Mamba lightweight building image super-resolution reconstruction method based on equation of state model

Abstract

The application discloses a Mamba lightweight building image super-resolution reconstruction method based on a state equation model, and specifically comprises the following steps: step 1, obtaining a high-definition building image, and performing down-sampling on the original high-resolution image to obtain a corresponding low-resolution image; step 2, performing data enhancement on the low-resolution image; step 3, constructing a Mamba lightweight building image super-resolution network model based on a state equation model; and step 4, training the super-resolution model to obtain a Mamba lightweight building image super-resolution reconstruction model based on the state equation model. The application can effectively restore the regular texture of a building image, solves the trade-off problem between an effective receptive field and performance of an existing lightweight image super-resolution network, realizes efficient feature extraction of an image to obtain global information without increasing the calculation overhead, and effectively improves the lightweight property and reconstruction quality of the lightweight building image super-resolution model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a Mamba lightweight building image super-resolution reconstruction method based on a state equation model. BACKGROUND

[0002] As a non-contact monitoring method, visual measurement method is widely used in structural surface defect detection, structural dynamic and static response measurement and other application scenarios, and is playing an increasingly important role in structural health monitoring. Due to the constraints of equipment cost, large coverage of building structure, limited camera distribution position, complex environmental conditions and other factors, the images collected by the visual measurement system often have the problem of insufficient resolution. Super-resolution reconstruction technology aims to reconstruct high-resolution (HR) images from corresponding low-resolution (LR) images, providing an effective way to solve this problem. However, the surface of building structure often has regular texture features, so the super-resolution reconstructed image is sensitive to artifacts and noise, and its reconstruction quality depends on the ability of the network to model long-range relationships. With the introduction of modern deep learning models such as convolutional neural networks and Transformer models, the performance of image super-resolution models has been continuously improved. However, the high computational cost and memory consumption of many existing models make them difficult to be applied in resource-limited edge and mobile devices. In order to improve the efficiency of super-resolution reconstruction of building structure surface images, it is urgent to design a lightweight and efficient SR model with long-range relationship modeling capability.

[0003] In recent years, researchers have proposed a large number of lightweight and efficient SR methods, but there is still a lack of enough convolutional layers to increase the overall receptive field, limiting their ability to capture global information, which is not conducive to extracting higher-level patterns and structures, and restricting their anchor pixel reconstruction performance.

[0004] For CNN-based super-resolution networks, although the effective receptive field is limited, due to the good efficiency of convolution parallel operation, it is suitable for deployment on resource-limited devices. In contrast, Transformer-based super-resolution methods have better performance than CNN-based methods due to their larger receptive field, but their computational complexity grows quadratically with the input size, which is not friendly to lightweight super-resolution networks. To apply Transformer to lightweight super-resolution, existing research has adopted some efficient attention mechanisms, such as shift window attention, which effectively reduces the time complexity, but usually at the cost of sacrificing the global effective receptive field, and does not essentially escape the trade-off between effective receptive field and model performance.

[0005] In recent years, Structured State-Space Sequence Models (S4), particularly the improved Mamba, have emerged as efficient and effective backbone networks for building deep networks. This development offers a potential solution for balancing global receptive field and computational efficiency in image inpainting. Specifically, the discretized state-space equations in Mamba can be formalized recursively and, through specially designed structure reparameterization, can model long-range dependencies. This means that Mamba-based inpainting networks can naturally activate more pixels, improving reconstruction quality. Furthermore, Mamba's parallel scan algorithm not only captures global features of an image with linear complexity but also processes each token in parallel, facilitating efficient training on modern hardware such as GPUs. These promising characteristics inspire us to explore Mamba's potential for efficient long-range modeling in lightweight image inpainting networks.

[0006] However, standard Mamba was originally designed for one-dimensional sequence data in Natural Language Processing (NLP), so its direct application to image inpainting tasks presents several challenges. First, because Mamba processes one-dimensional flattened image sequences recursively, distant pixels may be mistakenly identified as spatially close pixels, leading to the forgetting of local pixel information. Second, due to the need to capture long sequence dependencies, the number of hidden states in the state-space equations is often enormous, causing channel redundancy and affecting the effective learning of key channel representations. Furthermore, Mamba's image flattening scanning method needs further optimization to better capture global image information. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a lightweight Mamba super-resolution reconstruction method for building images based on a state equation model, so as to solve the problems of slow speed and insufficient image information extraction capability of existing image super-resolution methods, and to design a local feature enhancement module to effectively improve the shortcomings of existing Mamba models.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: a Mamba lightweight building image super-resolution reconstruction method based on a state equation model, comprising the following steps:

[0009] Step S1: Acquire high-resolution images of buildings and use bicubic interpolation algorithm to obtain low-resolution images corresponding to the lightweight image dataset;

[0010] Step S2: Perform data augmentation on the input low-resolution data;

[0011] Step S3: Construct a Mamba lightweight building image super-resolution reconstruction network based on a state equation model, i.e., a MambaESR network model, which is specifically divided into a Mamba feature modulation network MFM and a local feature reinforcement module LFRM;

[0012] Step S4: Train the MambaESR network model constructed in step S3;

[0013] Step S5: Input the test set into the MambaESR network model trained in step S4, and output a super-resolution image.

[0014] In a preferred embodiment: in step S1, a bicubic interpolation algorithm is used to downsample the building high-resolution image to obtain a corresponding low-resolution image.

[0015] In a preferred embodiment: in step S2, different scale scaling, flipping, and rotation are used for data enhancement to improve the generalization ability of the model.

[0016] In a preferred embodiment: the Mamba feature modulation network in step S3 replaces the traditional CNN and Transformer architecture with a network architecture based on a state equation model to increase the receptive field while reducing the computational cost according to the characteristics of building images that rely heavily on long-range information dependencies; the specific design process of the visual state equation model is as follows:

[0017] Step S3.1: First, perform a channel segmentation operation on the normalized input feature; given the input feature X, this process is represented as:

[0018] [X0,X1]=CHUNK(X) (1)

[0019] where CHUNK(·) is the channel segmentation operation, and [X0,X1] are the two segmented features of the input feature;

[0020] Step S3.2: In the first branch, the number of feature channels is expanded to γC through linear mapping, where γ is a predefined channel expansion factor, then transformed through deep convolution, SiLU activation function, state space module SSM layer, and layer normalization LayerNorm, represented as:

[0021]

[0022] where DW-Conv 3×3 (·) represents a deep separable 3x3 convolution kernel, φ(·) represents a GELU activation function, Linear(·) represents a linear mapping operation, SSM(·) represents a state space module, LN(·) represents a layer normalization operation, The intermediate feature of the first branch after the above operation;

[0023] Step S3.3: In the second branch, the feature channels are also expanded to γC through a linear layer and processed by a SiLU activation function; the feature representations of the two branches are fused by element-wise multiplication, denoted as:

[0024] X2 = φ(Linear(X2)) (3)

[0025] where X2 is the output feature of the second branch;

[0026] Step S3.4: The number of feature channels is mapped back to C to generate the output X out , which maintains the same shape as the input. This process can be represented as:

[0027] X out = Linear(X1 × X2) (4).

[0028] In a preferred embodiment: step S3 uses multiple different scanning image methods to spread an image into a one-dimensional image in different ways, obtaining global information of the image without increasing additional burden;

[0029] In a preferred embodiment: the specific design process of the local feature enhancement module is:

[0030] Step S3.4: In the local feature enhancement module, first use a 1 × 1 convolution to double the number of channels of the input feature, which is represented as:

[0031] X' = Conv 1×1 (X in ) (5)

[0032] Step S3.5: A depth separable convolution kernel is introduced to encode the spatial local context information, and a self-residual is added in the depth convolution kernel to improve the reconstruction result; then apply the GELU function to the hidden layer for nonlinear mapping, pass the result through an SE attention module, and finally reduce the channel to the original input dimension through a 1 × 1 convolution to obtain the output feature, which is represented as:

[0033] X out = Conv 1×1 (SE(φ(DWConv 3×3 (X')+X))) (6)

[0034] where X in and X out represent the input feature and the output feature respectively, and DWConv 3×3denotes a depthwise separable 3x3 convolution kernel, SE module denotes a spatial attention mechanism module, and φ(·) denotes a GELU activation function.

[0035] In a preferred embodiment: the specific process of step S4 is:

[0036] Step S4.1: input the building low-resolution image, generate a super-resolution image through the network model, calculate the regularization loss between the reconstructed image and the real image, use the mean absolute error loss and the FFT-based frequency loss function to optimize the network parameters, continuously train the super-resolution reconstruction network through backward propagation, train and learn to obtain the final trained super-resolution reconstruction network; the loss function is defined as:

[0037] Loss = ||I SR -I HR ||1+γ||F(I SR )-F(I HR )||1 (7)

[0038] where I HR is the high-resolution Ground-truth image, ||·||1 represents the L1 norm, F represents the fast Fourier transform, and γ is the weight factor, which is empirically set to 0.1;

[0039] Step S4.2: design and train three different sizes of super-resolution networks, named MambaESR-S, MambaESR-M and MambaESR-L respectively; MambaESR-S has a dimension of 36 and uses 8 modules stacked; MambaESR-M has a dimension of 48; MambaESR-L has a dimension of 48 and is composed of 12 modules; during training, each batch contains 16 LR images, and the input size is 256x256; the Adam optimizer with β1=0.9 and β2=0.99 is used to solve the proposed model; the initial learning rate is set to 1x10 -3 , the minimum value is set to 1x10 -6 , and the learning rate is updated by the cosine annealing algorithm; the weight of the exponential moving average line EMA is set to 0.999; in the implementation process, the model is trained for a total of 10 6 times of iterations on an NVIDIA GeForce RTX3080 GPU.

[0040] Compared with the prior art, the present application has the following beneficial effects: 1. The present application improves and perfects the traditional model based on the Mamba modulation mechanism, realizes a good balance between performance and efficiency, and overcomes the trade-off problem between computational complexity and reconstruction performance of the traditional super-resolution model. 2. The present application studies the principle and application of the Mamba core scanning mechanism, studies the influence of various scanning directions on the image super-resolution effect, thereby better modeling the long-range feature relationship, and significantly improving the detail expressiveness of image reconstruction. 3. The present application designs a local feature enhancement module for improving the Mamba network, which can effectively encode and enhance the local context information, and simultaneously perform channel mixing, effectively improving the local pixel forgetting and channel redundancy of the original Mamba, and effectively improving the model effect. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The flowchart of the Mamba lightweight building image super-resolution reconstruction method based on the state equation model in the preferred embodiment of the present application;

[0042] Figure 2 The overall flowchart and network internal structure diagram of the specific architecture of the preferred embodiment of the present application;

[0043] Figure 3 The comparison diagram of the local attribution map (LAM) and the diffusion index (DI) of various efficient super-resolution models;

[0044] Figure 4 The scanning method diagram of the image in the MFM;

[0045] Figure 5 The improvement effect diagram of the LFRN on the Mamba in the preferred embodiment of the present application;

[0046] Figure 6 The effect comparison diagram of the preferred embodiment of the present application and other lightweight super-resolution models;

[0047] Figure 7 The detail display comparison diagram of the preferred embodiment of the present application and the existing algorithm on the building image on the x4 super-resolution (SR) task. DETAILED DESCRIPTION

[0048] The present application will be further described below in conjunction with the drawings and embodiments.

[0049] It should be pointed out that the following detailed description is exemplary and is intended to provide further description of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as generally understood by those skilled in the art to which the present application belongs.

[0050] It should be noted that the terms used herein are only intended to describe specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, they indicate the presence of the features, steps, operations, devices, components and / or combinations thereof.

[0051] A Mamba lightweight building image super-resolution reconstruction method based on a state equation model, as shown in Figures 1-7 The low-resolution image block I LR is input to the network, and the high-resolution result image I SR .

[0052] Specifically, the following steps are implemented:

[0053] Step S1: Obtain a high-resolution building image, and downsample the original high-resolution building image to obtain a corresponding low-resolution image;

[0054] The specific process of the down-sampling is: using bicubic interpolation algorithm to process the high-resolution image, and obtaining the corresponding low-resolution image according to different reduction factors (such as 2, 3, 4, etc.).

[0055] Step S2: Data augmentation is performed on the low-resolution image;

[0056] The specific data augmentation operation includes using different scale scaling, flipping, rotating and other methods for data augmentation to improve the generalization ability of the model.

[0057] Step S3: In view of the trade-off between the effective receptive field and the performance of the traditional super-resolution model, and the characteristics of the building image having regular texture and strong long-range dependence, a Mamba lightweight building image super-resolution network model based on a state equation model is constructed, which is specifically divided into two modules, a Mamba feature modulation unit and a local feature enhancement module, as shown in Figure 2 The model based on the state equation has a significant effect on the effective receptive field compared with the model based on the traditional CNN and the Transformer, as shown in Figure 3 .

[0058] For details Figure 2 , the network model proposed by the present application is composed of three modules: shallow feature extraction, deep feature extraction and up-sampling feature reconstruction.

[0059] Step S3.1: Specifically, the network first applies a convolutional layer with a kernel size of 3x3 pixels to extract shallow features, converting the input LR image into a feature space and generating shallow features f0. The specific process can be represented as follows:

[0060] f0 = Conv 3×3 (I LR (1)

[0061] Among them, Conv 3×3 (·) represents a 3×3 convolutional layer.

[0062] Step S3.2: The shallow features f0 are passed to the deep feature extraction module, which consists of a series of Vision Mamba Modulation Blocks (VMMBs). This process can be described as follows:

[0063] f k =H k (f k-1 ),k=1,…,n (2)

[0064] Among them, H k (·) represents the k-th VMMB. k-1 and f k Let represent the input and output features of the k-th VMMB, respectively.

[0065] Step S3.3: To better recover the HR target image, a global residual connection is introduced. Finally, a lightweight upsampling layer consisting of a 3×3 convolutional layer and subpixel convolutions is used to quickly reconstruct the HR image. The specific process can be represented as follows:

[0066] I SR =P θ (f k +f0) (3)

[0067] Among them, I SR For the predicted HR image, P θ (·) represents the described upsampling reconstruction module.

[0068] Step S3.4: Building structures often exhibit regular texture features, making their super-resolution reconstructed images sensitive to artifacts and noise. The reconstruction quality particularly depends on the network's ability to model long-range relationships. The standard Mamba model processes input data causally, allowing for the retrieval of long-range dependencies within the image through different scanning methods. Therefore, a state-space model is introduced into the image restoration task to replace the traditional CNN and Transformer architectures, achieving more efficient global feature modeling. The architecture of the Mamba feature modulation unit is as follows...Figure 2 The specific design process is as follows:

[0069] Step S3.4.1: First, the normalized input feature X is subjected to a channel segmentation operation to generate two part features, and is processed through two parallel branches. It can be represented as:

[0070] [X0,X1]=CHUNK(X) (4)

[0071] Where CHUNK(·) is the channel segmentation operation.

[0072] Step S3.4.2: In the first branch, the feature channel number is expanded to γC through linear mapping, where γ is a predefined channel expansion factor, and then is transformed through deep convolution, SiLU activation function, state space module (SSM) layer and layer normalization (LayerNorm). It can be represented as:

[0073]

[0074] Where DW-Conv n×n (·) represents a depth separable 3x3 convolution kernel, φ(·) represents a GELU function, Linear(·) represents a linear layer, and SSM represents a state space model, which contains different scanning image methods such as Figure 4 As shown in the figure, our model uses 4 directions to expand the image to obtain the global information of the image.

[0075] Step S3.4.3: In the second branch, the feature channel is also expanded to γC through a linear layer and is processed through a SiLU activation function. The features of the two branches are fused through element-wise multiplication. It can be represented as:

[0076] X2=φ(Linear(X2)) (6)

[0077] Step S3.4.4: Finally, the feature channel number is mapped back to C to generate the output X out , which has the same shape as the input. This process can be represented as

[0078] X out =Linear(X1×X2) (7)

[0079] Step S3.5: Although the Mamba feature modulation unit can model the long-range relationship between pixels, it ignores the importance of local context information for super-resolution image reconstruction and has the problem of channel redundancy. Therefore, a local feature enhancement module is designed to be added after the Mamba feature modulation unit to supplement the local context information, and the local feature enhancement module is as shown in Figure 2as shown.

[0080] Step S3.5.1: In the local feature enhancement module, first use a 1x1 convolution kernel to double the channel number of the input feature, which can be represented as:

[0081] X' = Conv 1×1 (X in ) (8)

[0082] where X in represents the input image, and X' represents the intermediate feature obtained after the convolution kernel.

[0083] Step S3.5.2: A depth separable convolution kernel is introduced to encode the spatial local context information, and a self-residual is added in the depth convolution kernel to improve the reconstruction result, so that the model can further improve the performance without calculating the cost in the inference stage. Then apply the GELU function to the hidden layer for nonlinear mapping, pass the result through an SE attention module, and finally reduce the channel to the original input dimension through convolution to obtain the output feature. The whole process can be represented as:

[0084] X onv = Conv 1×1 (SE(φ(DW-Conv 3×3 (X')))) (15)

[0085] where X out represents the output feature, DW-Conv 3×3 represents the use of a depth separable 3x3 convolution kernel, SE module represents the spatial attention mechanism module, and φ(·) represents the activation function GELU. We apply RELU and global average pooling to the output of the last layer of the network to obtain the channel activation value, as shown in Figure 5 Figure 5 The left figure does not use the local feature enhancement module (LFRM). It is obvious that most channels are not activated, which indicates the channel redundancy problem in the original Mamba modulation block. Figure 5 The right figure uses LFRM, which effectively solves the channel redundancy problem. The local feature enhancement module effectively improves the channel redundancy phenomenon of MFM.

[0086] Step S3.6: Design the overall architecture. In recent lightweight image super-resolution models, the module design mainly follows the process of Transformer, that is, first extract global features, and then extract local feature information. As shown in Figure 2 The depth feature extraction module of the network mainly consists of two parts, which can be represented as

[0087] X' = MFM(LN(X in ))+α·X in ​(8)

[0088] X out = LFRM(LN(X')) + β · X' (9)

[0089] where X in and X out represent the input and output feature maps of the deep feature extraction module, LN(·) is the LayerNorm operation, and X' represents the intermediate feature. The Mamba Feature Modulation (MFM) and Local Feature Reinforcement Module (LFRM) play different roles in the network. The former effectively models long-range dependencies, while the latter focuses more on fine local contextual information. They complement each other and work together to achieve more accurate image reconstruction. Finally, we also add learnable scale factors a and β to dynamically adjust the information flow. The network can dynamically adjust the information transmission between different layers, making the model more flexible to handle different input features and task requirements.

[0090] Step S4: Train the network model constructed in step 3;

[0091] Step S4.1: First, define the loss function. This method optimizes our network parameters by using the mean absolute error loss and the FFT-based frequency loss function, which can be represented as:

[0092] Loss = ||I SR -I HR ||1 + γ||F(I SR )-F(I HR )||1 (12)

[0093] where I HR is the high-resolution ground-truth image, ||·||1 represents the L1 norm, F represents the Fast Fourier Transform, and γ is the weight factor, which is empirically set to 0.1.

[0094] S4.2: To embody the performance of the present application, three different sizes of super-resolution networks were designed and trained, named MambaESR-S, MambaESR-M and MambaESR-L, respectively. Among them, MambaESR-S has a dimension of 36, using 8 module stacks; MambaESR-M has a dimension of 48; while MambaESR-L has a dimension of 48, using 12 module groups. During the training process, each batch contains 16 LR images, with an input size of. We use the Adam optimizer with and to solve the proposed model. We set the initial learning rate to, the minimum value to, and the learning rate is updated by the cosine annealing algorithm. The weight of the exponential moving average line (EMA) is set to 0.999. In the implementation process, the model of the present application uses a total of iterations for training on an NVIDIA GeForce RTX 3080 GPU.

[0095] Step S5: input the test set into the network model trained in step 4, and output the super-resolution image.

[0096] To evaluate the performance of the present application, we compared it with the most advanced lightweight SR methods, including CNN-based methods CARN, LAPAR, ShufferMixer, SAFMN, PAN, Transformer-based methods LatticeNet, ELAN, SwinIR-light, SPIN, and existing Mamba-based methods MambaIR. Table 1 reports the quantitative comparison results on the benchmark dataset with scale factors of ×2 and ×4, including the building-specific dataset Urban100. In addition to the PSNR / SSIM indicators, we also list the parameter quantity (#params) and the floating-point computation quantity (#FLOPs), where #params is related to the consumption of memory, and #FLOPs is related to energy use. #FLOPs is calculated by the fvcore1 library after super-resolving the LR image to the HR image of 1280×720. Table 1 and Figure 6 The results of Table 1 and

[0097] To further illustrate the effectiveness of our method, we also compared the visual effects of ×4 magnification on the Urban100 dataset. As Figure 7As shown, most of the compared lightweight super-resolution methods cannot correctly recover the texture and details of the building image, and there are obvious blur artifacts and distortion phenomena. In contrast, the Mamba lightweight building image super-resolution reconstruction method based on the state equation model proposed by us can obtain the best visual quality and recover more building structure detail information.

[0098] In addition, in order to verify the effectiveness of the Mamba feature modulation block (MFM) and the local feature reinforcement module (LFRM) designed by us, we further conduct ablation experiments on MFM and LFRM. The results of Table 2 and Table 3 show that all components of MFM and LFRM are the key to performance improvement, and the results of Table 4 show that our scanning mechanism has significant advantages in performance and computational efficiency, proving its superiority and stability in feature extraction tasks, far exceeding other scanning strategies.

[0099] Table 1: PSNR / SSIM quantitative evaluation of super-resolution reconstruction performance of the method of the present application on four related data sets and performance comparison with other existing methods ("_" and "_" mark the best / second best performance), wherein Urban 100 is a set of building test (including 100 urban landscape images) ​

[0100]

[0101] Table 2: Ablation experiment verification of MFM and LFRM in the method of the present application

[0102]

[0103] Table 3: Ablation experiment verification of LFRM module performance in the method of the present application

[0104]

[0105] Table 4: Ablation experiment verification of the performance of the MFM module in the method of the present application to spread the image in different directions

[0106]

[0107] The above is the preferred embodiment of the present application, any changes made according to the technical solutions of the present application, as long as the generated function does not exceed the scope of the technical solutions of the present application, belongs to the protection scope of the present application.

Claims

1. A method for Mamba light-weight building image super-resolution reconstruction based on a state equation model, characterized in that, Comprise the following steps: Step S1: Obtain a high-definition image of a building and obtain a low-resolution image corresponding to a lightweight image dataset using a bicubic interpolation algorithm; Step S2: Data augmentation is performed on the input low-resolution data; Step S3: Construct a Mamba lightweight building image super-resolution reconstruction network based on a state equation model, namely a MambaESR network model, which is specifically divided into a Mamba feature modulation network MFM and a local feature reinforcement module LFRM; Step S4: Train the MambaESR network model constructed in step S3; Step S5: Input the test set into the MambaESR network model trained in step S4, and output a super-resolution image; The Mamba feature modulation network of step S3 is designed to address the characteristics of building images that rely heavily on long-range information dependencies. It uses a network architecture based on a state equation model to replace traditional CNN and Transformer architectures, thereby increasing the receptive field while reducing computational cost. The specific design process of the visual state equation model is as follows: Step S3.1: First, perform a channel segmentation operation on the normalized input features. Given input features X, this process is represented as: wherein, is a channel split operation, are two parts of the input feature after being split. Step S3.2: In the first branch, the number of feature channels is expanded by a linear mapping to where is a predefined channel expansion factor, followed by a transformation through a depth convolution, a SiLU activation function, a state space module SSM layer, and a layer normalization LayerNorm, denoted as: wherein, representing a depth separable convolution kernel, denotes a function, Linear( ) denotes a linear mapping operation, SSM( ) denotes a state space module, LN( ) denotes a layer normalization operation, is an intermediate feature of the first branch after the above operations; Step S3.3: In the second branch, the feature channels are also expanded through a linear layer to and processed by a SiLU activation function; the feature representations of the two branches are fused by element-wise multiplication, denoted as: wherein is the output feature of the second branch; Step S3.4: The number of features channels is mapped back to C, generating the output which maintains the shape of the input, this process is represented as: 。 2. The Mamba light-weight building image super-resolution reconstruction method based on a state equation model according to claim 1, characterized in that: In step S1, the bicubic interpolation algorithm is used to downsample the high-resolution building image to obtain the corresponding low-resolution image.

3. The Mamba light-weight building image super-resolution reconstruction method based on a state equation model according to claim 1, characterized in that: In step S2, different scale scaling, flipping, and rotation are used for data augmentation to improve the generalization ability of the model.

4. The Mamba light-weight building image super-resolution reconstruction method based on a state equation model according to claim 1, characterized in that: In step S3, multiple different scanning image methods are used to spread a single image into a one-dimensional image in different ways, obtaining global information without additional burden.

5. The Mamba light-weight building image super-resolution reconstruction method based on a state equation model according to claim 1, characterized in that: The specific design process of the local feature reinforcement module is as follows: Step S3.4: In the local feature enhancement module, first use Convolution doubles the number of channels of input features, and this process is represented as: Step S3.5: Introducing a depth separable convolution kernel to encode the spatial local context information, and adding a self-residual in the depth convolution kernel to improve the reconstruction result; then applying the GELU function to the hidden layer for nonlinear mapping, passing the result through an SE attention module, and finally passing the result through Convolution reduces the channel to the original input dimension to obtain the output feature, and this process is represented as: wherein, and represent input features and output features, respectively, represent depth separable convolution kernels, and the SE module represents a spatial attention mechanism module, represent activation functions.

6. The Mamba light-weight building image super-resolution reconstruction method based on a state equation model according to claim 1, characterized in that: The specific process of step S4 is as follows: Step S4.1: Input the low-resolution image of the building, and generate a super-resolution image through the network model. Calculate the regularization loss between the reconstructed image and the real image, and use the mean absolute error loss and the FFT-based frequency loss function to optimize the network parameters. Through continuous backpropagation training of the super-resolution reconstruction network, training and learning are performed to obtain the final trained super-resolution reconstruction network. The loss function is defined as: wherein is a high resolution Ground-truth image, denotes a norm, F denotes a Fast Fourier Transform, is a weight factor, empirically set to 0.1; Step S4.2: design and train three different sizes of super-resolution network, named MambaESR-S, MambaESR-M and MambaESR-L respectively; wherein MambaESR-S has a dimension of 36, using 8 module stacks; MambaESR-M has a dimension of 48; and MambaESR-L has a dimension of 48, using 12 module groups; during the training process, each batch contains 16 LR images, and the input size is ; the proposed model is solved using the Adam optimizer with and ; the initial learning rate is set to , the minimum value is set to , and the learning rate is updated by the cosine annealing algorithm; the weight of the exponential moving average line EMA is set to 0.999; in the implementation process, the model is trained a total of iterations on an NVIDIA GeForce RTX 3080 GPU.

Images