Lightweight building image super-resolution reconstruction method based on expansion feature modulation network

Abstract

The application discloses a light-weight building image super-resolution reconstruction method based on an inflation feature modulation network, and specifically comprises the following steps: step 1, obtaining a high-resolution building image, down-sampling an 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 light-weight building image super-resolution network model based on an inflation feature modulation network; and step 4, training the model to obtain a light-weight building image super-resolution reconstruction model based on the inflation feature modulation network. The application effectively recovers information such as texture details of a low-resolution building image, solves the deficiency of a long-range dependency relationship in existing light-weight image super-resolution network modeling, realizes accurate and efficient long-range dependency relationship modeling in a global feature range without increasing computational overhead, and effectively improves the light weight and reconstruction quality of a building image super-resolution model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of building image super-resolution reconstruction technology in the field of vision-based structural health monitoring, in particular to a lightweight building image super-resolution reconstruction method based on an inflation feature modulation network. BACKGROUND

[0002] Structural health monitoring is of great significance to ensure the safety, reliability and durability of various engineering structures. As a non-contact monitoring method, visual measurement is used in structural static and dynamic displacement measurement, crack measurement and other application scenarios, and is playing an increasingly important role in the field of structural health monitoring. In the process of visual measurement of structural health monitoring, due to the limitations of measurement environment, equipment cost and other factors, the acquired images often have 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, building structures often have regular texture features, so their super-resolution reconstruction images are sensitive to artifacts and noise, and their reconstruction quality depends on the network's ability to model long-range relationships. The present application proposes a lightweight and efficient building image super-resolution reconstruction method, which includes the design of an adaptive inflation feature modulation network, which uses multiple inflation convolutions to stack equivalent large convolution kernels to reduce computational cost while achieving large receptive field, and through simple element-wise multiplication operations, it realizes spatial attention mechanism, so that the model can adaptively focus on the key features in the image; further, a local feature enhancement module based on reparameterization operation is designed to reduce network inference time, ultimately achieving a good trade-off between performance and efficiency for regular texture building image super-resolution reconstruction.

[0003] Currently, there is no lightweight building image super-resolution method based on inflation feature modulation network in the published literature and patents. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a lightweight building image super-resolution reconstruction method based on an inflation feature modulation network, which solves the limitations of existing lightweight super-resolution models in modeling long-range dependencies, and realizes accurate and efficient long-range dependency modeling in the global feature range without increasing the number of parameters, effectively improving the robustness and reconstruction quality of lightweight image super-resolution models.

[0005] To achieve the above purpose, the present application adopts the following technical solution: a lightweight building image super-resolution reconstruction method based on an inflation feature modulation network, comprising the following steps:

[0006] Step S1: Obtain a high-resolution image of a building surface, and downsample the original high-resolution image of the building to obtain a corresponding low-resolution image;

[0007] Step S2: Perform data enhancement on the low-resolution image;

[0008] Step S3: A lightweight building image super-resolution network model based on the dilated feature modulation network is constructed for the characteristics of regular texture and large long-range dependence in the building image, which is specifically divided into a dilated separable modulation unit DSMU and a local feature enhancement module LFEM based on a reparameterization operation, aiming to restore a high-resolution image from a low-resolution building image;

[0009] Step S4: Train the lightweight building image super-resolution network model based on the dilated feature modulation network constructed in step S3;

[0010] Step S5: Input the test set of images to be super-resolved into the lightweight building image super-resolution network model based on the dilated feature modulation network trained in step S4, and output a super-resolution image.

[0011] In a preferred embodiment: the specific process of step S1 is to process the original high-resolution image of the building using a bicubic interpolation algorithm to obtain a corresponding low-resolution image; and the high-resolution image HR is cropped into a low-resolution image block I LR according to the super-resolution magnification, and the corresponding high-resolution image block is denoted as I HR .

[0012] In a preferred embodiment: in step S2, different scale scaling, blurring, flipping, rotation, and CutBlur methods are used for data enhancement to improve the generalization ability of the model.

[0013] In a preferred embodiment: the dilated separable modulation unit DSMU in step S3 is designed to aggregate spatial information from different receptive fields to better model long-range feature relationships for building images with large long-range dependence; in particular, multiple dilated convolutions are used to stack equivalent large convolution kernels to reduce computational cost while obtaining the effect of a large receptive field; the specific design process of the dilated separable modulation unit is as follows:

[0014] Step S3.1: First, perform a channel segmentation operation on the normalized input feature, and the process is represented as:

[0015] [X0,X1,X2,X3]=CHUNK(X) (1)

[0016] where CHUNK(·) is a channel split operation, and [X0, X1, X2, X3] are four parts of the input feature after being split.

[0017] Step S3.2: A large convolution kernel is obtained by using multiple dilated small convolution kernels, and the hyperparameters include the convolution kernel size k, the dilation rate d, and the reconstruction kernel size K. The equivalent process in the dilated separable modulation unit is as follows:

[0018] k = (3, 3), d = (1, 2) → K = 5, k = (5, 3, 3), d = (1, 2, 3) → K = 7, k = (5, 9, 3, 3, 3), d = (1, 2, 4, 5, 7) → K = 17, k = (5, 7, 3, 3, 3), d = (1, 2, 3, 4, 5) →

[0019] K = 13; the input feature X is input into the dilated separable modulation unit for multi-scale feature extraction; finally, the multi-scale features are connected and aggregated through 1 × 1 convolution, and the process can be represented as:

[0020]

[0021] DW-Conv n×n (·) represents a deep convolution kernel component for multi-scale feature extraction, where n is 5, 7, 13, or 17; Concat(·) represents a concatenation operation along the channel dimension, Conv 1×1 (·) is a 1 × 1 convolution, represents the feature after multi-scale feature extraction, represents different scale features obtained by four different values of n, represents the result obtained by aggregating different scale features;

[0022] Step S3.3: After obtaining the aggregated multi-scale feature, we normalize it through GELU nonlinearity, and then add it to the input feature X to generate the final output, which is represented as:

[0023]

[0024] where φ(·) represents the GELU activation function, and is an element-level product.

[0025] In a preferred embodiment, the local feature enhancement module LFEM based on the reparameterization operation is specifically designed as follows:

[0026] Step S3.4: In the LFEM, first, use a 1 × 1 convolution to double the number of channels of the input feature, which is represented as:

[0027] X' = Conv1×1 (X in ) (5)

[0028] Step S3.5: Then introduce a variety of multi-branch structure in parallel with 3x3 depth convolution to encode the spatial local context information, then apply 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 1x1 convolution to obtain the output feature, which is represented as:

[0029] X out =Conv 1×1 (SE(φ(Rep Conv3×3 (X')))) (6)

[0030] Where X in and X out represent the input feature and the output feature respectively, Rep Conv3×3 represents the 3x3 convolution kernel equivalent to the multi-branch structure using the reparameterization trick, SE represents the spatial attention mechanism module, and φ(·) represents the GELU activation function.

[0031] In a preferred embodiment: the specific process of step S4 is:

[0032] Step S4.1: First define the loss function, optimize our DCFMN parameters by using the mean absolute error loss and the FFT-based frequency loss function, which is represented as:

[0033] L=||I SR -I HR ||1+γ||F(I SR )-F(I HR )||1 (7)

[0034] 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;

[0035] Step S4.2: Train two different sizes of dilated feature modulation network, named DCFMN-S and DCFMN-L respectively, where DCFMN-S uses 10 modules and DCFMN-L uses 16 modules; During the training process, each batch contains 32 LR images, and the input size is 256x256; Use the Adam optimizer with β1=0.9 and β2=0.99 to solve the proposed model; Set the initial learning rate to 1x10 -3 , and the minimum value to 1x10 -6The cosine annealing algorithm is used for learning rate updating, the weight value of the exponential moving average line EMA is 0.999, and the model uses a total of 10 6 iterations for training on an NVIDIA GeForce RTX3080 GPU.

[0036] Compared with the prior art, the present application has the following beneficial effects: 1. The present application proposes a lightweight dilated feature modulation network, which realizes a good balance between performance and efficiency through an adaptive dilated feature modulation mechanism, overcoming the trade-off problem between computational complexity and reconstruction performance of traditional super-resolution models. 2. The present application designs a dilated separable modulation unit (DSMU), which is equivalent to a large convolution kernel through multiple dilated convolutions, aggregates spatial information from different scales at a lower computational cost, and thus better models the long-range feature relationship of building images, significantly improving the detail expressiveness of image reconstruction. 3. The present application designs a local feature enhancement module (LFEM) based on a reparameterization operation, which not only effectively encodes and enhances the local context information, but also incorporates more learnable parameters during the training phase to improve the model performance, and merges parameters through reparameterization during the inference phase to avoid additional computational overhead, thereby improving the SR performance of the model without increasing the inference burden. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The overall flowchart of the lightweight building image super-resolution method based on the dilated feature modulation network of the preferred embodiment of the present application is shown in the figure;

[0038] Figure 2 The comparison chart of the local attribution map (LAM) and the diffusion index (DI) of the preferred embodiment of the present application and other efficient super-resolution models is shown in the figure;

[0039] Figure 3 The overall framework chart of the lightweight dilated feature modulation network of the preferred embodiment of the present application is shown in the figure;

[0040] Figure 4 The design chart of the dilated separable modulation unit and the local feature enhancement module based on the reparameterization operation of the preferred embodiment of the present application is shown in the figure;

[0041] Figure 5 The reparameterization block equivalent chart of the local feature enhancement module based on the reparameterization operation of the preferred embodiment of the present application is shown in the figure;

[0042] Figure 6 The performance and model complexity comparison chart of the present application and existing algorithms in the ×4 super-resolution (SR) task on the benchmark dataset is shown in the figure;

[0043] Figure 7A detail display comparison chart of the present application and existing algorithms on a x4 super-resolution (SR) task on building images. DETAILED DESCRIPTION

[0044] The present application is further described below in conjunction with the accompanying drawings and examples.

[0045] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0046] It should be noted that the terms used herein are merely for the purpose of describing specific embodiments and are not intended to limit 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 this specification, they indicate the presence of the features, steps, operations, devices, components and / or combinations thereof.

[0047] The present application proposes a lightweight image super-resolution reconstruction method based on an expansion feature modulation network, as shown in Figures 1-7 . First, the low-resolution image block I LR is input to the network, and the super-resolution result image I SR is output.

[0048] Specifically, the following steps are implemented:

[0049] Step S1: Obtain a high-resolution building image, and downsample the original high-resolution image of the building to obtain a corresponding low-resolution image;

[0050] The specific process of the downsample is: using a bicubic interpolation algorithm to process the high-resolution image to obtain a corresponding low-resolution image. According to the super-resolution magnification, the high-resolution image HR is cropped to correspond to the low-resolution image block I LR , and the corresponding high-resolution image block is denoted as I HR .

[0051] Step S2: Perform data enhancement on the low-resolution image;

[0052] The specific data enhancement operation includes using different scale scaling, blurring, flipping, rotating, and CutBlur methods for data enhancement to improve the generalization ability of the model.

[0053] Step S3: A lightweight image super-resolution network model based on the inflation feature modulation network is constructed according to the characteristics of the building image with regular texture and large long-range dependence, and the model is specifically divided into two modules, an inflation separable modulation unit and a local feature enhancement module based on the reparameterization operation, as shown in Figure 3 and Figure 4 .

[0054] As shown in Figure 3 , the inflation feature modulation network proposed by the application is composed of three modules: shallow feature extraction, deep feature extraction and up-sampling feature reconstruction.

[0055] Step S3.1: Specifically, the network first applies a convolution layer with a kernel size of 3x3 pixels to extract shallow features, converts the input LR image into a feature space, and generates shallow features f0, and the specific process can be represented as:

[0056] f0=Conv 3×3 (I LR ) (1)

[0057] Where Conv 3×3 (·) represents a 3x3 convolution layer.

[0058] Step S3.2: The shallow feature f0 is passed into the deep feature extraction module composed of a series of inflation separable modulation blocks (DSMB). The specific process can be represented as:

[0059] f k =H k (f k-1 ),k=1,…,n (2)

[0060] Where H k (·) represents the kth DSMB. f k-1 and f k represent the input feature and output feature of the kth DSMB, respectively.

[0061] Step S3.3: In order to better restore the HR target image, a global residual connection is also introduced, and finally a lightweight up-sampling layer composed of a 3x3 convolution layer and a sub-pixel convolution is used to quickly reconstruct the HR image, and the specific process can be represented as:

[0062] I SR =P θ (f k +f0) (3)

[0063] Step S3.4: Long-term context information involves more feature interaction, in order to reconstruct more fine results in dense pixel prediction, so the application reconsiders the design of the basic feature extraction unit in the deep feature extraction module. Unlike self-attention mechanism, traditional feature pyramid, large kernel convolution, the application proposes a more lightweight alternative, an expanded separable modulation unit learns long-term dependencies from different scale feature representations in order to better explore useful feature information for building image reconstruction. As shown in Figure 4 the specific design process is as follows.

[0064] Step S3.4.1: First, the normalized input features are subjected to a channel segmentation operation to generate four-part features, and then the features are input into the expanded separable modulation unit (DSMU) for multi-scale feature extraction. Given the input feature X, this process can be represented as:

[0065] [X0,X1,X2,X3]=CHUNK(X) (4)

[0066]

[0067] where CHUNK(·) is the channel segmentation operation, DW-Conv n×n (·) represents a deep convolution kernel component for multi-scale feature extraction, where n takes the values 5, 7, 13, 17.

[0068] Step S3.4.2: For a pixel on the feature map, some distant pixels may be more relevant than its neighboring pixels, so in order to reduce the computational burden, we use the properties of ordinary convolution to equivalent large convolution kernel, and use multiple expanded small convolution kernels to replace a large convolution kernel to extract features. Its hyperparameters include convolution kernel size k, expansion rate d and reconstruction kernel size K. As shown in Figure 5 the case of k=(5,7,3,3,3), d=(1,2,3,4,5) is equivalent to a convolution kernel with kernel size K=13. The remaining equivalent process in the expanded separable modulation unit is as follows: k=(3,3), d=(1,2)→K=5, k=(5,3,3), d=(1,2,3)→K=7, k=(5,9,3,3,3), d=(1,2,4,5,7)→K=17. The results in Table 3 show that this operation reduces the parameter amount and floating point operation (FLOPs) while improving performance.

[0069] Step S3.4.3: Finally, we connect these multi-scale features together and aggregate different scale information through 1x1 convolution. This process can be represented as:

[0070]

[0071] where Concat(·) denotes concatenation along the channel dimension, Conv 1×1 (·) is a 1x1 convolution.

[0072] Step S3.4.4: After obtaining the aggregated multi-scale features, we normalize them by GELU nonlinearity and add them to the input features to generate the final output. This process can be represented as:

[0073]

[0074] where φ(·) denotes the GELU function, and is the element-wise product.

[0075] Step S3.5: Although the dilated separable modulation unit can model the long-range relationship between pixels by extracting and aggregating information of different scales, it ignores the importance of local context information for super-resolution image reconstruction. Therefore, the invention designs a local feature enhancement module (LFEM) to be added after the dilated separable modulation unit to supplement the local context information. The specific design of LFEM is as follows.

[0076] Step S3.5.1: As shown in Figure 4 , in the LFEM, first use a 1x1 convolution to double the channel number of the input feature. This process can be represented as:

[0077] X' = Conv 1×1 (X in ) (8)

[0078] Step S3.5.2: Then introduce multiple multi-branch structures parallel to the 3x3 deep convolution to encode the spatial local context information, then apply the GELU function to the hidden layer for nonlinear mapping, pass the result through a SE attention module, and finally reduce the channel to the original input dimension through a 1x1 convolution to obtain the output feature. This process can be represented as:

[0079] X out = Conv 1×1 (SE(φ(Rep Conv3×3 (X')))) (9)

[0080] where X in and X out represent the input feature and the output feature respectively, Rep Conv3×3 represents the 3x3 convolution kernel equivalent to the multi-branch structure using the reparameterization trick, the SE module represents the spatial attention mechanism module, and φ(·) represents the GELU activation function.

[0081] Step S3.6: Finally, the overall architecture is designed as Figure 3As shown, the dilation feature modulation block (DSMB) consists of two parts, DSUM and LFEM. This process can be represented as:

[0082] X' = DSUM(LN(X in ))+X (10)

[0083] X out = LFEM(LN(X'))+X' (11)

[0084] where X in and X out represent the input and output feature maps of the dilation feature modulation block, LN(·) is the LayerNorm operation, and X' represents the intermediate feature.

[0085] Step S4: training the network model constructed in step 3;

[0086] 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. This process can be represented as:

[0087] L = ||I SR -I HR ||1+γ||F(I SR )-F(I HR )||1 (12)

[0088] 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.

[0089] S4.2: the present application trains two models of different sizes, named DCFMN-S and DCFMN-L. DCFMN-S uses 10 modules, while DCFMN-L uses 16 modules. During training, each batch contains 32 LR images, with an input size of 256x256. We use the Adam optimizer with β1 = 0.9 and β2 = 0.99 to solve the proposed model. We set the initial learning rate to 1x10 -3 , the minimum value to 1x10 -6 , and use the cosine annealing algorithm to update the learning rate. The weight of the exponential moving average line (EMA) is set to 0.999. In the implementation process, the model of the present application is trained for a total of 10 6 times of iteration on the NVIDIA GeForce RTX 3080 GPU

[0090] Step S5: Input the test set into the network model trained in step 3, and output the super-resolution image.

[0091] To evaluate the performance of this invention, we compared it with state-of-the-art lightweight SR methods, including SRCNN, VDSR, IMDN, EDSR-baseline, LAPAR, PAN, RFDN, ShuffleMixer, HPUN, HNCT, RepECN-S, and SAFMN. Table 1 reports the quantitative comparison results on benchmark datasets with scale factors of ×2, ×3, and ×4. In addition to the PSNR / SSIM metric, we also list the number of parameters (#params) and floating-point computation (#FLOPs), where #params relates to memory consumption, and #FLOPs relates to energy usage. #FLOPs were calculated using the fvcore1 library after super-resolution of the LR image into a 1280×720 HR image. Table 1 and... Figure 6 The results demonstrate that our proposed lightweight super-resolution method for building images based on dilatation feature modulation networks achieves a good trade-off between model complexity and reconstruction performance.

[0092] To further illustrate the effectiveness of our method, we also compared the visual effects of ×4 magnification on the Urban100 dataset. Figure 7 As shown, most of the lightweight super-resolution methods used in comparison fail to accurately recover the texture and details of architectural images, exhibiting significant blurring artifacts and distortion. In contrast, our proposed lightweight super-resolution reconstruction method for architectural images based on dilated feature modulation networks achieves the best visual quality and recovers more detailed information about the architectural structure.

[0093] In addition, to verify the effectiveness of our designed DSMU and LFEM, we further conducted ablation experiments on the DSMU and LFEM. The results in Tables 2 and 3 show that all components of the DSMU and LFEM are key to performance improvement.

[0094] Table 1: Quantitative evaluation of PSNR / SSIM performance of the method of the present invention on four relevant datasets and performance comparison with other existing methods (“” and “” are marked as best)

[0095] (Second-best performance)

[0096]

[0097]

[0098] Table 2: Ablation Experiment Verification of DSMU Module Performance in the Method of the Invention

[0099]

[0100] Table 3: Ablation experiment verification of LFEM module performance in the method of the present application

[0101]

Claims

1. A lightweight building image super-resolution reconstruction method based on an expansion feature modulation network, characterized in that, Comprise the following steps: Step S1: Obtain the high-resolution image of the building surface, and downsample the original high-resolution image of the building to obtain the corresponding low-resolution image; Step S2: Data enhancement is performed on the low-resolution image; Step S3: A lightweight building image super-resolution network model based on the inflation feature modulation network is constructed for the characteristics of regular texture and large long-range dependence in the building image, which is specifically divided into an inflation separable modulation unit DSMU and a local feature enhancement module LFEM based on a reparameterization operation, aiming to recover a high-resolution image from a low-resolution building winter image; Step S4: The lightweight building image super-resolution network model based on the inflation feature modulation network constructed in step S3 is trained; Step S5: The test set of the image to be super-resolved is input into the lightweight building image super-resolution network model based on the inflation feature modulation network trained in step S4, and a super-resolution image is output. The specific design process of the local feature enhancement module LFEM based on the reparameterization operation is: Step S3.4: In the LFEM, first use Convolution doubles the number of channels of the input features, this process is represented as: Step S3.5: Then introduce multiple branches in parallel with Deep convolutional parallel multi-branch structure to encode spatial local context information, then apply GELU function to the hidden layer for nonlinear mapping, pass the result through an SE attention module, and finally pass it through Convolution reduces the channel to the original input dimension to get the output feature, and this process is represented as: wherein, and denote input and output features, respectively, denotes a multi-branch structure which is equivalent to a convolution kernel using a reparameterization trick, SE denotes a spatial attention mechanism module, denotes an activation function.

2. The method of claim 1, wherein the method is based on a dilated feature modulation network. The specific process of step S1 is: using bicubic interpolation algorithm to process the original high-resolution image of the building to obtain the corresponding low-resolution image; according to the super-resolution magnification, the high-resolution image HR is cropped into a low-resolution image block I LR The corresponding high-resolution image block is denoted as I HR .

3. The method of claim 1, wherein the method is based on a dilated feature modulation network. In step S2, different scale scaling, blurring, flipping, rotation, and CutBlur methods are used for data enhancement to improve the generalization ability of the model.

4. The method of claim 1, wherein the method is a method of lightweight building image super-resolution reconstruction based on an expansion feature modulation network. The inflation separable modulation unit DSMU in step S3 is designed to aggregate spatial information from different receptive fields to better model long-range feature relationships. Multiple inflation convolutions are used to stack equivalent large convolution kernels to reduce computational cost while achieving large receptive field effects. The specific design process of the inflation separable modulation unit is as follows: Step S3.1: First, perform channel segmentation on the normalized input features. Given the input feature X, the process is represented as: wherein, is a channel split operation, is the four-part feature after splitting the input feature. Step S3.2: Obtain the effect of a large convolution kernel with multiple dilated small convolution kernels, whose hyperparameters include the convolution kernel size , the dilation rate , and the reconstruction kernel size ; the equivalent process in the dilated separable modulation unit is as follows: ; input features X are passed into the dilated separable modulation unit for multi-scale feature extraction; finally, these multi-scale features are connected and aggregated by convolution, which can be represented as: wherein, a deep convolution kernel component representing multi-scale feature extraction, wherein the value of n is 5, 7, 13, 17; represents a concatenation operation along the channel dimension, is convolution, represents the features after multi-scale feature extraction, represents different scale features obtained by four different values of n, represents the result obtained by aggregating different scale features; Step S3.3: After obtaining the aggregated multi-scale features, we normalize them through GELU nonlinearity, and then add them to the input features X to generate the final output. The process is represented as: wherein represents function, for element-wise multiplication.

5. The method of claim 1, wherein the method is based on a dilated feature modulation network. The specific process of step S4 is as follows: Step S4.1: First, define the loss function. By using the mean absolute error loss and the FFT-based frequency loss function, the DCFMN parameters are optimized, and the process is represented as: wherein is a high resolution Ground-truth image, denotes norm, F denotes the Fast Fourier Transform, is a weight factor; Step S4.2: Train two different sizes of dilated feature modulation network, respectively named as DCFMN-S and DCFMN-L, where DCFMN-S uses 10 modules and DCFMN-L uses 16 modules to compose; during the training process, each batch contains 32 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; the model is trained for a total of iterations on an NVIDIA GeForce RTX 3080 GPU.

Images