampler.

[0028] ; Where C is the number of channels, H is the height of the feature map, and W is the width of the feature map, initialized as a Gaussian distribution with a mean of 0 and a standard deviation σ of 0.02. By varying the parameters at different spatial locations, the encoding provides location information, thereby helping the network distinguish features in different regions.

[0029] The spatial location-aware encoding matrix is ​​initialized with a Gaussian distribution, and the Gaussian distribution is combined with the original convolutional features. By adding the features point by point, we obtain the explicit spatial perception features. Compared to using sine coding or additional calculation modules, this method is simply an addition operation and does not increase computational complexity. The calculation process is as follows: ; like Figure 3 As shown, the schematic diagram illustrates the principle of computationally learnable positional encoding. In this way, the model can obtain more explicit positional guidance in the early stages, thereby improving the generation quality of complex structures. By introducing position-aware encoding, the PA-STN module can explicitly model the global and local positional information of each spatial location in the feature map, thereby enhancing the spatial transformation network's ability to perceive and model spatial structures, strengthening the model's robustness to spatial deformation of torch combustion states under small sample conditions, and improving generation quality.

[0030] Then, using parametric regression and a sampler, affine transformations, translations, rotations, and scaling are applied to the explicit spatial perception features to obtain a spatial feature map with explicit positional information and geometric constraints.

[0031] Step 4, as follows Figure 4 As shown, a center mask deformable convolution (CMDC module) is designed. The center mask deformable convolution is used to process the upsampling stage of the spatial feature map to obtain the overall structural adaptability feature map; step 4 includes: Based on the spatial features, basic representation features are extracted using standard convolution, and offset and mask prediction branches are introduced to obtain deformable convolution offset and dynamic mask; wherein, the channel dimension of the offset is 2K², K is the convolution kernel size, the offset channels are normalized and constrained by the Tanh activation function to control the offset range and avoid excessive deformation causing feature distortion, and the initial mask is K² channels.

[0032] After mask convolution generates an initial mask image for K² channels, a central region guided coding based on Gaussian distribution is introduced to obtain Gaussian coding. The Gaussian coding is smoothly decayed from the center to the edge to highlight the central structural changes and suppress edge noise. The Gaussian coding is then multiplied point by point with the initial mask to obtain the final central mask, thereby improving the natural transition effect and structural stability of the deformed region.

[0033] After obtaining the offset and center mask, the deformable convolution offset is used to adaptively deform the input spatial feature map, and the deformation intensity is adjusted using the dynamic mask to obtain a deformable convolution feature map. Then, the deformable convolution feature map and the ordinary feature map output by standard convolution are weighted and fused to generate a structurally adaptive total feature map.

[0034] Step 5: Combining the spatial feature map and the structural adaptability total feature map, design the generator side, discriminator side, and loss constraints to construct a lightweight adversarial network (MDCGAN model); Step 5 includes: like Figure 5As shown, based on the generator side, a position-aware spatial transformation network is introduced into the low-resolution feature layers (8×8 and 16×16 stages) to enhance the model's ability to reconstruct high-quality features of torch burning images. Skip connections are used to stabilize the spatial structure of the low-resolution feature layers, and a center-masked deformable convolution is introduced in the stage of upsampling from 128×128 to 256×256 resolution, completing the generator-side design. Wherein: Generator offset regularization (L2 regularization) is used to constrain the amplitude of the offset field and prevent unstable deformations from affecting the generation quality. The formula is: ; in, The offsets learned by the deformable convolutional layer are given. Furthermore, during the training of the generator G, we employ adversarial loss to maximize the discriminator's score, encouraging the generator to produce more realistic images. The expression for the generator's optimization objective function is: ; in, This is a hyperparameter used to balance the effects of various losses.

[0035] Based on the discriminator side, hinge loss and perceptual loss are introduced to perform multi-scale discrimination, learn global and local features to distinguish complex textures, and complete the design of the discriminator side. Combining the generator side and the discriminator side, a lightweight adversarial network is obtained; where: An improved hinge loss is used as the adversarial loss to ensure the discriminator can effectively distinguish between real and generated samples. The improved hinge loss function for a real sample *x* is defined as: Introducing perceptual loss into the discriminator This ensures that the discriminator can learn global and local features at multiple scales. The discriminator reconstructs at different scales and calculates the perceptual error between the original input and the new input, with the loss function as follows: in, For the feature extraction function based on VGG network, and and Let represent the image reconstructed by the discriminator and the corresponding ground truth image, respectively. Combining the above losses, the objective function for discriminator optimization is expressed as: ; in, For the hinge loss of the real sample, To generate sample hinge loss, In order to perceive loss, This is a hyperparameter.

[0036] Step 6: Combining the basic sample data and the latent representation map, training data is obtained. The lightweight adversarial network is trained in stages using the training data, and model parameters are output to obtain the generative model. Step 6 includes: By combining the basic sample data and the latent representation map, training data is obtained. The center mask deformable convolution is frozen, and the training data is used for feature extraction and adversarial training using standard convolution to complete the first stage of training.

[0037] The deformable convolution with the central mask is gradually unfrozen to capture non-rigid structures and subtle changes in flames and smoke, completing the second stage of training. During both the first and second training stages, offset regularization terms and the Adam optimizer are introduced for constraints to output model parameters, resulting in the generated model. The offset range is constrained by calculating the L2 regularization loss of the offset, thus preventing unreasonable deformation. During training, the Adam optimizer iteratively updates the network parameters, setting appropriate learning rates and batch sizes.

[0038] Step 7: Use the generative model to sample the potential representation map to generate a torch combustion state map, and then use FID, KID and LPIPS to quantitatively evaluate the torch combustion state map to obtain the image generation quality.

[0039] A trained generator network is used to generate high-resolution images of torch combustion status by inputting a noise vector that has been normalized and corrected through variational inference. The generated images are quantitatively evaluated using FID, KID, and LPIPS metrics, and compared with results from existing methods. FID assesses the similarity between the generated and real images by calculating the distribution distance between them in the Inception network feature space. The formula is as follows: in, and These are the mean values ​​of the generated image and the real image in the Inception feature space, respectively. and These are the covariance matrices of the generated image and the real image in the feature space, respectively. This represents the trace of the matrix. Since the sample size of the real-world dataset is too small, KID is also used as a metric. The calculation steps for KID are similar to FID, the main difference being that KID uses the kernel trick to calculate a more stable distance and typically does not require calculating the square root of the covariance matrix. The formula for calculating KID is as follows: in, , These represent the feature vectors of the real image and the generated image in the feature space, respectively. , These are samples extracted from the real image set and the generated image set, respectively. It is the feature mapping function for calculating kernel distance, using a Gaussian kernel.

[0040] To more accurately evaluate the quality of generated images, perceptual loss LPIPS is used as an additional evaluation metric. LPIPS does not rely on the overall statistical distribution but calculates the distance between images based on local perceptual differences in the deep feature space. A pre-trained deep neural network, VGG, is used to extract multi-layer features from the images. Feature channels are normalized to ensure that features from different layers have equal contributions. The Euclidean distance between the generated and ground images is calculated in the feature space, and the final LPIPS value is obtained by weighted averaging. The LPIPS formula is as follows: in, , Representing real images respectively and generating images In the neural network Representation in the layer feature space. These are the perceptual weights obtained during training, used to weight the feature contributions of different layers. Finally, this invention evaluates the generative performance of different models using three evaluation metrics: FID, KID, and LPIPS.

[0041] To fully verify the effectiveness of MDCGAN in generating images of torch combustion states from a small sample size, the results of the comparative experiments are shown in Table 1. The comparison of the generated image results from the different models is shown in the following figure. Figure 6 As shown.

[0042] Table 1 Comparison of experimental results

[0043] The comparison between the image generated by the MDCGAN model and the real image in the torch combustion state generation task is shown in the figure below. Figure 7 As shown, the generated images conform to the physical characteristics of the actual combustion process, while exhibiting good realism in terms of flame details, smoke dynamics, and the structural integrity of the torch.

[0044] To verify the effectiveness of the PA-STN and CMDC modules, corresponding ablation experiments were designed, and the results are shown in Table 2. The final MDCGAN model outperformed the baseline model in both FID and KID, indicating that its generated images are closer to real data in terms of overall distribution and texture details. Meanwhile, the decrease in LPIPS score indicates that the images generated by MDCGAN are more perceptually similar to real samples. This verifies the effectiveness of the PA-STN and CMDC modules in improving generation performance.

[0045] Table 2 Ablation Experiment Results

[0046] Evaluation results show that the images generated by the method of the present invention have significant advantages in terms of structural integrity and texture detail. The present invention can generate high-fidelity images of the torch combustion state.

[0047] Therefore, this invention provides a spatial and texture collaborative image generation method for small sample torch combustion states, which can achieve high-quality generation of complex texture structures of flames and smoke in torch combustion state images under small sample conditions, effectively improving the spatial consistency and detail restoration capability of the generated images, and obtaining more visually realistic torch combustion state images.

[0048] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0049] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for generating spatial and texture co-images of small-sample torch combustion states, characterized in that, include: The combustion video of the torch test base was acquired, and image sets of different combustion states were collected. The image sets were preprocessed to obtain basic sample data. Based on the aforementioned basic sample data, noise correction and latent space standardization are performed to obtain the latent representation map; A location-aware spatial transformation network is constructed, and the latent representation map is used to drive the generation of spatial feature maps with explicit location information and geometric constraints; Design a center mask deformable convolution, and use the center mask deformable convolution to process the upsampling stage of the spatial feature map to obtain the overall structural adaptability feature map; By combining the spatial feature map and the overall structural adaptability feature map, generator-side, discriminator-side, and loss constraints are designed to construct a lightweight adversarial network. By combining the basic sample data and the latent representation map, training data is obtained. The lightweight adversarial network is trained in stages using the training data, and the model parameters are output to obtain the generative model. The latent representation map is sampled using the generative model to generate a torch combustion state map. Then, FID, KID, and LPIPS are used to quantitatively evaluate the torch combustion state map to obtain the image generation quality. The design incorporates a center-mask deformable convolution. This center-mask deformable convolution is used to process the upsampling stage of the spatial feature map, yielding a total structurally adaptive feature map, including: Based on the aforementioned spatial features, standard convolution is used to extract fundamental representation features, and offset and mask prediction branches are introduced to obtain deformable convolutional offsets and dynamic masks; wherein, the channel dimension of the offset is 2K. 2 The offset channels are normalized using the Tanh activation function, with an initial mask of K. 2 Channels; K is the kernel size; A Gaussian distribution-based central region guided coding is introduced to obtain Gaussian coding. The Gaussian coding is then smoothly decayed from the center to the edge to highlight the central structural changes and suppress edge noise. Finally, the Gaussian coding is multiplied point by point with the initial mask to obtain the final center mask. The deformable convolution offset is used to adaptively deform the input spatial feature map, and the deformation intensity is adjusted using the dynamic mask to obtain a deformable convolution feature map. Then, the deformable convolution feature map and the ordinary feature map output by standard convolution are weighted and fused to generate a structurally adaptive total feature map.

2. The spatial and texture co-generation method for small-sample torch combustion states according to claim 1, characterized in that, The combustion video of the torch test site was acquired, and image sets of different combustion states were collected. The image sets were preprocessed to obtain basic sample data, including: The combustion video of the torch test base was acquired, and a set of 80 images under different combustion states was collected. The image set was converted to tensor, pixel normalized and standardized, and mean and variance were unified to obtain a standard image. For the standard image, a first layer of data augmentation is performed, including rotation, scaling, and color perturbation. Then, a differential augmentation strategy is used to apply color transformation and translation to the real image and the generated image to complete the second layer of data augmentation and obtain the basic sample data.

3. The spatial and texture co-generation method for small-sample torch combustion states according to claim 2, characterized in that, Based on the aforementioned basic sample data, noise correction and latent space standardization are performed to obtain a latent representation map, including: Based on the aforementioned basic sample data, a variational inference mechanism is introduced, and the original noise is input into the variational encoder to obtain the potential distribution parameters including the mean and variance. The latent distribution parameters are reparameterized to obtain normalized noise, which is then input into a generator to obtain a latent representation map. The variational encoder maintains consistency between the posterior and prior a prior by minimizing the KL divergence, the formula for minimizing the KL divergence being: ； in, Let KL divergence be the KL divergence. For the input sample, As latent variables, Let be the prior distribution of the latent variables. This is the posterior distribution obtained by the variational encoder based on the input samples.

4. The spatial and texture co-generation method for small-sample torch combustion states according to claim 3, characterized in that, Constructing a location-aware spatial transformation network, and using the latent representation map to drive the generation of spatial feature maps with explicit location information and geometric constraints, including: By combining a localization network, parameter regression, a sampler, and a position-aware encoder, a position-aware spatial transformation network is obtained, and the position-aware spatial transformation network is inserted into the low-resolution feature layer of the generator. For the potential representation map, the localization network adaptively learns and regresses spatial transformation parameters. Based on the spatial transformation parameters, the sampler performs spatial resampling on the original feature map, and integrates a learnable spatial location-aware coding matrix between the input features and the sampler. The spatial location-aware encoding matrix is ​​initialized with a Gaussian distribution. The Gaussian distribution is then added point-by-point to the original convolutional features to obtain explicit spatial-aware features. Then, using parametric regression and a sampler, affine transformations, translations, rotations, and scaling are applied to the explicit spatial-aware features to obtain a spatial feature map with explicit location information and geometric constraints.

5. The spatial and texture co-generation method for small-sample torch combustion states according to claim 4, characterized in that, Combining the spatial feature map and the structural adaptive total feature map, generator-side, discriminator-side, and loss constraints are designed to construct a lightweight adversarial network, including: On the generator side, a position-aware spatial transformation network is introduced into the low-resolution feature layer, and skip connections are used to stabilize the spatial structure of the low-resolution feature layer. Furthermore, a center-masked deformable convolution is introduced during the upsampling from 128×128 to 256×256 resolution to complete the generator-side design. Specifically: The expression for calculating the generator offset regularity is: ； in, The offset learned by the deformable convolutional layer; For generator offset regularization terms; The expression for calculating the generator optimization objective function is: ； in, For hyperparameters; Optimize the objective function for the generator. For generator, For discriminators; Based on the discriminator side, hinge loss and perceptual loss are introduced to perform multi-scale discrimination, learn global and local features to distinguish complex textures, and complete the design of the discriminator side. Combining the generator side and the discriminator side, a lightweight adversarial network is obtained; where: The objective function for discriminator optimization is expressed as follows: ； in, For the hinge loss of the real sample, To generate sample hinge loss, In order to perceive loss, This is a hyperparameter.

6. The spatial and texture co-generation method for small-sample torch combustion states according to claim 5, characterized in that, By combining the basic sample data and the latent representation map, training data is obtained. The lightweight adversarial network is then trained in stages using this training data, outputting model parameters to obtain a generative model, including: By combining the basic sample data and the latent representation map, training data is obtained. The center mask deformable convolution is frozen, and the training data is used to perform feature extraction and adversarial training using standard convolution to complete the first stage of training. The deformable convolution with the center mask is gradually unfrozen to capture non-rigid structures and detailed changes in flames and smoke, completing the second stage of training. During the first and second stages of training, offset regularization terms and Adam optimizers are introduced for constraints to output model parameters, resulting in a generated model.

Images