Image Deblurring Method Based on Multi-Scale Attention Feature Fusion

Abstract

This invention belongs to the field of image processing technology, specifically relating to an image deblurring method based on multi-scale attention feature fusion. It includes constructing a residual attention module (RAM) to capture key features using coordinate attention, increasing attention to relatively blurred areas of the image and enhancing the network's representational ability; constructing a feature pyramid structure based on attention residuals (RA-FPN) to increase detailed features in the feature map using a bottom-up pyramid structure, and achieving image deblurring through encoding and decoding; and employing a multi-scale efficient channel attention (MECA) feature fusion structure, introducing an efficient channel attention (ECA) module to increase the network's receptive field, using the attention model to capture key features, and improving the model's ability to recover image details. This invention solves the problem of poor local image deblurring results caused by feature redundancy easily generated when using residual blocks for feature extraction in SRN networks.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to an image deblurring method based on multi-scale attention feature fusion. Background Technology

[0002] Image deblurring has been applied in various fields and plays an important role in medical imaging, astronomical remote sensing, and other fields. Image deblurring technology can improve the clarity of images and restore lost information in images, which is of great research significance for visual tasks in special environments.

[0003] Currently, extensive research has been conducted both domestically and internationally on motion image deblurring techniques. Traditional deblurring methods estimate the image's blur kernel and then perform deconvolution to obtain the restored image. Xu et al. proposed using a non-natural L0 norm for image deblurring, achieving good results. Shan et al. used a piecewise function with an alternating iterative approach to estimate the blur kernel. Liu et al. proposed using the L0 norm of the image gradient to estimate the blur kernel, improving the deblurring effect by increasing the accuracy of the blur kernel. While these methods can achieve some deblurring effect, they are all highly sensitive to noise, and the image restoration results are not ideal.

[0004] With the deepening research of scholars on deep learning, deep learning has been applied to the task of image deblurring. Nah et al. proposed to use multi-scale convolutional networks to train images and use an end-to-end approach to deblur. Kupyn et al. proposed generative adversarial networks, which use VGG19-based perceptual loss to optimize the network. The image restoration effect is relatively ideal. However, adversarial networks can only take fixed-size images as input, have a long training time, and have limited ability to restore image details.

[0005] It is evident that the above method, which uses residual blocks in the existing SRN network for feature extraction, is prone to feature redundancy, resulting in poor local deblurring of the image. Therefore, further improvements are needed. Summary of the Invention

[0006] To address the aforementioned technical problem that existing SRN networks using residual blocks for feature extraction are prone to feature redundancy, resulting in poor local image deblurring performance, the present invention aims to propose an image deblurring method based on multi-scale attention feature fusion that is beneficial for improving the restoration of image details and textures.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0008] The image deblurring method based on multi-scale attention feature fusion includes the following steps:

[0009] S1. Construct a residual attention module (RAM) to capture key features using coordinate attention, improve attention to relatively blurred areas of the image, and increase the network's representation ability.

[0010] S2. Construct a feature pyramid structure based on attention residuals (RA-FPN), use the bottom-up pyramid structure to increase the detailed features in the feature map, and achieve image deblurring through encoding and decoding;

[0011] S3. Construct a multi-scale efficient channel attention (MECA) feature fusion structure, introduce an efficient channel attention (ECA) module to increase the receptive field of the network, use the attention model to capture key features, and improve the model's ability to recover image details;

[0012] S4. Based on the Feature Pyramid Structure (RA-FPN) based on Attention Residual and the Multi-Scale Efficient Channel Attention (MECA) feature fusion structure, a Multi-Scale Attention Feature Fusion Model (MAFF-SRN) is constructed to achieve image deblurring.

[0013] Preferably, in step S1, the residual attention module (RAM) reduces feature loss during convolution through residual structure, while using the coordinate attention (CA) module to focus on the channels, orientations and details of the feature map, thereby increasing attention to relatively blurred areas of the image, increasing the network's representation ability, and thus improving the network's deblurring effect.

[0014] The construction process of the residual attention module (RAM) is as follows:

[0015] S11. Input a feature M with dimensions H×W×C. c Perform a 2D convolution;

[0016] S12. Use two one-dimensional global average pooling methods to aggregate the feature layers along the height and width directions into two independent directional sensing feature maps.

[0017] S13. Perform 1×1 convolution transformation on the two feature maps respectively, and concatenate the two transformation results into channels.

[0018] S14. The ReLU activation function is used to increase its nonlinearity to obtain the intermediate feature map f;

[0019] S15. Divide the intermediate feature map f into two separate tensors along the spatial dimension. h and f w The input is transformed to the same number of channels as the input using a 1×1 convolution, and then multiplied with the original feature map to obtain the final attention output, followed by a 2D convolution.

[0020] S16. Sum the corresponding elements of the 2D convolution result with the original input to obtain the output of the residual attention module (RAM).

[0021] Preferably, in step S2, the feature pyramid structure based on attention residuals (RA-FPN) fuses shallow, high-resolution detail features with deep semantic information, thereby reducing the loss of feature information during convolution, enhancing the representation ability of high-level features, and improving the network's performance and ability to recover image details.

[0022] The construction process of the attention residual-based feature pyramid structure (RA-FPN) is as follows:

[0023] S21. Perform a convolution on the input image;

[0024] S22. Perform three residual attention module feature extractions on the convolution result, and fuse the high-resolution shallow feature map with the output of the next convolution block after 2x downsampling.

[0025] S23. Repeat steps S21 to S22 twice to obtain a top-down feature pyramid structure based on attention residuals (RA-FPN).

[0026] Furthermore, in S22, before feature fusion, a 1x1 convolution is used to adjust the number of channels so that the number of channels in the two feature maps after 2x downsampling remains consistent. Repeating this process can obtain a feature layer with rich semantics and details, thereby improving the network's feature extraction capability.

[0027] Furthermore, the method of adjusting the number of channels using 1x1 convolution specifically includes: setting the stride to 2, directly adjusting the length, width, and number of channels of the shallow feature map simultaneously, and then fusing the features by summing the corresponding elements (Add).

[0028] Preferably, in step S3, the multi-scale efficient channel attention (MECA) feature fusion structure introduces an efficient channel attention (ECA) module. This increases the network's receptive field while utilizing the attention mechanism to improve the network's feature extraction capability, thereby enhancing the model's deblurring effect. Its construction process is as follows:

[0029] S31. Convert the network input into three sizes: Input1, Input2, and Input3, which are 256×256, 512×512, and 1024×1024 respectively.

[0030] S32. The initial image restoration is achieved by encoding and decoding the input Input1 through a convolution process.

[0031] S33. Introduce an efficient channel attention (ECA) module to perform global average pooling on the preliminary image restoration result to obtain a 1-dimensional vector representing the channel;

[0032] S34. Calculate the size of the convolution kernel using an adaptive function. The formula for calculating the convolution kernel is as follows:

[0033]

[0034] Where k is the size of the convolution kernel, b is 1, γ is 2, and C is the number of channels;

[0035] S35. Calculate the weights of each channel through 1D convolution. The size of the convolution kernel is the kernel k calculated by the adaptive function. Then multiply the obtained weights with the feature map of the original input channel by channel.

[0036] S36. Upsampling is used to convert the output of the efficient channel attention into a feature map with the same dimension as the input Input2. The attention output and the input Input2 are fused by summing corresponding elements.

[0037] S37. Repeating the above S31 to S36 processes for inputs Input2 and Input3 will yield a multi-scale, efficient channel attention feature fusion structure.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] This invention, based on the proposed Residual Attention Module (RAM), utilizes the Coordinate Attention (CA) module to focus on important channels and positional information in the image, enhancing the network's attention to relatively blurred regions and improving the overall deblurring effect. Secondly, by constructing a Residual Attention-Based Feature Pyramid Structure (RA-FPN), shallow features are integrated into feature maps rich in deep semantic information, ensuring that the feature maps contain not only positional and detailed features but also semantic features, thus solving the problem of key feature loss due to convolution. Finally, by constructing a Multi-Scale Efficient Channel Attention (MECA) feature fusion structure and introducing an Efficient Channel Attention (ECA) module, the receptive field of the network is increased, while the attention model captures key features, improving the model's deblurring effect.

[0040] In summary, by adopting the technical solution of this invention, the problem of poor local image deblurring effect caused by feature redundancy due to the use of residual blocks for feature extraction in SRN networks is solved. Through testing on the GOPRO dataset, PSNR and SSIM can reach 34.84dB and 0.98 respectively, which has a great advantage in detail texture recovery ability compared with other existing methods. Attached Figure Description

[0041] Figure 1 This is a diagram of the overall network structure of the present invention in an embodiment.

[0042] Figure 2 This is a residual structure diagram from the embodiment.

[0043] Figure 3 This is a structural diagram of the coordinate attention (CA) module in the embodiment.

[0044] Figure 4 This is a structural diagram of the residual attention module (RAM) in the embodiment.

[0045] Figure 5 This is a structural diagram of the High Efficiency Channel Attention (ECA) module in the embodiment.

[0046] Figure 6 The diagram shows the structure of the Multi-Scale Efficient Channel Attention (MECA) feature fusion structure in the embodiment.

[0047] Figure 7 This is a structural diagram of the feature pyramid structure based on attention residuals (RA-FPN) in the embodiment.

[0048] Figure 8 This is a comparison diagram of the application effects of the present invention and other methods in the prior art in scenario one.

[0049] Figure 9 This is a comparison diagram showing the application effects of the present invention and other methods in the prior art in scenario two. Detailed Implementation

[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0051] Example

[0052] like Figure 1 As shown, the image deblurring method based on multi-scale attention feature fusion proposed in this embodiment includes the following steps:

[0053] S1. Construct a residual attention module (RAM) to capture key features using coordinate attention, improve attention to relatively blurred areas of the image, and increase the network's representation ability.

[0054] S2. Construct a feature pyramid structure based on attention residuals (RA-FPN), use the bottom-up pyramid structure to increase the detailed features in the feature map, and achieve image deblurring through encoding and decoding;

[0055] S3. Construct a multi-scale efficient channel attention (MECA) feature fusion structure, introduce an efficient channel attention (ECA) module to increase the receptive field of the network, use the attention model to capture key features, and improve the model's ability to recover image details;

[0056] S4. Based on the Feature Pyramid Structure (RA-FPN) based on Attention Residual and the Multi-Scale Efficient Channel Attention (MECA) feature fusion structure, a Multi-Scale Attention Feature Fusion Model (MAFF-SRN) is constructed to achieve image deblurring.

[0057] Specifically, in step S1, the residual attention module (RAM) reduces feature loss during convolution through the residual structure, while simultaneously utilizing the coordinate attention (CA) module to focus on the channels, orientations, and details of the feature map, thereby increasing attention to relatively blurred areas of the image, enhancing the network's representational ability, and ultimately improving the network's deblurring effect. The residual structure is as follows: Figure 2 As shown, the coordinate attention (CA) module is as follows: Figure 3 As shown.

[0058] Among them, such as Figure 4 As shown, the construction process of the residual attention module (RAM) is as follows:

[0059] S11. Input a feature M with dimensions H×W×C. c Perform a 2D convolution;

[0060] S12. Use two one-dimensional global average pooling methods to aggregate the feature layers along the height and width directions into two independent directional sensing feature maps.

[0061] S13. Perform 1×1 convolution transformation on the two feature maps respectively, and concatenate the two transformation results into channels.

[0062] S14. The ReLU activation function is used to increase its nonlinearity to obtain the intermediate feature map f;

[0063] S15. Divide the intermediate feature map f into two separate tensors along the spatial dimension. h and f w The input is transformed to the same number of channels as the input using a 1×1 convolution, and then multiplied with the original feature map to obtain the final attention output, followed by a 2D convolution.

[0064] S16. Sum the corresponding elements of the 2D convolution result with the original input to obtain the output of the residual attention module (RAM).

[0065] Specifically, in step S2, the feature pyramid structure based on attention residuals (RA-FPN) fuses shallow, high-resolution detail features with deep semantic information, thereby reducing the loss of feature information during convolution, enhancing the representation ability of high-level features, and improving the network's performance and ability to recover image details.

[0066] Among them, such as Figure 7 As shown, the construction process of the attention residual-based feature pyramid structure (RA-FPN) is as follows:

[0067] S21. Perform a convolution on the input image;

[0068] S22. Perform three residual attention module feature extractions on the convolution result, and fuse the high-resolution shallow feature map with the output of the next convolution block after 2x downsampling.

[0069] S23. Repeat steps S21 to S22 twice to obtain a top-down feature pyramid structure based on attention residuals (RA-FPN).

[0070] Furthermore, in S22, before feature fusion, a 1x1 convolution is used to adjust the number of channels so that the number of channels in the two feature maps after 2x downsampling remains consistent. Repeating this process can obtain a feature layer with rich semantics and details, thereby improving the network's feature extraction capability.

[0071] Furthermore, the method of adjusting the number of channels using 1x1 convolution specifically includes: setting the stride to 2, directly adjusting the length, width, and number of channels of the shallow feature map simultaneously, and then fusing the features by summing the corresponding elements (Add).

[0072] Specifically, such as Figure 6 As shown, in step S3, the Multi-Scale Efficient Channel Attention (MECA) feature fusion structure introduces an Efficient Channel Attention (ECA) module. While increasing the receptive field of the network, it utilizes the attention mechanism to improve the network's feature extraction capability, thereby enhancing the model's deblurring effect. Its construction process is as follows:

[0073] S31. Convert the network input into three sizes: Input1, Input2, and Input3, which are 256×256, 512×512, and 1024×1024 respectively.

[0074] S32. The initial image restoration is achieved by encoding and decoding the input Input1 through a convolution process.

[0075] S33, Introduce an efficient channel attention (ECA) module (such as...) Figure 5 As shown, the preliminary image restoration result is subjected to global average pooling to obtain a 1-dimensional vector representing the channel;

[0076] S34. Calculate the size of the convolution kernel using an adaptive function. The formula for calculating the convolution kernel is as follows:

[0077]

[0078] Where k is the size of the convolution kernel, b is 1, γ is 2, and C is the number of channels;

[0079] S35. Calculate the weights of each channel through 1D convolution. The size of the convolution kernel is the kernel k calculated by the adaptive function. Then multiply the obtained weights with the feature map of the original input channel by channel.

[0080] S36. Upsampling is used to convert the output of the efficient channel attention into a feature map with the same dimension as the input Input2. The attention output and the input Input2 are fused by summing corresponding elements.

[0081] S37. Repeating the above S31 to S36 processes for inputs Input2 and Input3 will yield a multi-scale, efficient channel attention feature fusion structure.

[0082] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0083] This invention, based on the proposed Residual Attention Module (RAM), utilizes a Coordinate Attention module to focus on important channels and positional information in the image, enhancing the network's attention to relatively blurred regions and improving the overall deblurring effect. Secondly, by constructing a Residual Attention-based Feature Pyramid Structure (RA-FPN), shallow features are integrated into feature maps rich in deep semantic information, ensuring that the feature maps contain not only positional and detailed features but also semantic features, thus solving the problem of key feature loss due to convolution. Finally, by constructing a Multi-Scale Efficient Channel Attention (MECA) feature fusion structure and introducing an Efficient Channel Attention (ECA) module, the receptive field of the network is increased, while the attention model captures key features, improving the model's deblurring effect.

[0084] In summary, by adopting the technical solution of this invention, the problem of poor local image deblurring effect caused by feature redundancy due to the use of residual blocks for feature extraction in SRN networks is solved. Through testing on the GOPRO dataset, PSNR and SSIM can reach 34.84dB and 0.98 respectively, which has a great advantage in detail texture recovery ability compared with other methods.

[0085] To verify the above technical solution, this embodiment conducts the following experiments in the following manner:

[0086] Experimental environment and dataset:

[0087] Using the GoPro public dataset provided by Nah et al. in 2017, experiments were conducted using 2103 clear and blurry street view images. The computer's GPU configuration was an RTX 2080 Ti with 11GB of video memory. The Adam optimizer was used to optimize the loss function, with momentum parameters of 0.9 and 0.999, and a learning rate of 0.000001.

[0088] Evaluation indicators:

[0089] To verify the deblurring effect of the network model proposed in this embodiment, several other classic deblurring algorithms were selected for comparison, and peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) were used as evaluation criteria for the deblurring effect.

[0090] Given a sharp image I of size m×n and a deblurred image K, the mean square error (MSE) is as follows:

[0091]

[0092] The peak signal-to-noise ratio (PSNR) is shown in the following formula:

[0093]

[0094] Structural similarity (SSIM) is a metric that measures the similarity between two images, and its structure is shown in the following formula:

[0095]

[0096] Where, μ x It is the average value of x, μ y It is the average value of y. It is the variance of x. σ is the variance of y. xy Let c1 be the covariance of x and y, and c1 = (k1L). 2 c2=(k2L) 2These are constants used to maintain stability, L is the dynamic range of pixel values, k1 = 0.01, k2 = 0.03.

[0097] Experimental results and analysis:

[0098] On the GOPRO Street View dataset, this embodiment is compared with eight existing classic algorithms. Table 1 shows the test results of different deblurring algorithms.

[0099] Table 1 Evaluation results of the deblurring algorithm for the GOPRO dataset.

[0100] algorithm PSNR (dB) SSIM Pan et al. 23.50 0.84 Sun 24.64 0.84 Deblur GAN 26.12 0.89 Deblur GAN_v2 27.34 0.87 SRN 32.35 0.96 BANet 33.68 0.97 MIMO_UNet 33.96 0.97 MPRnet 34.42 0.97 This embodiment 34.84 0.98

[0101] The image deblurring method based on feature fusion SRN is compared with other advanced algorithms on the GoPro Street View dataset. Table 1 shows the PSNR, SSIM, and time comparison results of several algorithms. Among them, SRN, BANet, MIMO_UNet_main, and MPRnet are advanced algorithms in the field of image deblurring in recent years. As can be seen from Table 1, the deblurring method proposed in this embodiment achieves good results in both PSNR and SSIM. Compared with the original SRN algorithm, PSNR is improved by 2.49 dB, and SSIM is improved by 0.02.

[0102] Figure 8 , Figure 9 This is a comparison chart of the deblurring effects of the Deblur GAN_V2, SRN, BANet, MIMO_UNet, MPRnet algorithms and this embodiment. From... Figure 8 , Figure 9 It is evident that the Deblur GAN_V2 algorithm has the worst deblurring effect, with the license plate, headlights, and fabric images appearing blurry. The MIMO_UNet, BANet, SRN, and MPRnet algorithms perform better than Deblur GAN_V2 in image restoration. The deblurring effect of this embodiment is superior to other methods, and it largely avoids the distortion issues present in the original SRN algorithm. In summary, while the MIMO_UNet, BANet, SRN, and MPRnet algorithms can achieve some effect in restoring image texture and detail, they still have some shortcomings. The method proposed in this application provides the best deblurring and achieves clearer and more realistic restoration of image details and textures.

[0103] ablation experiment

[0104] To demonstrate the effectiveness of each module in this embodiment, ablation experiments were conducted using the GOPRPO dataset under the same experimental conditions. The experimental results are shown in Table 2:

[0105] Table 2 Ablation Experiment Results

[0106] ARM FPN CA PSNR SSIM 32.35 0.96 √ 33.52 0.97 √ 33.72 0.97 √ 33.64 0.97 √ √ √ 34.84 0.98

[0107] As shown in Table 2, the PSNRs after adding the Attention Convolutional Module (ARM), Feature Pyramid Module (FPN), and Multi-Scale Coordinate Attention Module (CA) are 33.52, 33.72, and 33.64, respectively, while the SSIMs are 0.97, 0.97, and 0.97. Compared with the original SRN network, both PSNR and SSIM show some improvement. After fusing the three structures, the PSNR and SSIM are 34.84 and 0.98, respectively, achieving a good result.

[0108] The embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. An image deblurring method based on multi-scale attention feature fusion, characterized in that, Includes the following steps: S1. Construct a residual attention module (RAM) to capture key features using coordinate attention, thereby increasing the attention to relatively blurred areas of the image and enhancing the network's representational ability. The residual attention module (RAM) reduces feature loss during convolution through residual structures, while using a coordinate attention (CA) module to focus on the channels, orientations, and details of the feature map. Furthermore, it uses two one-dimensional global average pooling methods to aggregate the feature layer along the height and width directions into two independent orientation-aware feature maps, thereby increasing the attention to relatively blurred areas of the image, enhancing the network's representational ability, and thus improving the network's deblurring effect. S2. Construct a feature pyramid structure based on attention residuals (RA-FPN), which uses a bottom-up pyramid structure to increase the detail features in the feature map and achieves image deblurring through encoding and decoding; the feature pyramid structure based on attention residuals (RA-FPN) fuses the shallow, high-resolution detail features with deep semantic information, thereby reducing the loss of feature information during convolution, strengthening the representation ability of high-level features, and improving the network performance and the ability to recover image details; S3. Construct a multi-scale efficient channel attention (MECA) feature fusion structure, introduce an efficient channel attention (ECA) module to increase the receptive field of the network, use the attention model to capture key features, and improve the model's ability to recover image details; S4. Based on the Feature Pyramid Structure (RA-FPN) based on Attention Residual and the Multi-Scale Efficient Channel Attention (MECA) feature fusion structure, a Multi-Scale Attention Feature Fusion Model (MAFF-SRN) is constructed to achieve image deblurring.

2. The image deblurring method according to claim 1, characterized in that, The construction process of the residual attention module (RAM) is as follows: S11, Input a dimension as Features Perform a 2D convolution; S12. Use two one-dimensional global average pooling methods to aggregate the feature layers along the height and width directions into two independent directional sensing feature maps. S13. Perform separate operations on the two feature maps. Perform convolution transformation and concatenate the two transformation results into one channel; S14. Use the ReLU activation function to increase its nonlinearity to obtain intermediate feature maps. ; S15. Transfer the intermediate feature map Divided along the spatial dimension into two separate tensors and and utilize The convolution transforms it to the same number of channels as the input, then multiplies it with the original feature map to obtain the final attention output, followed by a 2D convolution. S16. Sum the corresponding elements of the 2D convolution result with the original input to obtain the output of the residual attention module (RAM).

3. The image deblurring method according to claim 1, characterized in that, The construction process of the attention residual-based feature pyramid structure (RA-FPN) is as follows: S21. Perform a convolution on the input image; S22. Perform three residual attention module feature extractions on the convolution result, and fuse the high-resolution shallow feature map with the output of the next convolution block after 2x downsampling. S23. Repeat steps S21 to S22 twice to obtain a top-down feature pyramid structure based on attention residuals (RA-FPN).

4. The image deblurring method according to claim 3, characterized in that, In S22, before feature fusion, a 1x1 convolution is used to adjust the number of channels so that the number of channels in the two feature maps after 2x downsampling remains consistent. Repeating this process can obtain a feature layer with rich semantics and details, thereby improving the network's feature extraction capability.

5. The image deblurring method according to claim 4, characterized in that, The method of adjusting the number of channels using 1x1 convolution specifically includes: setting the stride to 2, directly adjusting the length, width, and number of channels of the shallow feature map simultaneously, and then fusing the features by summing the corresponding elements (Add).

6. The image deblurring method according to claim 1, characterized in that, In step S3, the Multi-Scale Efficient Channel Attention (MECA) feature fusion structure introduces an Efficient Channel Attention (ECA) module. This increases the network's receptive field while enhancing its feature extraction capabilities through an attention mechanism, thereby improving the model's deblurring performance. The construction process is as follows: S31. Convert the network input to... Three input sizes: Input1, Input2, and Input3; S32. The initial image restoration is achieved by encoding and decoding the input Input1 through a convolution process. S33. Introduce an efficient channel attention (ECA) module to perform global average pooling on the preliminary image restoration result to obtain a 1-dimensional vector representing the channel; S34. Calculate the size of the convolution kernel using an adaptive function. The formula for calculating the convolution kernel is as follows: in, The size of the convolution kernel. =1, It is 2. Number of channels; S35. Calculate the weights of each channel using 1D convolution, with the kernel size being the kernel size calculated using the adaptive function. Then, the obtained weights are multiplied channel by channel with the feature map of the original input. S36. Upsampling is used to convert the output of the efficient channel attention into a feature map with the same dimension as the input Input2. The attention output and the input Input2 are fused by summing corresponding elements. S37. Repeating the above S31~S36 process for inputs Input2 and Input3 will yield a multi-scale efficient channel attention feature fusion structure.

Images