Blast furnace top infrared image restoration method, system, medium and equipment

Abstract

The invention relates to the technical field of infrared image processing, and discloses a blast furnace top infrared image restoration method, system, medium and equipment, and the method comprises the following steps: obtaining furnace top infrared images when a blast furnace does not work and works as clear images and noise images respectively, carrying out image pairing, and establishing image pairs; a blast furnace top infrared image restoration model is constructed, an encoder of the model comprises at least one layer of infrared enhancement module, and position coding information of the blast furnace top image is fused into each layer of infrared enhancement module; the infrared enhancement module comprises an embedded frequency domain feature enhanced self-attention module for extracting features of different frequencies, a deformable convolution-based discrete wavelet transform module for extracting local features and multi-scale features, and a spatial detail enhancement module for optimizing non-uniform features; and training the model, and inputting the blast furnace top infrared image to be restored into the trained model to obtain a restored image. According to the method, complex noise can be processed, the non-uniform enhancement problem is effectively solved, and the restoration precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of infrared image processing technology, and in particular to a blast furnace top infrared image restoration method, system, medium and equipment. Background Art

[0002] Blast furnace ironmaking, a core process in the modern steel industry, accounts for over 95% of global pig iron production. Infrared images are formed by a thermal infrared scanner receiving and recording the thermal radiation energy emitted by a target object. High-quality infrared images provide data support for downstream computer vision-based inspections, such as chute angle measurement, improving the accuracy and responsiveness of production decisions. Therefore, infrared images of the blast furnace roof are crucial for monitoring blast furnace conditions.

[0003] Infrared images of blast furnace interiors provide real-time visualization of charge distribution, lining wear, and chute trajectory, serving as a crucial basis for on-site workers to adjust blast furnace operations. However, due to the harsh conditions within the blast furnace, during the charge-feeding phase, the coordinated operation of the top charge-feeding system and the bottom hot air delivery system generates intense glare from smoke, dust, vapor, and complex heat sources, resulting in severe non-uniform noise contamination in the captured infrared images. Noisy infrared images can significantly impact image recognition accuracy. For example, using the Res-LSTM model for chute angle detection, while its recognition accuracy reaches 99.79% for clear images, it plummets to 54.13% for noisy images. Therefore, to effectively utilize infrared images within the blast furnace, it is necessary to denoise and enhance the quality of unclear blast furnace top infrared images.

[0004] At present, there are mainly the following infrared image enhancement and restoration methods in industrial scenarios:

[0005] (1) Traditional infrared image enhancement methods. Typical methods include: Contrast Limited Adaptive Histgram Equalization (CLAHE) based on grayscale mapping, frequency domain filtering based on transform domain, and wavelet enhancement algorithm. These methods generate enhanced images by directly operating on image pixel values or calculating transform coefficients in a specific domain and then converting them back to the original spatial domain. However, under the interference of high dynamic noise (such as smoke, water vapor, and glare inside a blast furnace), problems such as edge blurring, artifact amplification, and non-uniform enhancement are prone to occur.

[0006] (2) Infrared image enhancement methods based on deep learning. Typical methods include: low-light enhancement to improve image quality through multi-scale feature fusion and illumination perception module, improving image resolution by combining perceptual loss and adversarial training, and achieving super-resolution reconstruction of high-frequency details. These methods obtain a large amount of historical data through sensors and use end-to-end neural networks to learn the mapping relationship from noisy images to clear images. Although they can improve the contrast of infrared images, they lack the ability to process complex noise, and are prone to noise enhancement amplification during super-resolution reconstruction and low-light enhancement.

[0007] (3) Visible light dehazing cross-domain migration method: Typical methods include: dehazing and restoration methods based on the atmospheric scattering model (ASM), dehazing methods based on deep learning architectures such as convolutional neural networks (CNN), generative adversarial networks (GAN) and transformers, and dehazing methods based on data-driven enhancement strategies. This method draws on the visible light image dehazing and restoration method, migrates to the infrared image restoration task through architecture improvement and domain adaptation, reduces the dataset distribution bias through preprocessing transformation, and improves the model generalization by combining cross-domain visual matching. However, since infrared images are different from visible light images and are pseudo-color single-channel images, they lack color texture information, have low contrast and clarity, and the environment in which infrared images are formed naturally has complex lighting conditions. Therefore, the fundamental difference in data modality makes it impossible for dehazing algorithms based on ASM and deep neural networks to adapt to the infrared image restoration problem. Summary of the Invention

[0008] To this end, the technical problem to be solved by the present invention is to overcome the deficiencies in the prior art and provide a blast furnace top infrared image restoration method, system, medium and equipment, which can realize the processing of complex noise in blast furnace infrared images, effectively solve the problem of non-uniform enhancement, and improve the image restoration accuracy.

[0009] To solve the above technical problems, the present invention provides a blast furnace top infrared image restoration method, comprising:

[0010] Obtain an infrared image of the furnace top when the blast furnace is not working as a clear image, and an infrared image of the furnace top when the blast furnace is working as a noise image, pair the clear image and the noise image, and establish an image pair of the clear image and the noise image as a data set;

[0011] Constructing a blast furnace top infrared image restoration model, wherein the encoder of the blast furnace top infrared image restoration model includes at least one layer of infrared enhancement modules, each layer of the infrared enhancement modules incorporates position coding information of the blast furnace top image; the infrared enhancement module includes a self-attention module embedded in frequency domain feature enhancement, a discrete wavelet transform module based on deformable convolution, and a spatial detail enhancement module, wherein the self-attention module embedded in frequency domain feature enhancement extracts features of different frequencies, the discrete wavelet transform module based on deformable convolution extracts local features and multi-scale features, and the spatial detail enhancement module optimizes non-uniform features;

[0012] The blast furnace top infrared image restoration model is trained using the data set, and the infrared image of the blast furnace top to be restored is input into the trained blast furnace top infrared image restoration model to obtain a restored image.

[0013] Furthermore, each layer of the infrared enhancement module is integrated into the position coding information of the blast furnace top image, specifically:

[0014] Segment the pixels of the furnace top infrared image and calculate the centroid of the segmented pixels; divide the furnace top infrared image into blocks and determine the block number p corresponding to the centroid;

[0015] Get the M domain image blocks of the pth image block, denoted as N M (p), the sine and cosine functions are used to position encode the image blocks. The position encoding rule is:

[0016]

[0017] Where W and H represent the width and height of the furnace top infrared image respectively, x b and y b are the horizontal and vertical coordinates of the center of mass, N M+1 (p) represents the area covered by the centroid and its M domain image blocks, i represents the dimension index, and pos represents N M+1 Any feature point in (p), PE(pos,i) represents N M+1 The position encoding information of the i-th dimension corresponding to the pos-th feature point in (p); β represents the frequency coefficient, C represents the dimension size of the feature; DC represents the even dimension index, and SC represents the odd dimension index;

[0018] The features output by the infrared enhancement module of the previous layer are added to the infrared image of the furnace top encoded according to the position encoding rule as the input of the infrared enhancement module of the next layer.

[0019] Furthermore, the self-attention module embedded in the frequency domain feature enhancement extracts features of different frequencies, specifically:

[0020] The self-attention module embedded in the frequency domain feature enhancement includes a frequency domain enhancement module and an attention module. The input feature of the self-attention module embedded in the frequency domain feature enhancement is denoted as x, and the input feature is converted into amplitude and phase through Fourier transform, denoted as F(x), where F() represents Fourier transform;

[0021] F(x) is passed to the frequency domain enhancement module for frequency domain enhancement processing. The frequency domain enhancement module includes: a first layer of depthwise separable convolution layer, a second layer to calculate the mean and maximum value of F(x), a third layer of depthwise separable convolution layer, and then a Sigmoid activation function is used to obtain a normalized attention map, denoted as α; F(x) is multiplied by α to obtain a weighted feature vector, denoted as f′; f′ is subjected to inverse Fourier transform and then enters the attention module;

[0022] The query, key, and value obtained by the feature map after inverse Fourier transform are denoted as Q f , K f 、V f , the final output feature is obtained by adding the residual connection to the original input:

[0023] fs=SA(Q f +K f +V f )+x, where fs is the final output feature of the self-attention module embedded with frequency domain feature enhancement, and SA() is the window self-attention calculation.

[0024] Furthermore, the discrete wavelet transform module based on deformable convolution extracts local features and multi-scale features, specifically:

[0025] The output feature of the self-attention module embedded in the frequency domain feature enhancement is denoted as fs, and fs obtains feature maps of different scales through two deformable convolutions of different scales, denoted as fs1 and fs2, and the scale feature of fs1 is smaller than that of fs2;

[0026] Transform fs1 into four different wavelet subbands through wavelet transform, and the feature vector obtained by splicing the four subbands along the channel dimension is recorded as fs wavelet ;

[0027] After the feature modulation module, fs wavelet As a modulation parameter applied to fs2, the large-scale feature with local detail enhancement is obtained, recorded as fs2′, fs2′=β′·fs2, β′ represents the modulation parameter;

[0028] fs wavelet The output feature obtained by combining the convolution operation with fs1 and fs2′ is recorded as fs fusion , fs fusion The calculation method is:

[0029] fs fusion =τ⊙[fs1,fs wavelet ]+fs2′,

[0030] Where ⊙ represents the Hadamard product, τ represents the local dynamic range coefficient;

[0031] The local dynamic range coefficient is calculated as:

[0032] τ=Tanh(LocalMaxMin(fs2)),

[0033] Among them, τ represents the local dynamic range coefficient, LocalMaxMin() means calculating the difference between the maximum pixel and the minimum pixel in the local window, and Tanh() means limiting the value to [-1,1].

[0034] Furthermore, the spatial detail enhancement module optimizes the non-uniform features, specifically:

[0035] The output feature of the self-attention module embedded in the frequency domain feature enhancement is denoted as fs; after being processed by the spatial detail enhancement module, fs is combined with the input fs to obtain the output feature fs′ with edge high-frequency feature enhancement;

[0036] The spatial detail enhancement module includes at least one detail enhancement convolution integrating different differential convolutions, and the weight of the detail enhancement convolution is:

[0037] w DEConv =Scale(w CDC +w RDC )+Softmax(w HDC +w VDC +w ADC ), where w DEConv The weight of detail enhancement convolution, w CDC 、w RDC 、w HDC 、w VDC 、w ADC They are the weights of the center difference convolution, the weights of the rotation difference convolution, the weights of the horizontal difference convolution, the weights of the vertical difference convolution, and the weights of the diagonal difference convolution; Scale is a trainable scaling factor, and Softmax() is the activation function.

[0038] Furthermore, the output of the infrared enhancement module is:

[0039]

[0040] in, is the output vector of the infrared enhancement module, fs is the output feature of the self-attention module embedded in the frequency domain feature enhancement, and fs fusion is the output feature of the discrete wavelet transform module based on deformable convolution, fs′ is the output feature of the spatial detail enhancement module; g is the weighting coefficient, and the calculation method of g is:

[0041] g=sigmoid(MLP([fs fusion ,fs′])),

[0042] Among them, MLP() is the gating mechanism operation and sigmoid() is the activation function.

[0043] Furthermore, when training the blast furnace top infrared image restoration model, the total loss function is:

[0044] L = α′L combined +β″L edge +γL local ,

[0045] Among them, L is the total loss function, L combined is the global loss function, L edge is the marginal loss function, L local is the local loss function of the key area; α′, β″, γ are control parameters, α′+β″+γ=1;

[0046] The calculation method of the global loss function is:

[0047] L combined =ρ1·(1-SSIM(O i ,T i ))+ρ2·MSE(O i ,T i ),

[0048] Among them, ρ1 and ρ2 are weight coefficients, SSIM() is the structural similarity loss function, and MSE() is the mean square error loss function; i 、T i Respectively represent the output i-th restored image and i-th real clear image;

[0049] The calculation method of the edge loss function is:

[0050] L edge =∑|E(O i )-E(T i )|,

[0051] Where, E() represents the edge extraction operation;

[0052] The calculation method of the local loss function of the key area is:

[0053] L local =∑W(x,y)|O(x,y)-T(x,y)|,

[0054] Among them, W(x,y) is the Gaussian weight. The closer to the center of the key area, the higher the Gaussian weight. The key area is the area where the various parts of the blast furnace body are located in the infrared image of the blast furnace top; O(x,y) represents the pixel value output by the model at the coordinate (x,y), and T(x,y) represents the target value at the coordinate (x,y).

[0055] The present invention also provides a blast furnace top infrared image restoration system, comprising:

[0056] A data acquisition module is used to acquire an infrared image of the furnace top when the blast furnace is not working as a clear image and an infrared image of the furnace top when the blast furnace is working as a noise image, pair the clear image and the noise image, and establish an image pair of the clear image and the noise image as a data set;

[0057] A model construction module is used to construct an infrared image restoration model of a blast furnace top. The encoder of the infrared image restoration model of the blast furnace top includes at least one layer of infrared enhancement modules, and each layer of the infrared enhancement modules incorporates position coding information of the blast furnace top image; the infrared enhancement module includes a self-attention module embedded in frequency domain feature enhancement, a discrete wavelet transform module based on deformable convolution, and a spatial detail enhancement module. The self-attention module embedded in frequency domain feature enhancement extracts features of different frequencies, the discrete wavelet transform module based on deformable convolution extracts local features and multi-scale features, and the spatial detail enhancement module optimizes non-uniform features.

[0058] A model training module, configured to train the blast furnace top infrared image restoration model using a data set;

[0059] The image restoration module is used to input the infrared image of the blast furnace top to be repaired into the trained blast furnace top infrared image restoration model to obtain a restored image.

[0060] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the blast furnace top infrared image restoration method is implemented.

[0061] The present invention also provides a blast furnace top infrared image restoration device, comprising a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the blast furnace top infrared image restoration method is implemented.

[0062] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0063] The present invention improves the reconstruction quality of the blast furnace top image by incorporating the position coding information of the image; improves the model's ability to process different frequency features, especially the coupled noise of different frequencies in the image, by embedding a self-attention module for frequency domain feature enhancement; enhances the processing capabilities of local features and multi-scale features by adopting a discrete wavelet transform module based on deformable convolution, and is suitable for the reconstruction of targets with variable shapes in blast furnace images; optimizes important features such as edges through a spatial detail enhancement module, effectively solving problems such as edge blur, artifact amplification and non-uniform enhancement; realizes multi-scale encoding and decoding through a multi-layer infrared enhancement module, and improves the ability to gradually learn the complex feature mapping relationship in clear images and noisy images, thereby effectively processing complex noise and improving the accuracy of blast furnace top infrared image restoration. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0065] Figure 1 Flowchart of the method in the preferred embodiment of the present invention.

[0066] Figure 2 1 is an example diagram of an image pair of a clear image and a noisy image constructed in a preferred embodiment of the present invention.

[0067] Figure 3 This is the overall architecture diagram of the blast furnace top infrared image restoration model in the preferred embodiment of the present invention.

[0068] Figure 4 This is a structural diagram of the encoder (decoder) in a preferred embodiment of the present invention.

[0069] Figure 5 Schematic diagram of the structure of the self-attention module embedded with frequency domain feature enhancement in a preferred embodiment of the present invention.

[0070] Figure 6 Schematic diagram of the structure of the frequency domain enhancement module in the self-attention module embedded with frequency domain feature enhancement in a preferred embodiment of the present invention.

[0071] Figure 7 Schematic diagram of the structure of a discrete wavelet transform module based on deformable convolution in a preferred embodiment of the present invention.

[0072] Figure 8 Schematic diagram of the structure of the spatial detail enhancement module in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0073] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0074] Reference Figure 1 As shown, the present invention discloses a blast furnace top infrared image restoration method, comprising the following steps:

[0075] S1: Obtain an infrared image of the furnace top when the blast furnace is not working as a clear image, and an infrared image of the furnace top when the blast furnace is working as a noise image, pair the clear image and the noise image, and establish an image pair of the clear image and the noise image as a dataset.

[0076] In this example, the chute inside a blast furnace is used as a reference. During the charging phase of the blast furnace, the top charging system generates intense smoke, dust, and water vapor, which results in noise pollution in the captured infrared image. In contrast, during the interval between charging, the internal operation of the blast furnace is stable, and the captured image is clear. Therefore, in this example, images are continuously captured during the charging interval and during charging as clear images and noise images. Figure 2 As shown, excerpts of chute images in three different groups are selected to construct three image pairs: image pair 1, image pair 2, and image pair 3.

[0077] S2: Build Figure 3 The blast furnace top infrared image restoration model shown in FIG. has a backbone network of an improved U-Net architecture. The overall architecture is a three-layer U-Net structure, and the model accepts multi-scale input. The core of the blast furnace top infrared image restoration model is an encoder (decoder). The encoder of the blast furnace top infrared image restoration model includes at least one layer of infrared enhancement modules, and each layer of the infrared enhancement modules incorporates the position coding information of the blast furnace top image. In this embodiment, the encoder includes the following: Figure 4 The three-layer infrared enhancement module shown, each layer incorporates the position coding information of the blast furnace chute guide, which can provide the model with more discriminative expert information of the blast furnace top image.

[0078] Each of the infrared enhancement modules includes three identical modules: a self-attention module embedded in frequency domain feature enhancement (represented by FFE-SA), a discrete wavelet transform module based on deformable convolution (represented by DWT), and a spatial detail enhancement module (represented by SDE). The self-attention module embedded in frequency domain feature enhancement extracts features of different frequencies, especially coupled noise of different frequencies in the image; the discrete wavelet transform module based on deformable convolution extracts local features and multi-scale features, which is suitable for the reconstruction of targets with variable shapes in blast furnace images; the spatial detail enhancement module optimizes important non-uniform features such as edges, which can effectively solve the problems of edge blur and artifact amplification.

[0079] S2-1: The chute is the main moving part of the blast furnace top, and the chute rotates at a constant angular velocity, that is, the time it takes to rotate one circle is fixed. Taking the chute rotation period of 8 seconds / week and the video frame rate of 50 frames / second as an example, 400 frames of images of the chute in different positions can be obtained when the chute rotates one circle, that is, the chute runs to the same position every 400 frames of continuous video recording. When the chute rotates to the same position, its chute shape, relative position, occlusion relationship, etc. in the furnace top infrared image have very strong commonalities. Based on the above observations, by constructing the position coding information of the blast furnace chute guide, more discriminative blast furnace top image expert information can be provided to the model. The same characteristics are also possessed when selecting other parts in the blast furnace top as reference objects. Therefore, the image input into the blast furnace top infrared image restoration model in the present invention is first integrated into the position coding information of the blast furnace top image before entering each layer of infrared enhancement module, specifically:

[0080] The pixels of the chute in the furnace top infrared image are segmented by image segmentation and other methods, and the centroid of the segmented pixels is calculated and recorded as the centroid of the chute (x b ,y b ), x b and y b are the horizontal and vertical coordinates of the centroid; divide the furnace top infrared image into blocks and determine the block number p corresponding to the centroid of the chute.

[0081] Since the input is a fuzzy image, the segmentation result is inaccurate, so the M domain image blocks of the p-th image block are obtained, denoted as N M (p), in this embodiment, eight neighborhood image blocks N8(p) are obtained; the image blocks are position-coded using sine and cosine functions, and the position coding rule is:

[0082]

[0083] Where W and H represent the width and height of the furnace top infrared image respectively, x b / W,y b / H represents the weight of the encoding position using the normalized result of the chute centroid, which is intended to provide unique discriminant information related to the chute centroid; sin() and cos() represent the sine function and cosine function, respectively. M+1 (p) represents the area covered by the centroid of the chute and its M domain image blocks (i.e., the set of pixels covered on the original image, or the set of feature points covered at corresponding positions on the feature map). In this embodiment, it is the area N9(p) covered by the p-th image block and the eight neighborhood image blocks; i represents the dimension index, and pos represents the covered area N M+1 Any feature point in (p), PE(pos,i) represents N M+1The position encoding information of the i-th dimension corresponding to the pos-th feature point in (p); β represents the frequency coefficient, C represents the dimension size of the feature; DC represents the even dimension index, and SC represents the odd dimension index.

[0084] like Figure 3 As shown in the figure, before entering each layer of infrared enhancement module for encoding, the position encoding information PE of the blast furnace chute guide is added, where the size of PE is the same as the original image or the corresponding feature map, and each element in PE is calculated according to the position encoding rule. That is, the features output by the infrared enhancement module of the previous layer are added to the infrared image of the furnace top encoded according to the position encoding rule of the blast furnace chute guide, and used as the input of the infrared enhancement module of the next layer to achieve information fusion. It should be emphasized that although the feature maps of the infrared enhancement modules at different layers are of different sizes, according to the principle of convolution position invariance, the position of the chute position on the feature map input by the corresponding encoder layer can be determined according to the downsampling specifications executed by the encoder, thereby obtaining the block number corresponding to the chute position and its eight-neighborhood number. Substituting them into the above encoding formula can obtain the position encoding result of the corresponding layer. The feature after adding the position encoding information fusion is recorded as x and enters the infrared enhancement module.

[0085] S2-2: After adding the position encoding information fusion, x enters the self-attention module embedded in the frequency domain feature enhancement. The self-attention module embedded in the frequency domain feature enhancement extracts features of different frequencies. The specific process is as follows:

[0086] S2-2-1: The structure of FFE-SA is as follows: Figure 5 As shown in the figure, in FFE-SA, the core is the frequency domain enhancement module (denoted by FFE) embedded in the self-attention module. The structure of FFE is as follows Figure 6 As shown. The self-attention module embedded in frequency domain feature enhancement includes a frequency domain enhancement module and an attention module. The input image of the self-attention module embedded in frequency domain feature enhancement is denoted as x. The input image is converted into amplitude and phase through Fourier transform and denoted as F(x), where F() represents Fourier transform.

[0087] S2-2-2: F(x) is passed into the frequency domain enhancement module for frequency domain enhancement processing. The frequency domain enhancement module includes: the first layer of depthwise separable convolution (DSConv) layer, which consists of 3×3 convolution, ReLU activation function and 1×1 convolution; the second layer obtains the mean and maximum value of F(x) respectively, and the third layer is also a depthwise separable convolution (DSConv) layer for DSConv processing, and then a normalized attention map is obtained through the Sigmoid activation function layer, denoted as α; F(x) is multiplied by α to obtain a weighted feature vector, denoted as f′, to enhance important features in the frequency domain and suppress interference features such as noise; f′ enters the attention module after inverse Fourier transform.

[0088] S2-2-3: Let the feature after inverse Fourier transform be x f , x f =F -1 (α·F(x)), x f The query, key, and value obtained by mapping are respectively denoted as Q f , K f 、V f , the final output feature is obtained by adding the residual connection to the original input:

[0089] fs=SA(Q f +K f +V f )+x, where fs is the final output feature of the self-attention module embedded with frequency domain feature enhancement, and SA() is the window self-attention calculation.

[0090] S2-3: The output features of the self-attention module embedded with frequency domain feature enhancement are fed into the discrete wavelet transform module based on deformable convolution to extract local features and multi-scale features. Specifically:

[0091] S2-3-1: The structure of the DWT module is as follows Figure 7 As shown, fs first obtains feature maps of different scales through two deformable convolutions of different scales, denoted as fs1 and fs2, and the scale feature of fs1 is smaller than that of fs2; in this embodiment, specifically: fs1 = DConv 3×3 (fs),fs2=DConv 7×7 (fs), DConv 3×3 () represents a deformable convolution operation with a convolution kernel size of 3×3, DConv 7×7 () represents a deformable convolution operation with a convolution kernel size of 7×7.

[0092] S2-3-2: Transform the small-scale feature map fs1 into four different wavelet sub-bands by wavelet transform, specifically using Haar wavelet basis function in this embodiment, and transform the low-frequency approximate sub-band fs LL , horizontal high frequency subband fs LH , vertical high frequency sub-band fs HL and diagonal high frequency subband fs HH Splicing along the channel dimension, the feature vector obtained by splicing four sub-bands is recorded as fs wavelet ;fs wavelet =Concat(fs LL +fs HH +fs LH +fs HL ), fs LL 、fs HH、fs LH 、fs HL They are four different subbands obtained by wavelet transform of fs1, and Concat() represents the vector concatenation operation.

[0093] S2-3-3: Small-scale features fs wavelet The modulation parameter β′ is obtained through the feature modulation module, and β′ is applied to the large-scale feature fs2 to obtain the large-scale feature with local detail enhancement, recorded as fs2′, fs2′=β′·fs2, β′ represents the modulation parameter;

[0094] S2-3-4: fs wavelet The output feature obtained by combining the 1×1 convolution operation with the small-scale fs1 and fs2′ is recorded as fs fusion , fs fusion The calculation of the local dynamic range coefficient is used to enhance the highlight area, s fusion The calculation method is:

[0095] fs fusion =τ⊙[fs1,fs wavelet ]+fs2′,

[0096] Where ⊙ represents the Hadamard product, τ represents the local dynamic range coefficient;

[0097] The local dynamic range coefficient is calculated as:

[0098] τ=Tanh(LocalMaxMin(fs2)),

[0099] Among them, τ represents the local dynamic range coefficient, LocalMaxMin() means calculating the difference between the maximum pixel and the minimum pixel in the local window to highlight the highlight and shadow areas; Tanh() means limiting the value to [-1,1] to avoid over-enhancement.

[0100] Because the wavelet transform has excellent localization properties, the introduction of the DWT module gives the model stronger ability to process local detail features. Furthermore, the DWT module incorporates multi-scale feature processing to adapt to the constantly changing scales of objects in blast furnace roof images, achieving better reconstruction results. Furthermore, after denoising, to avoid excessively low contrast in the smoothed image, a local dynamic range factor, τ, is introduced to enhance the highlight areas of the chute, dynamically balancing detail preservation between the highlight areas (chute surface) and the dark areas (background).

[0101] S2-4: The output features of the self-attention module embedded in the frequency domain feature enhancement are fed into the spatial detail enhancement module to optimize the non-uniform features. This step and S2-3 are in no particular order. Specifically:

[0102] In SDE, in order to reconstruct high-frequency features (such as edges, contours, etc.) that are of great significance in the image, this embodiment introduces detail enhancement convolution (denoted by DEConv) that integrates different differential convolutions. Differential convolution combines the ability of traditional edge detection operators to extract high-frequency features such as edges and the powerful feature learning ability of convolutional networks. It can better focus on the gradient information of the image and improve the quality of the model's reconstruction of high-frequency edge features. In addition, the edges of the chute are prone to gradient diffusion due to motion blur, and the axial (horizontal / vertical / diagonal) edges need to be strengthened. The shape and size of the chute changes continuously during movement, and multi-directional gradient features need to be retained. Therefore, an axial edge weight priority strategy is designed in this embodiment.

[0103] like Figure 8 As shown, the spatial detail enhancement module includes at least one detail enhancement convolution integrating different differential convolutions, and the weight of the detail enhancement convolution is:

[0104] w DEConv =Scale(w CDC +w RDC )+Softmax(w HDC +w VDC +w ADC ), where w DEConv is the weight of detail enhancement convolution, CDC, RDC, HDC, VDC, ADC represent five different differential convolutions: center difference, rotation difference, horizontal difference, vertical difference, and diagonal difference, respectively. CDC 、w RDC 、w HDC 、w VDC 、w ADC They are the weights of the center difference convolution, the weights of the rotation difference convolution, the weights of the horizontal difference convolution, the weights of the vertical difference convolution, and the weights of the diagonal difference convolution; Scale is a trainable scaling factor, and Softmax() is the activation function.

[0105] In this embodiment, the detail enhancement convolution includes two detail enhancement convolutions integrating different differential convolutions and a ReLU activation function. After fs is processed by the spatial detail enhancement module, it is combined with the original input feature fs passed through the residual connection to obtain the output feature fs′ with enhanced edge high-frequency features. The calculation formula of fs′ is:

[0106] fs′=fs+DEConv(ReLU(DEConv(fs),

[0107] Among them, ReLU() is the activation function, DEConv() is the detail enhancement convolution, and the weight of the detail enhancement convolution (i.e., the convolution kernel) is w DEConvDEConv prioritizes enhancing horizontal / vertical / diagonal edges (chute contour structure), and combines center / rotation edge enhancement to adapt to the ever-changing contour shape of the chute.

[0108] This process effectively captures the details and edge information in the image and improves the model's ability to process key features.

[0109] S2-5: Finally, in order to improve the model's ability to handle complex features and coupled noise, feature interaction between the wavelet domain and the spatial domain is introduced. Figure 4 As shown in the figure, the input features are processed by FFE-SA to obtain the feature fs. Then the feature fs is processed by DWT and SDE respectively. The feature fs processed by the two modules is fusion Combined with fs′ and weighted by the gated mechanism MLP, combined with the feature fs introduced by the residual connection, the final output vector of the infrared enhancement module is:

[0110]

[0111] in, is the output vector after being encoded by the infrared enhancement module, fs is the output feature of the self-attention module embedded in the frequency domain feature enhancement, and fs fusion is the output feature of the discrete wavelet transform module based on deformable convolution, fs′ is the output feature of the spatial detail enhancement module; g is the weighting coefficient, and the calculation method of g is:

[0112] g=sigmoid(MLP([fs fusion ,fs′])),

[0113] Among them, MLP() is the gating mechanism operation, and sigmoid() is the sigmoid function.

[0114] The encoder in this embodiment includes three layers of infrared enhancement modules, that is, the final prediction result is obtained after executing S2-1 to S2-5 three times.

[0115] S3: Use the dataset to train the blast furnace roof infrared image restoration model. During training, multi-scale outputs are returned as additional supervisory signals, allowing the model to iteratively refine global and local features. During training, a global loss function based on SSIM and MSE is constructed, and model training is implemented using a gradient descent backpropagation strategy. Since the U-net has three output branches of different depths, the loss of each output network is calculated separately, and the constructed total loss function is:

[0116] L = α′L combined +β″L edge +γL local ,

[0117] Among them, L is the total loss function, L combined is the global loss function, L edge is the marginal loss function, L local is the local loss function of the key area; α′, β″, and γ are control parameters used to adjust the proportion between different losses, α′+β″+γ=1;

[0118] The calculation method of the global loss function is:

[0119] L combined =ρ1·(1-SSIM(O i ,T i ))+ρ2·MSE(O i ,T i ),

[0120] Among them, ρ1 and ρ2 are weight coefficients used to balance the influence of the two losses; SSIM() is the structural similarity loss function, which is used to evaluate the structural similarity between the generated image and the target image. The closer its value is to 1, the more similar the two images are; MSE() is the mean square error loss function, which is used to calculate the average square difference between the pixels of the predicted image and the real image, so as to measure the degree of difference between the two. i 、T i They represent the output i-th restored image and the i-th real clear image respectively; by implementing supervision at different depth layers, the model can more accurately learn complex mapping relationships and gradually restore the features in the detailed images between blurred and clear images.

[0121] Given that there is a large amount of important edge information in the blast furnace chute image, edge detection is introduced as an additional supervisory signal to help the model better capture these key features. The calculation method of the edge loss function is:

[0122] L edge =∑|E(O i )-E(T i )|, where E() represents edge extraction operation;

[0123] Since the chute is the only key target in the chute image, and the chute is located in a specific area, a local loss function for the key area is introduced to make the model pay more attention to the changes in these areas, thereby improving the restoration quality of the key parts. The calculation method of the local loss function for the key area is:

[0124] L local =∑W(x,y)|O(x,y)-T(x,y)|,

[0125] Among them, W(x,y) is the Gaussian weight. The closer to the center of the key area, the higher the Gaussian weight. The Gaussian weight in this embodiment satisfies the Gaussian distribution with the center of the key area as the mean and the standard deviation as 1 / 5 of the image width. The key area is the area where the various parts of the blast furnace body are located in the infrared image of the blast furnace top; O(x,y) represents the pixel value output by the model at the coordinate (x,y), and T(x,y) represents the target value at the coordinate (x,y).

[0126] S4: Input the infrared image of the blast furnace roof to be restored into the trained blast furnace roof infrared image restoration model to obtain a restored image. In this embodiment, the collected dataset is divided into a training set and a test set in an 8:2 ratio. The training set is used to train the blast furnace roof infrared image restoration model, and the test set is used to verify the restoration effect of the blast furnace roof infrared image restoration model.

[0127] The invention also discloses a blast furnace top infrared image restoration method system, which includes a data acquisition module, a model construction module, a model training module and an image restoration module.

[0128] The data acquisition module is used to obtain the infrared image of the furnace top when the blast furnace is not working as a clear image and the infrared image of the furnace top when the blast furnace is working as a noise image, pair the clear image and the noise image, and establish the image pair of the clear image and the noise image as a data set.

[0129] The model construction module is used to construct an infrared image restoration model of the blast furnace top. The encoder of the infrared image restoration model of the blast furnace top includes at least one layer of infrared enhancement modules, and each layer of the infrared enhancement module integrates the position coding information of the blast furnace top image; the infrared enhancement module includes a self-attention module embedded in frequency domain feature enhancement, a discrete wavelet transform module based on deformable convolution, and a spatial detail enhancement module. The self-attention module embedded in frequency domain feature enhancement extracts features of different frequencies, the discrete wavelet transform module based on deformable convolution extracts local features and multi-scale features, and the spatial detail enhancement module optimizes non-uniform features.

[0130] The model training module is used to train the blast furnace top infrared image restoration model using the data set.

[0131] The image restoration module is used to input the infrared image of the blast furnace top to be repaired into the trained blast furnace top infrared image restoration model to obtain a restored image.

[0132] The present invention also discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, a method for restoring an infrared image of a blast furnace top is implemented.

[0133] The present invention also discloses a blast furnace top infrared image restoration device, which includes a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, a blast furnace top infrared image restoration method is implemented.

[0134] This paper addresses the issues of low infrared image resolution, insufficient contrast, and a lack of color and texture information. By combining the advantages of infrared enhancement and visible light dehazing methods, a series of infrared enhancement modules are embedded into the dehazing network framework. These infrared enhancement modules specifically enhance the model's ability to extract and process infrared image features. Furthermore, the dehazing network leverages its powerful feature learning capabilities and multi-scale input and output capabilities to gradually learn the complex mapping from noisy images to clear images, completing the task of restoring infrared images of blast furnace roofs.

[0135] This method can denoise and restore infrared noisy images captured during the material distribution process, improving not only the image's resolution and contrast but also its color and texture information, thereby providing high-quality data support for real-time monitoring of blast furnace conditions and subsequent image-based visual tasks. Compared with existing image restoration methods, this method can more effectively address the complex noise inherent in blast furnace roof infrared images, as well as the inherent pseudo-color and low texture diversity of infrared images. Specific advantages include:

[0136] (1) By integrating the position coding information of the blast furnace top image, the image reconstruction quality is improved.

[0137] (2) This invention achieves feature enhancement in the frequency domain by embedding a self-attention module for frequency domain feature enhancement. This improves the model's ability to process features of different frequencies, especially coupled noise of different frequencies in the image, and enhances the model's ability to process global features and long-range dependencies in the image. Furthermore, a windowed self-attention mechanism is used to fuse spatial and frequency domain features.

[0138] (3) By adopting a discrete wavelet transform module based on deformable convolution, the processing capabilities of local features and multi-scale features are enhanced, which is suitable for the reconstruction of targets with variable shapes in blast furnace images.

[0139] (4) The spatial detail enhancement module is used to optimize the processing of important features such as edges, effectively solving problems such as edge blur, artifact amplification, and non-uniform enhancement, and enhancing the model's ability to process important high-frequency detail features.

[0140] (5) The blast furnace top infrared image restoration model accepts multi-scale input features and outputs output features of different scales at different decoding layers of the network, which improves the model's ability to gradually learn the complex feature mapping relationships in clear images and noisy images.

[0141] (6) The infrared enhancement module combines the enhanced features of the frequency domain, wavelet domain, and spatial domain to generate a more comprehensive and richer feature representation, improving the model's ability to handle complex features and coupled noise. By integrating multi-domain feature processing, high-quality images can be restored from low-quality infrared images. It can effectively manage multi-scale complex features and highlight local important features, significantly improving the model's ability to solve infrared image noise problems and achieving high-precision blast furnace roof infrared image restoration.

[0142] To further illustrate the beneficial effects of the present invention, this example collected 2160 images of noise during the distribution process and clear images of the distribution gaps at a blast furnace site. This yielded 2160 image pairs. The training and test sets were divided into 8:2 ratios, resulting in 1728 pairs of training data and 432 pairs of test data. Peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) were used as evaluation metrics. The calculation formulas for PSNR and SSIM are as follows:

[0143]

[0144]

[0145] Among them, MAX I μ represents the maximum possible value of the image pixel (255 for 8-bit grayscale images), and MSE is the mean square error between the original image and the processed image. x and μ y are the mean values of image x and y respectively, and are the variances of images x and y, σ xy is the covariance of images x and y, and C1 and C2 are small constant values introduced to stabilize the divisor.

[0146] The definition of MSE is as follows:

[0147]

[0148] Among them, I(i,j) represents the pixel value of the original image at position (i,j), K(i,j) represents the pixel value of the processed image at the same position (i,j), and m and n are the height and width of the image respectively.

[0149] The proposed method and the existing models PSRGAN, HWMNet, and DehazeFormer were trained on the training set for 100 rounds respectively. The trained models were used to restore the test set data. The results are shown in Table 1.

[0150] Table 1 Comparison of restoration results of different models

[0151] Model Name PSNR SSIM PSRGAN 15.01 0.442 HWMNet 22.53 0.699 DehazeFormer 22.54 0.697 The present invention 25.40 0.816

[0152] As can be seen from Table 1, the present invention achieves optimal performance, with a PSNR of 25.40 and an SSIM of 0.816, both significantly higher than other existing image restoration methods. This demonstrates the superiority of the present invention for the specific task of restoring infrared images of blast furnace roofs.

[0153] The present invention has also been applied at a blast furnace operation site. Directly applying noisy images to visual inspection yielded an accuracy rate of only 54.13%. However, using the present invention for image restoration significantly improved the accuracy to 99.65%. This demonstrates that the present invention can address the problem of harsh blast furnace operating conditions introducing complex noise into infrared images, significantly reducing the accuracy of image-based visual inspection tasks.

[0154] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0155] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0156] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0157] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0158] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A method for restoring an infrared image of a blast furnace roof, characterized in that: include: Obtain an infrared image of the furnace top when the blast furnace is not working as a clear image, and an infrared image of the furnace top when the blast furnace is working as a noise image, pair the clear image and the noise image, and establish an image pair of the clear image and the noise image as a data set; Constructing a blast furnace top infrared image restoration model, wherein an encoder of the blast furnace top infrared image restoration model includes at least one layer of infrared enhancement modules, and each layer of the infrared enhancement modules incorporates position coding information of the blast furnace top image; The infrared enhancement module includes a self-attention module embedded in frequency domain feature enhancement, a discrete wavelet transform module based on deformable convolution, and a spatial detail enhancement module. The self-attention module embedded in frequency domain feature enhancement extracts features of different frequencies, the discrete wavelet transform module based on deformable convolution extracts local features and multi-scale features, and the spatial detail enhancement module optimizes non-uniform features. The blast furnace top infrared image restoration model is trained using the data set, and the infrared image of the blast furnace top to be restored is input into the trained blast furnace top infrared image restoration model to obtain a restored image.

2. The blast furnace top infrared image restoration method according to claim 1, characterized in that: The position coding information of each layer of the infrared enhancement module integrated into the blast furnace top image is specifically: Segment the pixels of the furnace top infrared image and calculate the centroid of the segmented pixels; divide the furnace top infrared image into blocks and determine the block number p corresponding to the centroid; Get the M domain image blocks of the pth image block, denoted as N M (p), the sine and cosine functions are used to position encode the image blocks. The position encoding rule is: Where W and H represent the width and height of the furnace top infrared image respectively, x b and y b are the horizontal and vertical coordinates of the center of mass, N M+1 (p) represents the area covered by the centroid and its M domain image blocks, i represents the dimension index, and pos represents N M+1 Any feature point in (p), PE(pos,i) represents N M+1 The position encoding information of the i-th dimension corresponding to the pos-th feature point in (p); β represents the frequency coefficient, C represents the dimension size of the feature; DC represents the even dimension index, and SC represents the odd dimension index; The features output by the infrared enhancement module of the previous layer are added to the infrared image of the furnace top encoded according to the position encoding rule as the input of the infrared enhancement module of the next layer.

3. The blast furnace top infrared image restoration method according to claim 1, characterized in that: The self-attention module embedded in the frequency domain feature enhancement extracts features of different frequencies, specifically: The self-attention module embedded in the frequency domain feature enhancement includes a frequency domain enhancement module and an attention module. The input feature of the self-attention module embedded in the frequency domain feature enhancement is denoted as x, and the input feature is converted into amplitude and phase through Fourier transform, denoted as F(x), where F() represents Fourier transform; F(x) is passed to the frequency domain enhancement module for frequency domain enhancement processing. The frequency domain enhancement module includes: a first layer of depthwise separable convolution layer, a second layer to calculate the mean and maximum value of F(x), a third layer of depthwise separable convolution layer, and then a Sigmoid activation function is used to obtain a normalized attention map, denoted as α; F(x) is multiplied by α to obtain a weighted feature vector, denoted as f′; f′ is subjected to inverse Fourier transform and then enters the attention module; The query, key, and value obtained by the feature map after inverse Fourier transform are denoted as Q f , K f 、V f , the final output feature is obtained by adding the residual connection to the original input: fs=SA(Q f +K f +V f )+x, Among them, fs is the final output feature of the self-attention module embedded in the frequency domain feature enhancement, and SA() is the window self-attention calculation.

4. The blast furnace top infrared image restoration method according to claim 1, characterized in that: The discrete wavelet transform module based on deformable convolution extracts local features and multi-scale features, specifically: The output feature of the self-attention module embedded in the frequency domain feature enhancement is denoted as fs, and fs obtains feature maps of different scales through two deformable convolutions of different scales, denoted as fs1 and fs2, respectively. The scale feature of fs1 is smaller than that of fs2; Transform fs1 into four different wavelet subbands through wavelet transform, and the feature vector obtained by splicing the four subbands along the channel dimension is recorded as fs wavelet ; After the feature modulation module, fs wavelet As a modulation parameter applied to fs2, the large-scale feature with local detail enhancement is obtained, recorded as fs2′, fs2′=β′·fs2, β′ represents the modulation parameter; fs wavelet The output feature obtained by combining the convolution operation with fs1 and fs2′ is recorded as fs fusion , fs fusion The calculation method is: fs fusion =τ⊙[fs1,fs wavelet ]+fs2′, Where ⊙ represents the Hadamard product, τ represents the local dynamic range coefficient; The local dynamic range coefficient is calculated as: τ=Tanh(LocalMaxMin(fs2)), Among them, τ represents the local dynamic range coefficient, LocalMaxMin() means calculating the difference between the maximum pixel and the minimum pixel in the local window, and Tanh() means limiting the value to [-1,1].

5. The blast furnace top infrared image restoration method according to claim 1, characterized in that: The spatial detail enhancement module optimizes non-uniform features, specifically: The output feature of the self-attention module embedded in the frequency domain feature enhancement is denoted as fs; after being processed by the spatial detail enhancement module, fs is combined with the input fs to obtain the output feature fs′ with edge high-frequency feature enhancement; The spatial detail enhancement module includes at least one detail enhancement convolution integrating different differential convolutions, and the weight of the detail enhancement convolution is: w DEConv =Scale(w CDC +w RDC )+Softmax(w HDC +w VDC +w ADC ), where w DEConv The weight of detail enhancement convolution, w CDC 、w RDC 、w HDC 、w VDC 、w ADC They are the weights of the center difference convolution, the weights of the rotation difference convolution, the weights of the horizontal difference convolution, the weights of the vertical difference convolution, and the weights of the diagonal difference convolution; Scale is a trainable scaling factor, and Softmax() is the activation function.

6. The blast furnace top infrared image restoration method according to claim 1, characterized in that: The output of the infrared enhancement module is: in, is the output vector of the infrared enhancement module, fs is the output feature of the self-attention module embedded in the frequency domain feature enhancement, and fs fusion is the output feature of the discrete wavelet transform module based on deformable convolution, fs′ is the output feature of the spatial detail enhancement module; g is the weighting coefficient, and the calculation method of g is: g=sigmoid(MLP([fs fusion ,fs′])), Among them, MLP() is the gating mechanism operation and sigmoid() is the activation function.

7. The blast furnace top infrared image restoration method according to any one of claims 1 to 6, characterized in that: When training the blast furnace top infrared image restoration model, the total loss function is: L=α′L combined +β″L edge +γL local , Among them, L is the total loss function, L combined is the global loss function, L edge is the marginal loss function, L local is the local loss function of the key area; α′, β″, γ are control parameters, α′+β″+γ=1; The calculation method of the global loss function is: L combined =ρ1·(1-SSIM(o i ,T i ))+ρ2·MSE(O i ,T i ), Among them, ρ1 and ρ2 are weight coefficients, SSIM() is the structural similarity loss function, and MSE() is the mean square error loss function; i 、T i Respectively represent the output i-th restored image and i-th real clear image; The calculation method of the edge loss function is: L edge ∑|E(O i )-E(T i )|, Where, E() represents the edge extraction operation; The calculation method of the local loss function of the key area is: L local =∑W(x,y)|O(x,y)-T(x,y)|, Among them, W(x,y) is the Gaussian weight. The closer to the center of the key area, the higher the Gaussian weight. The key area is the area where the various parts of the blast furnace body are located in the infrared image of the blast furnace top; O(x,y) represents the pixel value output by the model at the coordinate (x,y), and T(x,y) represents the target value at the coordinate (x,y).

8. A blast furnace roof infrared image restoration system, characterized in that: include: A data acquisition module is used to acquire an infrared image of the furnace top when the blast furnace is not working as a clear image and an infrared image of the furnace top when the blast furnace is working as a noise image, pair the clear image and the noise image, and establish an image pair of the clear image and the noise image as a data set; A model construction module is used to construct a blast furnace top infrared image restoration model, wherein the encoder of the blast furnace top infrared image restoration model includes at least one layer of infrared enhancement modules, and each layer of the infrared enhancement modules is integrated with position coding information of the blast furnace top image; The infrared enhancement module includes a self-attention module embedded in frequency domain feature enhancement, a discrete wavelet transform module based on deformable convolution, and a spatial detail enhancement module. The self-attention module embedded in frequency domain feature enhancement extracts features of different frequencies, the discrete wavelet transform module based on deformable convolution extracts local features and multi-scale features, and the spatial detail enhancement module optimizes non-uniform features. A model training module, configured to train the blast furnace top infrared image restoration model using a data set; The image restoration module is used to input the infrared image of the blast furnace top to be repaired into the trained blast furnace top infrared image restoration model to obtain a restored image.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the blast furnace top infrared image restoration method according to any one of claims 1 to 7 is implemented.

10. A blast furnace roof infrared image restoration device, characterized by: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for restoring the infrared image of the blast furnace top as claimed in any one of claims 1 to 7 is implemented.

Images