A polymeric infrared image super-resolution reconstruction network for aluminum reduction cells

Abstract

The application discloses a kind of polymeric infrared image super-resolution reconstruction networks for aluminum electrolytic cell, comprising: shallow feature extraction subnetwork, for shallow feature extraction to the infrared image after pre-processing, obtain shallow feature;Deep feature extraction subnetwork includes n mutually connected residual blocks, wherein residual block includes: multi-scale compression diffusion Token aggregation module MS-CDTA and local area attention module;Local area attention module carries out local attention processing to the feature output by MS-CDTA finally forms feature flow;Image reconstruction subnetwork is used to reconstruct the image result consistent with the size of real high-resolution infrared image based on feature flow output.

Description

Technical Field

[0001] This invention relates to the field of infrared image super-resolution technology, and more specifically to a polymerized infrared image super-resolution reconstruction network for aluminum electrolytic cells. Background Technology

[0002] Infrared super-resolution imaging technology has significant application value in monitoring the operational status of aluminum electrolytic cells. As a core piece of equipment in non-ferrous metal electrolytic production, the thermal state of the aluminum electrolytic cell directly affects production safety and energy efficiency. During operation, the electrolytic cell is prone to localized overheating due to problems such as cathode plate deformation and anode short circuits, making comprehensive and accurate monitoring difficult with traditional contact-based temperature measurement methods. While infrared thermal imaging technology can acquire temperature distribution non-contactly, the resulting images are often low-resolution and blurry due to hardware costs, optical limitations, and noise interference in the production environment, affecting early fault identification and accurate location. Therefore, using deep learning algorithms to perform super-resolution reconstruction of low-resolution infrared images to enhance image details and spatial resolution is of great significance for achieving refined temperature analysis and fault early warning in electrolytic cells. Current deep learning-based super-resolution methods mainly revolve around convolutional neural networks, generative adversarial networks, and Transformer models. Through end-to-end mapping, adversarial training, or global attention mechanisms, they exhibit their respective advantages in local feature extraction, texture generation, and long-range dependency modeling, providing multiple technical paths for improving the quality of infrared images of electrolytic cells.

[0003] However, existing methods still face significant challenges when dealing with the specific industrial scenario of aluminum electrolysis cells. Convolutional neural networks (CNNs) are limited by their local receptive field, making it difficult to model the long-distance relationship between large hot areas within the cell and the background. At high magnification, they are prone to incomplete contours and texture distortion. Generative adversarial networks (GANs) are unstable during training, exhibiting weak generalization ability on limited infrared datasets of aluminum electrolysis cells, and may generate false hot spots, interfering with actual temperature measurement and fault diagnosis. While Transformer models can capture global context, their high computational complexity, large number of parameters, and reliance on large-scale training data make them prone to overfitting under limited data conditions. Even after lightweighting, they are still difficult to deploy on low-computing-power devices such as embedded systems, limiting their application in industrial real-time monitoring scenarios. Therefore, it is necessary to address the characteristics of infrared images from aluminum electrolysis cells, improving resolution while maintaining algorithm stability, lightweight design, and detail fidelity to meet the multiple requirements of real-time performance, reliability, and accuracy in actual production environments. Summary of the Invention

[0004] In view of this, the present invention provides an aggregated infrared image super-resolution reconstruction network for aluminum electrolytic cells, which at least solves the problems of single-scale token aggregation in existing infrared image super-resolution reconstruction methods, which makes it difficult to take into account both global structure and local details and lacks global information interaction.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A super-resolution reconstruction network for polymerized infrared images of aluminum electrolytic cells includes: a shallow feature extraction subnetwork, a deep feature extraction subnetwork, and an image reconstruction subnetwork; A shallow feature extraction subnetwork is used to extract shallow features from the preprocessed infrared image to obtain shallow features. ; The deep feature extraction subnetwork consists of n interconnected residual blocks, and the shallow feature extraction subnetwork consists of n interconnected residual blocks. After passing through n residual blocks, the final feature stream is output. Each residual block includes a multi-scale compression-diffusion token aggregation module (MS-CDTA) and a local region attention module. MS-CDTA generates token representations of different granularities based on the input features through multi-scale adaptive pooling and low-dimensional orthogonal subspace projection. It also introduces a self-attention mechanism to compress, refine, and diffuse the token representations of different granularities, outputting feature information at different scales. After feature fusion and detail enhancement, the feature set is output. Local attention module Local attention processing is performed to form a feature flow; An image reconstruction subnetwork is used to reconstruct an image result with the same size as the real high-resolution infrared image based on the final feature flow.

[0006] Preferably, the preprocessing includes the following: Median filtering was used to remove thermal noise and salt-and-pepper noise from infrared images; The image is normalized to map the pixel values ​​to the [0,1] range, thus eliminating the magnitude difference between different images; Data augmentation is performed by randomly flipping the registered low-resolution LR and high-resolution HR image pairs horizontally and vertically. The training set, validation set, and test set are divided proportionally to ensure that the test set images of aluminum electrolytic cells cover scenarios not included in the training set.

[0007] Preferably, the shallow feature extraction subnetwork includes the following: The preprocessed infrared image data is input into a shallow feature extraction subnetwork, which extracts shallow features of the image through convolution operations. ,in, , These are the height and width of the low-resolution image, respectively. To extract the number of feature channels.

[0008] Preferably, the MS-CDTA includes multiple compressed diffusion token aggregation modules (CDTAs) of different scales, wherein each CDTA specifically includes: a token representing an initialization unit, which specifically includes the following: CDTA acquires the features after normalizing the input features. , will feature Features projected onto a set of low-dimensional orthogonal subspaces. Dimensions The number of channels is , The deep feature extraction subnetwork represents the first... One residual block, Indicates the first The first residual block A subspace of attention head projection; Through average pooling layer Dimensional reduction Each token represents , , The three scales are represented, and the specific tokens for the three scales are 16, 32, and 64 respectively. Through cross attention pairs Perform iterative optimization, for and Perform a linear transformation. As a query token vector , As a key token vector Sum Value Token Vector ; pass Generates a weight matrix for cross-attention by interacting with global information. To measure Input Token Pair The degree of contribution; weight matrix and Multiplication and addition The final token is obtained by adding them together. : ; ; in, For learnable parameters, This indicates element-wise addition.

[0009] Preferably, each CDTA also specifically includes: a Token representing a compression, refining, and diffusion unit, specifically including the following: right Perform a linear transformation to obtain the corresponding query token vector. Token vector Value Token Vector By introducing a self-attention mechanism, token representations of different granularities are compressed, refined, and diffused. The final feature information output by CDTA is... for: ; in, Represents the subspace of all attention head projections. For the compression process, and Multiplication is a diffusion process, through Left multiplication projects the output of CDTA back into higher-dimensional space. It should remain consistent with the input features.

[0010] Preferably, MS-CDTA also includes: a feature fusion module: The feature information output by all CDTA modules is used as input features for the feature fusion module. The input features are weighted and fused, and the fused multi-scale features are added to the input features through residual connections to avoid the loss of low-level feature information. ; in, For normalization processing, The quantity represented by the token. This represents three scales, with the corresponding token quantities being 16, 32, and 64, respectively. Learnable weight coefficients for CDTA representing different token quantities; GeLU is the activation function. The feature information output by CDTA for different tokens representing a certain number of tokens.

[0011] Preferably, the specific content of the local region attention module includes: Input features Normalization process is performed to obtain ,in, , These are the height and width of the low-resolution image, respectively. To extract the number of feature channels; Will According to fixed dimensions Divide into non-overlapping sliding windows, then pass through a fully connected layer to output deformable offsets: ; in Indicates the sampling location characteristics, This indicates a local area. Indicates the predicted offset. This represents a lightweight offset prediction module; The features at the current sampling position are processed by the weight matrix. Get the query vector The deformed neighborhood sampling features are processed by the weight matrix. Obtain the key vector Attention weights are calculated using the softmax normalization function. : ; in, It is an exponential function. Adjusted local sampling location features and belong Offset in different directions; The size is The window; With the weight matrix Obtain the value vector Perform element-wise multiplication. It is the Kronecker product; Features after attention processing via deformable sliding window for: .

[0012] Preferably, the deep feature extraction subnetwork further includes a first frequency domain forward feedback network and a second frequency domain forward feedback network, which are respectively connected to the MS-CDTA and the local region attention module, and respectively process the multi-scale features after feature fusion and the local region attention module. The feature stream formed after local attention processing is used as the input features for detail enhancement: After normalizing the input features, we obtain ,in, , These are the height and width of the low-resolution image, respectively. To extract the number of feature channels; Spatial feature preprocessing: Integrating cross-channel information from infrared images through pointwise convolution PConv yields... Subsequently, the local spatial correlation of infrared features is enhanced by deep convolution (DConv), and the features are refined and denoised to obtain... ; After the second PConv adjusts the number of feature channels to the adaptation dimension for frequency domain processing, the preprocessed features are obtained. ; right The spatial feature flattening operation, frequency domain filtering operation, and quantization matrix filtering operation are performed sequentially to obtain the features. ; Finally Features are obtained by converting back to the spatial domain sequence. and to Perform a folding operation to restore the spatial positions in sequence, and obtain... ; Input features are combined with residual connections. By combining these features, we obtain a feature set with enhanced details.

[0013] Preferably, the specific content of the image reconstruction subnetwork includes: Construct a convolutional layer to remove redundant information from the final feature stream output by the deep feature extraction subnetwork and unify the feature dimensions; The output features of the convolutional layer are added to the deep feature extraction subnetwork through residual connections. The added features are then upsampled and input into the convolutional layer. Upsampling enlarges the size of the feature map to the target super-resolution factor. The convolutional layer restores the spatial details of the image, and finally reconstructs the output image with the same size as the real high-resolution infrared image.

[0014] Preferably, it also includes a training sub-network for training the aggregated infrared image super-resolution reconstruction network, specifically including the following: Training is supervised using a dual-domain constraint loss function, where the loss function is: ; in, For the total loss, L1 loss in the spatiotemporal domain refers to the difference between the reconstructed image and the real image. For frequency domain L1 loss, the difference between the reconstructed image (transformed from the spatiotemporal domain image to the frequency domain) and the real image is... To reconstruct the image, For real images, Indicates Fourier transform, These are the weighting coefficients for the frequency domain loss. This represents the mean absolute error (MAE).

[0015] Preferably, the training process specifically includes: Input LR-HR infrared image data pairs of aluminum electrolysis cells, set training parameters, and train the network in a loop: S1. Initialize network parameters; S2. Input the data into the network to obtain the reconstructed values; S3. Calculate the total loss and backpropagate the network parameters; S4. If the training count is reached, stop training; otherwise, repeat S2-S4.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a polymeric infrared image super-resolution reconstruction network for aluminum electrolytic cells, which has the following beneficial effects: This invention proposes an infrared image super-resolution network based on multi-scale token aggregation. The network adopts an end-to-end architecture of shallow feature extraction → deep feature extraction enhancement → image reconstruction. By supervising the network training with a loss function, the shallow feature extraction module, deep feature extraction module, and image reconstruction module can be effectively coordinated during training, improving the infrared image super-resolution reconstruction effect. Addressing the shortcomings of existing super-resolution methods in capturing insufficient long-distance dependent information and the loss of local information due to fixed window attention, this invention achieves efficient global feature interaction and context compression with near-linear complexity through a multi-scale compressed diffusion token aggregation module. It utilizes a frequency-domain forward feedback network for directional high-frequency information enhancement and noise suppression in the frequency domain, and leverages a region deformable attention module for adaptive local information refinement of edges and texture structures. This allows for collaborative modeling of the image at multiple levels, including global structure, local details, and geometric structure, comprehensively improving reconstruction quality. Experiments were conducted on an aluminum electrolytic cell infrared image super-resolution dataset, and the results show that the proposed network achieves excellent performance in two objective metrics: Peak Signal-to-Noise Ratio (PNSR) and Structural Similarity (SSIM). Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0018] Figure 1 This is a structural diagram of the multi-scale token aggregation module provided in an embodiment of the present invention; Figure 2 This is a structural diagram of the compressed diffusion token aggregation module provided in an embodiment of the present invention; Figure 3 This is a diagram of the frequency domain forward feedback network structure provided in an embodiment of the present invention; Figure 4 This is a structural diagram of a local region attention module provided in an embodiment of the present invention; Figure 5 This is a diagram of the attention structure of a deformable sliding window provided in an embodiment of the present invention; Figure 6 This is a diagram of a multi-scale token aggregation infrared image super-resolution network structure provided in an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] This invention provides a super-resolution reconstruction network for polymerized infrared images of aluminum electrolytic cells, comprising: a shallow feature extraction subnetwork, a deep feature extraction subnetwork, and an image reconstruction subnetwork; A shallow feature extraction subnetwork is used to extract shallow features from the preprocessed infrared image to obtain shallow features. ; The deep feature extraction subnetwork consists of n interconnected residual blocks, and the shallow feature extraction subnetwork consists of n interconnected residual blocks. After passing through n residual blocks, the final feature stream is output. Each residual block includes a multi-scale compression-diffusion token aggregation module (MS-CDTA) and a local region attention module. MS-CDTA generates token representations of different granularities based on the input features through multi-scale adaptive pooling and low-dimensional orthogonal subspace projection. It also introduces a self-attention mechanism to compress, refine, and diffuse the token representations of different granularities, outputting feature information at different scales. After feature fusion and detail enhancement, the feature set is output. Local attention module Local attention processing is performed to form a feature flow; An image reconstruction subnetwork is used to reconstruct an image result with the same size as the real high-resolution infrared image based on the final feature flow.

[0021] To further implement the above technical solution, the specific content of preprocessing includes: Median filtering was used to remove thermal noise and salt-and-pepper noise from infrared images; The image is normalized to map the pixel values ​​to the [0,1] range, thus eliminating the magnitude difference between different images; Data augmentation is performed by randomly flipping the registered low-resolution LR and high-resolution HR image pairs horizontally and vertically. The training set, validation set, and test set are divided proportionally to ensure that the test set images of aluminum electrolytic cells cover scenarios not included in the training set.

[0022] It should be noted that: The specific calculation formula for normalization is as follows: ; in and These are the minimum and maximum pixel values ​​for a single image, respectively. The image pixel values ​​are normalized. In data augmentation, the flip angle can be selected as 90 degrees, 180 degrees, or 270 degrees to expand the dataset size. The dataset can be divided as needed. In this embodiment, the training set, validation set, and test set are divided in a 6:2:2 ratio.

[0023] To further implement the above technical solution, the shallow feature extraction subnetwork includes the following: The preprocessed infrared image data is input into a shallow feature extraction subnetwork, which extracts shallow features of the image through convolution operations. ,in, , These are the height and width of the low-resolution image, respectively. To extract the number of feature channels.

[0024] It should be noted that: The shallow feature extraction subnetwork consists of multiple convolutional layers, which in this embodiment are composed of 3×3 convolutional layers.

[0025] To further implement the above technical solution, MS-CDTA includes multiple compressed diffusion token aggregation modules (CDTAs) of different scales, wherein each CDTA specifically includes: a token representative initialization unit, which specifically includes the following content; The deep feature extraction module of this sub-network consists of n interconnected residual blocks. The number of residual blocks, n, is an adjustable hyperparameter; in this embodiment, n is set to 8. Each residual block consists of a multi-scale compressed-diffusion token aggregation module (MS-CDTA) and a local region attention module (in practical applications, a convolutional layer is usually set after the local region attention module to refine local features and form a higher-quality feature stream). Finally, the input features and the extracted features are added and fused through residual connections. The MS-CDTA module of this invention generates three sets of tokens with different granularities through multi-scale adaptive pooling and low-dimensional orthogonal subspace projection. This accurately covers the local thermal details, mid-level structure, and global context of infrared images, while suppressing thermal noise in aluminum electrolysis cells and reducing computational redundancy. The compressed-diffusion mechanism then achieves efficient interaction of global information with low complexity, avoiding detail loss. The Local Region Attention Module (LRSA) takes over global features and dynamically adjusts the attention sampling position by learning content-aware offsets. This breaks through the limitations of fixed windows, enabling flexible connections across window features, eliminating block artifacts while focusing on high-value infrared regions, and accurately optimizing local thermal details and edges. Furthermore, a Frequency Domain Feedforward Network (EDFFN) is added to the ends of the MS-CDTA module and the Local Region Attention Module. Unlike the Feedforward Networks (FFNs) used in most infrared image super-resolution networks, this network can move the frequency domain filtering operation to a lower channel stage, significantly reducing computational costs. It strengthens local correlations through spatial convolution before performing frequency domain filtering, directionally preserving high-frequency details and mid-frequency structures while suppressing redundant noise.

[0026] Each CDTA specifically includes: Token representing an initialization unit, such as... Figure 1 As shown, the features extracted by the shallow feature pre-module are first... Input the MS-CDTA module, which contains three Compressed Diffusion Token Aggregation Modules (CDTAs) at different scales. Specifically, each CDTA has a different number of token representations. Different scales represent features of different scales. In this network, the scale sizes (tokens represent the number of tokens) are set to 16, 32, and 64 respectively. The CDTA calculation process is as follows: Figure 2 As shown, each CDTA acquires the features after normalizing the input features. Then the features Projecting features onto a set of low-dimensional orthogonal subspaces decomposes high-dimensional features into multiple sets of independent subspaces, reducing linear complexity and computational redundancy. Features projected onto the subspaces... Dimensions The number of channels is , Indicates the first The first layer The subspace of the attention head projection is then used to reduce the dimensionality of the input projection features through an average pooling layer. Each token represents , , This represents three scales, with corresponding token counts of 16, 32, and 64 respectively. The average pooling calculation formula is as follows: ; in This indicates that gradient operators should be stopped to avoid affecting subsequent processes. Cross-attention is used to represent... Perform iterative optimization. and Perform a linear transformation. As a query token vector , As a key token vector Sum Value Token Vector This is used for subsequent attention calculations. Generates a weight matrix for cross-attention by interacting with global information. To measure Input Token Pair The contribution of each component. Finally, the weight matrix is... and ( Multiply and combine The final token is obtained by adding them together. The specific calculation formula is as follows: ; ; in It is a learnable parameter that can adaptively adjust and optimize the token representation during the iteration process. This indicates element-wise addition.

[0027] To further implement the above technical solution, each CDTA also specifically includes: Token representing a compression, refining, and diffusion unit, specifically including the following: After completing the representative initialization, for Perform a linear transformation and use it as the query token vector. Token vector Sum Value Token Vector This is used for subsequent attention calculations. A self-attention mechanism allows representatives to interact and initiate a representative contraction operation to refine the representatives and strengthen structured features. After compression, the transpose of the cross-attention weight matrix from the representative extraction stage is directly reused. This ensures that the association weight between the representative and the input token is completely consistent throughout the extraction stage, preventing weight misalignment during information transmission. The specific calculation formula is as follows: ; in Represents the subspace of all attention head projections. For the compression process, and Multiplication is a diffusion process, and finally it passes through... Left multiplication projects the output of the CDTA module back into a higher-dimensional space. It maintains consistency with the input features. This compression-diffusion process allows all tokens to share the structured information they represent, preserving local details while integrating into the global structure, thus avoiding the limitations of local attention.

[0028] To further implement the above technical solution, MS-CDTA also includes: a feature fusion module. The multi-scale feature set is weighted and fused through a feature fusion module. The fused multi-scale features are then added to the input features via residual connections to avoid losing low-level feature information. The specific calculation formula is as follows: ; in, To normalize and stabilize the feature distribution at different scales, The quantity represented by the token. This represents three scales, with the corresponding token quantities being 16, 32, and 64, respectively. The learnable weight coefficients of the compressed diffusion token aggregation module, which represent different token quantities, reflect the importance of features at different scales. These coefficients can be continuously optimized through training to improve the fusion effect of multi-scale features. GeLU is used as the activation function to improve the nonlinear expressive power of features; These are the feature information output by the compressed diffusion token aggregation module, representing different quantities of tokens.

[0029] To further implement the above technical solution, the specific content of the local region attention module includes: Input features Normalization process is performed to obtain ,in, , These are the height and width of the low-resolution image, respectively. To extract the number of feature channels; Will According to fixed dimensions Divide into non-overlapping sliding windows, then pass through a fully connected layer to output deformable offsets: ; in Indicates the sampling location characteristics, This indicates a local area. Indicates the predicted offset. This represents a lightweight offset prediction module; The features at the current sampling position are processed by the weight matrix. Get the query vector The deformed neighborhood sampling features are processed by the weight matrix. Obtain the key vector Attention weights are calculated using the softmax normalization function. : ; in, It is an exponential function. Adjusted local sampling location features and belong Offset in different directions; The size is The window; With the weight matrix Obtain the value vector Perform element-wise multiplication. It is the Kronecker product; Features after attention processing via deformable sliding window for: .

[0030] It should be noted that: like Figure 4 As shown, the feature set processed by the multi-scale token aggregation module is fed into the local region attention module to enhance cross-window feature interaction. First, the input features... Normalization process is performed to obtain To stabilize the feature distribution at different scales. Figure 5 The calculation process for the attention of a deformable sliding window is first... According to fixed dimensions The data is divided into non-overlapping sliding windows, and then passed through a fully connected layer to output a deformable offset. The specific calculation formula is as follows: ; in Indicates the sampling location characteristics, This indicates a local area. Indicates the predicted offset. This represents a lightweight offset prediction module. Subsequently, features from the current sampling position are processed by a weight matrix. Get the query vector The deformed neighborhood sampling features are processed by the weight matrix. Obtain the key vector The attention weights are calculated using the softmax normalization function, and the specific calculation formula is as follows: ; in For attention weights, It is an exponential function. Adjusted local sampling location features , belong Offset in different directions. The size is The window. With the weight matrix Obtain the value vector Perform element-wise multiplication. It is the Kronecker product, and the specific calculation formula is as follows: ; in This indicates the features processed through a deformable sliding window attention mechanism. This mechanism not only reduces the number of parameters and computational cost during network training but also preserves more local feature details. Features are stabilized at a scale through a normalization module and updated using an EDFFN module, reducing redundant features while further preserving key features. Finally, residual connections are used to prevent the loss of low-level features.

[0031] To further implement the above technical solution, the deep feature extraction subnetwork also includes a first frequency domain forward feedback network and a second frequency domain forward feedback network, which are respectively connected to the MS-CDTA and the local region attention module, and respectively process the multi-scale features after feature fusion and the local region attention module. The feature stream formed after local attention processing is used as the input features for detail enhancement: After normalizing the input features, we obtain ,in, , These are the height and width of the low-resolution image, respectively. To extract the number of feature channels; Spatial feature preprocessing: Integrating cross-channel information from infrared images through pointwise convolution PConv yields... Subsequently, the local spatial correlation of infrared features is enhanced by deep convolution (DConv), and the features are refined and denoised to obtain... ; After the second PConv adjusts the number of feature channels to the adaptation dimension for frequency domain processing, the preprocessed features are obtained. ; right The spatial feature flattening operation, frequency domain filtering operation, and quantization matrix filtering operation are performed sequentially to obtain the features. ; Finally Features are obtained by converting back to the spatial domain sequence. and to Perform a folding operation to restore the spatial positions in sequence, and obtain... ; Input features are combined with residual connections. By combining these features, we obtain a feature set with enhanced details.

[0032] It should be noted that: This embodiment uses the multi-scale features obtained through feature fusion as input features. like Figure 3 As shown, multi-scale fusion features After normalization again, we get The input to the EDFFN module first preprocesses the spatial features, integrating cross-channel information from the infrared image through pointwise convolution (PConv), and then enhances the local spatial correlation of the infrared features through depthwise convolution (DConv). The specific calculation formula is as follows: ; ; in , The input features are PConv and DConv, respectively. , For bias. , They are PConv's Convolution kernel and DConv Convolution kernel. The activation function is GeLU. Then, element-wise multiplication is performed. The obtained features are refined and denoised, and then processed by a second PConv to adjust the number of feature channels to the adaptation dimension for frequency domain processing, resulting in preprocessed features. After completing the spatial feature preprocessing, the 2D spatial features are converted into a 1D sequence and adapted to the Fourier transform input format to obtain the adapted features. , This is a flattening operation that arranges features according to their spatial location. The flattened features are then subjected to frequency domain filtering, and a Fourier transform is performed on the 1D sequence to obtain the features. , It is a Fourier transform operator that converts the spatial infrared feature distribution into a frequency domain distribution, making it easier to distinguish between detail frequencies and noise frequencies. The features are then filtered through a quantization matrix. The infrared detail features and noise are assigned corresponding high and low weights, among which Is with A learnable matrix of the same dimension is continuously iteratively optimized during training. It is element-wise multiplication. Finally, the filtered frequency features are converted back to spatial sequences to obtain the features. , It is the inverse Fourier transform operator. It also restores the 1D spatial sequence to a 2D feature map. , This is a folding operation that restores the spatial location in sequence. Finally, residual connections are used to combine the input features with the frequency-filtered features, preserving low-level feature details and avoiding gradient explosion. The specific calculation formula is as follows: ; in This is the feature set after processing by the multi-scale compression-diffusion token aggregation module. The EDFFN module reduces a significant amount of computation by shifting the Fourier transform to the lower channel stage, while simultaneously enhancing spatial features to accurately capture local frequency domain details of the image and reduce redundant noise.

[0033] To further implement the above technical solution, the specific content of the image reconstruction subnetwork includes: like Figure 6 As shown, a convolutional layer is first constructed to remove redundant information from the final feature stream output by the deep feature extraction sub-network, unifying the feature dimensions. Then, the output features of the convolutional layer are added to the deep feature extraction sub-network via residual connections to avoid losing low-level information. This indicates element-wise addition. The features are then fed into the reconstruction module, which includes an upsampling module and a convolutional layer. Upsampling uses a pixel shuffle technique to enlarge the feature map size to the target super-resolution factor, while the convolutional layer restores the spatial details of the image. The final reconstructed output is an image with the same size as the true high-resolution infrared image. This result maximizes the recovery of details from the original image and reduces noise and artifacts in the reconstructed image.

[0034] To further implement the above technical solution, a training sub-network is also included for training the aggregated infrared image super-resolution reconstruction network, specifically including the following: Training is supervised using a dual-domain constraint loss function, where the loss function is: ; in, For the total loss, L1 loss in the spatiotemporal domain refers to the difference between the reconstructed image and the real image. For frequency domain L1 loss, the difference between the reconstructed image (transformed from the spatiotemporal domain image to the frequency domain) and the real image is... To reconstruct the image, For real images, Indicates Fourier transform, These are the weighting coefficients for the frequency domain loss. This represents the mean absolute error (MAE).

[0035] It should be noted that: This loss function design calculates the loss in both the spatiotemporal and frequency domains, balancing global image sharpness with local detail. It ensures the overall brightness and structure of the reconstructed image remain consistent with the real image, resulting in a high-quality image reconstruction. In this embodiment, Setting it to 0.1 optimizes high-frequency details while maintaining pixel-level alignment. Through supervised training using this loss function, the network learns how to efficiently reconstruct low-resolution images into high-resolution images while preserving the original details and structure.

[0036] To further implement the above technical solution, the training process specifically includes: Input LR-HR infrared image data pairs of aluminum electrolysis cells, set training parameters, and train the network in a loop: S1. Initialize network parameters; S2. Input the data into the network to obtain the reconstructed values; S3. Calculate the total loss and backpropagate the network parameters; S4. If the training count is reached, stop training; otherwise, repeat S2-S4.

[0037] It should be noted that: During network training, the Adam optimizer is used to calculate and update the gradients of the network weight parameters, while the MultiStepLR learning rate scheduler adjusts the learning rate. Adam can provide an independent adaptive learning rate for each parameter, improving learning efficiency, while MultiStepLR can flexibly decay the learning rate at any training node. The network weights are continuously iteratively optimized during training to minimize loss and achieve optimal infrared image reconstruction.

[0038] The invention will be further illustrated by the following experiments: Experimental setup: During network training, the Adam optimizer was used to calculate and update the gradients of the network weight parameters. The number of iterations was set to 100,000, and the batch size for each iteration was set to 8. The decay rate of Adam's first-order moment estimate was set to 0.9, and the decay rate of the second-order moment estimate was set to 0.99. The initial learning rate was set to 0.0002, and the MultiStepLR learning rate scheduler was used to adjust the learning rate appropriately during training to stabilize the network training.

[0039] The number of RTB residual groups in the network structure is set to 8, the feature dimension is 48, the number of multi-head attention heads in CDTA is set to 8, and the window size of DSSA is set to 7.

[0040] The experimental platform used in this embodiment employs an RTX 4090 GPU with 24GB of video memory. The CPU is a 16 vCPU Intel(R) Xeon(R) Gold 6430. The operating system used is Ubuntu 22.04, PyTorch 2.8.0, and Python 3.10.

[0041] Through experimental quantitative analysis, the performance of a series of image super-resolution algorithms is evaluated. The image reconstruction quality is evaluated by the metrics PSNR and SSIM, while the model complexity is measured by the number of model parameters (Params) and the number of multiply-accumulate operations (MultiAdds).

[0042] The specific comparison methods in this embodiment include RCAN, EDSR, AWSRAN, SwinIR, SPIN, ATD, and the method provided by this invention. By conducting comparative experiments with these algorithms, the performance of this invention in different scenarios can be comprehensively evaluated, further verifying its feasibility and effectiveness.

[0043] The specific experimental results are shown in Tables 1 and 2: Table 1: Comparison of experimental results for different models when the magnification factor is 2 ; As shown in Table 1, when the amplification factor is 2, the PNSR results of this invention on the IND1K and Alum datasets reach 35.41 and 32.96, respectively, and the SSIM results reach 0.9441 and 0.9714, both reaching the highest values. Compared with the second-best performing ATD dataset, this invention improves the PNSR performance by 0.01 on IND1K and remains on par with Alum. For the SSIM metric, it improves by 0.0005 on IND1K and by 0.0002 on Alum.

[0044] Table 2: Comparison of experimental results for different models when the magnification factor is 4 ; When the magnification factor is 4, the reconstruction effect of this invention still achieves the best results, as shown in Table 2. Compared with the second-best performing ATD, PNSR is improved by 0.08 and 0.09 respectively, and SSIM is improved by 0.0007 and 0.0011 respectively. Meanwhile, compared with ATD, the present invention has approximately 10% less Params and approximately 12.5% ​​less MultiAdds, demonstrating significant superiority in balancing computational resources and model performance.

[0045] This invention proposes a multi-scale token-aggregated infrared image super-resolution reconstruction network for aluminum electrolytic cells, constructing an end-to-end efficient infrared image reconstruction architecture. This improves the reconstruction performance of low-resolution infrared images of aluminum electrolytic cells. Through joint training using both spatiotemporal and frequency domain losses, this training method enhances the network's ability to learn image data features, improving both the network's training effect and image reconstruction performance.

[0046] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A polymeric infrared image super-resolution reconstruction network for aluminum electrolytic cells, characterized in that, include: Shallow feature extraction subnetwork, deep feature extraction subnetwork, and image reconstruction subnetwork; A shallow feature extraction subnetwork is used to extract shallow features from the preprocessed infrared image to obtain shallow features. ; The deep feature extraction subnetwork consists of n interconnected residual blocks, and the shallow feature extraction subnetwork consists of n interconnected residual blocks. After passing through n residual blocks, the final feature stream is output. Each residual block includes a multi-scale compression-diffusion token aggregation module (MS-CDTA) and a local region attention module. MS-CDTA generates token representations of different granularities based on the input features through multi-scale adaptive pooling and low-dimensional orthogonal subspace projection. It also introduces a self-attention mechanism to compress, refine, and diffuse the token representations of different granularities, outputting feature information at different scales. After feature fusion and detail enhancement, the feature set is output. Local attention module Local attention processing is performed to form a feature flow; An image reconstruction subnetwork is used to reconstruct an image result with the same size as the real high-resolution infrared image based on the final feature flow.

2. The polymeric infrared image super-resolution reconstruction network for aluminum electrolytic cells according to claim 1, characterized in that, The specific contents of preprocessing include: Median filtering was used to remove thermal noise and salt-and-pepper noise from infrared images; The image is normalized to map the pixel values ​​to the [0,1] range, thus eliminating the magnitude difference between different images; Data augmentation is performed by randomly flipping the registered low-resolution LR and high-resolution HR image pairs horizontally and vertically. The training set, validation set, and test set are divided proportionally, and the images in the test set are ensured to cover all operating conditions of the aluminum electrolysis cell.

3. The polymeric infrared image super-resolution reconstruction network for aluminum electrolytic cells according to claim 1, characterized in that, The shallow feature extraction subnetwork includes the following: The preprocessed infrared image data is input into a shallow feature extraction subnetwork, which extracts shallow features of the image through convolution operations. ,in, , These are the height and width of the low-resolution image, respectively. To extract the number of feature channels.

4. The polymeric infrared image super-resolution reconstruction network for aluminum electrolytic cells according to claim 1, characterized in that, MS-CDTA includes multiple compressed diffusion token aggregation modules (CDTAs) at different scales. Each CDTA specifically includes: a token representing an initialization unit, which specifically includes the following: CDTA acquires the features after normalizing the input features. , will feature Features projected onto a set of low-dimensional orthogonal subspaces. Dimensions The number of channels is , The deep feature extraction subnetwork represents the first... One residual block, Indicates the first The first residual block A subspace of attention head projection; Through average pooling layer Dimensional reduction Each token represents , , The three scales are represented, and the specific tokens for the three scales are 16, 32, and 64 respectively. Through cross attention pairs Perform iterative optimization, for and Perform a linear transformation. As a query token vector , As a key token vector Sum Value Token Vector ; pass Generates a weight matrix for cross-attention by interacting with global information. To measure Input Token Pair The degree of contribution; weight matrix and Multiplication and addition The final token is obtained by adding them together. : ； ； in, For learnable parameters, This indicates element-wise addition.

5. The polymeric infrared image super-resolution reconstruction network for aluminum electrolytic cells according to claim 4, characterized in that, Each CDTA also specifically includes: Tokens representing compression, refining, and diffusion units, which specifically include the following: right Perform a linear transformation and use it as the query token vector. Token vector Sum Value Token Vector By introducing a self-attention mechanism, token representations of different granularities are compressed, refined, and diffused. The final feature information output by CDTA is... for: ； in, Represents the subspace of all attention head projections. For the compression process, and Multiplication is a diffusion process, through Left multiplication projects the output of CDTA back into higher-dimensional space. It should remain consistent with the input features.

6. The polymeric infrared image super-resolution reconstruction network for aluminum electrolytic cells according to claim 1, characterized in that, MS-CDTA also includes: a feature fusion module. The feature information output by all CDTA modules is used as input features for the feature fusion module. The input features are weighted and fused, and the fused multi-scale features are added to the input features through residual connections to avoid the loss of low-level feature information. ； in, For normalization processing, The quantity represented by the token. This represents three scales, with the corresponding token quantities being 16, 32, and 64, respectively. Learnable weight coefficients for CDTA representing different token quantities; GeLU is the activation function. The feature information output by CDTA for different tokens representing a certain number of tokens.

7. The polymeric infrared image super-resolution reconstruction network for aluminum electrolytic cells according to claim 1, characterized in that, The specific content of the local region attention module includes: Input features Normalization process is performed to obtain ,in, , These are the height and width of the low-resolution image, respectively. To extract the number of feature channels; Will According to fixed dimensions Divide into non-overlapping sliding windows, then pass through a fully connected layer to output deformable offsets: ； in Indicates the sampling location characteristics, This indicates a local area. Indicates the predicted offset. This represents a lightweight offset prediction module; The features at the current sampling position are processed by the weight matrix. Get the query vector The deformed neighborhood sampling features are processed by the weight matrix Obtain the key vector Attention weights are calculated using the softmax normalization function. : ； in, It is an exponential function. Adjusted local sampling location features and belong Offset in different directions; The size is The window; With the weight matrix Obtain the value vector Perform element-wise multiplication. It is the Kronecker product; Features after attention processing via deformable sliding window for: 。 8. The polymeric infrared image super-resolution reconstruction network for aluminum electrolytic cells according to claim 1, characterized in that, The deep feature extraction subnetwork also includes a first frequency domain feedforward network and a second frequency domain feedforward network, which are connected to the MS-CDTA and the local region attention module, respectively. These networks respectively integrate the multi-scale features after feature fusion and the local region attention module. The feature stream formed after local attention processing is used as the input features for detail enhancement: After normalizing the input features, we obtain... ,in, , These are the height and width of the low-resolution image, respectively. To extract the number of feature channels; Spatial feature preprocessing: Integrating cross-channel information from infrared images through pointwise convolution PConv yields... Subsequently, the local spatial correlation of infrared features is enhanced by deep convolution (DConv), and the features are refined and denoised to obtain... ; After the second PConv adjusts the number of feature channels to the adaptation dimension for frequency domain processing, the preprocessed features are obtained. ; right The spatial feature flattening operation, frequency domain filtering operation, and quantization matrix filtering operation are performed sequentially to obtain the features. ; Finally Features are obtained by converting back to the spatial domain sequence. and to Perform a folding operation to restore the spatial positions in sequence, and obtain... ; Input features are combined with residual connections. By combining these features, we obtain a feature set with enhanced details.

9. The polymeric infrared image super-resolution reconstruction network for aluminum electrolytic cells according to claim 1, characterized in that, The specific content of the image reconstruction subnetwork includes: Construct a convolutional layer to remove redundant information from the final feature stream output by the deep feature extraction subnetwork and unify the feature dimensions; The output features of the convolutional layer are added to the deep feature extraction subnetwork through residual connections. The added features are then upsampled and input into the convolutional layer. Upsampling enlarges the size of the feature map to the target super-resolution factor. The convolutional layer restores the spatial details of the image, and finally reconstructs the output image with the same size as the real high-resolution infrared image.

10. The polymeric infrared image super-resolution reconstruction network for aluminum electrolytic cells according to claim 1, characterized in that, It also includes training a sub-network for training the aggregated infrared image super-resolution reconstruction network, specifically including the following: Training is supervised using a dual-domain constraint loss function, where the loss function is: ； in, For the total loss, L1 loss in the spatiotemporal domain refers to the difference between the reconstructed image and the real image. For frequency domain L1 loss, the difference between the reconstructed image (transformed from the spatiotemporal domain image to the frequency domain) and the real image is... To reconstruct the image, For real images, Indicates Fourier transform, These are the weighting coefficients for the frequency domain loss. This represents the mean absolute error (MAE).

11. The polymeric infrared image super-resolution reconstruction network for aluminum electrolytic cells according to claim 10, characterized in that, The training process specifically includes: Input LR-HR infrared image data pairs of aluminum electrolysis cells, set training parameters, and train the network in a loop: S1. Initialize network parameters; S2. Input the data into the network to obtain the reconstructed values; S3. Calculate the total loss and backpropagate the network parameters; S4. If the training count is reached, stop training; otherwise, repeat S2-S4.

Images