Image moiré noise removal method based on multi-scale dynamic feature aggregation

Abstract

The present invention discloses a method for removing image moiré noise based on multi-scale dynamic feature aggregation. The method comprises constructing a residual dense block, a scale feature extraction and fusion module, and a main bypass dynamic coding module. The residual dense block and the scale feature extraction and fusion module are fused into a combined module, and a jump connection is adopted for the combined module. The main bypass dynamic coding module structure adopts a residual connection. The combined module and the main bypass dynamic coding module are coupled in a preset manner to form a deep convolutional neural network. The beneficial effects of the present invention are as follows: in response to the insufficient multi-scale feature fusion of moiré noise, a multi-scale feature extraction and fusion module is proposed to better utilize multi-scale feature information and better realize information exchange and fusion. At the same time, in response to the dynamic texture characteristics of moiré noise, a main bypass dynamic coding module is proposed to specifically encode sample-adaptive moiré noise. Finally, the two modules are integrated into a framework to achieve moiré noise removal.

Description

Technical Field

[0001] The present invention relates to the technical field of image denoising and restoration, and in particular to a method for removing image moiré noise based on multi-scale dynamic feature aggregation. Background Art

[0002] Electronic screens are a ubiquitous information medium, used to convey computer visual information to users. To instantly record and share information, capturing screen content using software or portable cameras like smartphones has become commonplace. However, these captured screen images are often affected by moiré noise, which degrades their visual quality. Moiré noise, unlike common noise introduced by sensors and other components, is caused by widespread frequency aliasing between the camera's color filter array (CFA) and the screen's LCD subpixels. Moiré noise manifests as multi-scale patterns such as stripes, ripples, and waves, accompanied by spatially varying color distortion, severely degrading image quality. Moiré is also a common problem in scanners and printing. Demoiré technology can help reduce noise and interference, improving clarity and detail in scans and prints. In medical imaging, CT scans, MRIs, and digital X-rays often produce unwanted moiré, impacting image quality and diagnostic results. Demoiré technology can improve image quality and readability, facilitating accurate analysis and diagnosis of medical images. Therefore, removing moiré noise from images has great practical significance.

[0003] Image moiré removal methods fall into two main categories: pre-processing, which reduces moiré by adding special optical filters or lenses to the camera and adjusting screen settings to change color temperature, brightness, and other factors. Post-processing, however, uses specific algorithms to eliminate moiré noise from images. Given the convenience and immediacy of capturing screen content, as well as the widespread demand and cost-effectiveness of moiré removal, it's more appropriate to embed the second type of moiré removal method into the device's processing module via software.

[0004] Current moiré denoising algorithms have two shortcomings. First, there is a lack of communication and fusion of multi-scale information. Although most methods have noticed the multi-scale characteristics of moiré noise and adopted corresponding multi-scale network architectures or modules, these structures and modules are very rough in processing multi-scale information. Common methods include direct splicing and addition of multi-scale features, lacking communication and fusion between information; second, there is a lack of encoding of the dynamic texture of moiré noise. Although some people have begun to make certain attempts to address this feature, the de-moiré noise effect is not satisfactory. Summary of the Invention

[0005] In response to the above problems, the present invention proposes an image moiré noise removal method based on multi-scale dynamic feature aggregation, which mainly solves the problem that the existing moiré denoising algorithm lacks the communication and fusion of multi-scale information and the encoding of dynamic texture, resulting in poor moiré noise removal effect.

[0006] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0007] A method for removing image moiré noise based on multi-scale dynamic feature aggregation, comprising the following steps:

[0008] Selecting a de-moiré noise dataset, and generating image pairs for training according to properties of the de-moiré noise dataset;

[0009] Separately constructing a residual dense block, a scale feature extraction and fusion module, and a main bypass dynamic encoding module, fusing the residual dense block and the scale feature extraction and fusion module into a combined module, and adopting a skip connection for the combined module. The main bypass dynamic encoding module structure adopts a residual connection, and the combined module and the main bypass dynamic encoding module are coupled in a preset manner to form a deep convolutional neural network;

[0010] Inputting the moiré-noised image in the image pair into the deep convolutional neural network, and outputting a moiré-noised image;

[0011] Calculating a loss function between the de-moiré noised image and the clear image in the image pair, and training the deep convolutional neural network;

[0012] A test set is selected from the de-moiré noise dataset, and the moiré noise images in the test set are input into the trained deep convolutional neural network, and the corresponding de-moiré noise images are output.

[0013] The present invention has the following beneficial effects: To address the lack of multi-scale feature fusion in moiré noise, a multi-scale feature extraction and fusion module is proposed to better utilize multi-scale feature information and achieve better information exchange and fusion. Furthermore, a main-bypass dynamic encoding module is proposed to specifically encode sample-adaptive moiré noise, targeting the dynamic texture characteristics of moiré noise. Ultimately, these two modules are integrated into a single framework to achieve moiré noise removal. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A schematic flow chart of a method for removing image moiré noise based on multi-scale dynamic feature aggregation disclosed in an embodiment of the present invention;

[0015] Figure 2 This is a schematic diagram of the structure of the residual dense block and multi-scale feature extraction fusion module disclosed in an embodiment of the present invention;

[0016] Figure 3 This is a schematic structural diagram of the main bypass dynamic encoding module disclosed in an embodiment of the present invention;

[0017] Figure 4 This is a schematic diagram of the improved DDF module structure disclosed in an embodiment of the present invention;

[0018] Figure 5 A schematic diagram of the structure of a deep convolutional neural network disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the present invention more clear and distinct, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all of the present invention.

[0020] This embodiment proposes a method for removing image moiré noise based on multi-scale dynamic feature aggregation. To address the limitations of multi-scale feature fusion in moiré noise, a multi-scale feature extraction and fusion module is proposed to better utilize multi-scale feature information and achieve better information exchange and fusion. Furthermore, a main-bypass dynamic encoding module is proposed to specifically encode sample-adaptive moiré noise, targeting the dynamic texture characteristics of moiré noise. Ultimately, these two modules are integrated into a single framework to achieve moiré noise removal.

[0021] like Figure 1 As shown, the following steps S1-S5 are included:

[0022] S1, select a de-moiré noise dataset and generate image pairs for training according to the properties of the de-moiré noise dataset.

[0023] In this embodiment, the above-mentioned moiré noise removal dataset is an image dataset publicly available on the Internet, which is specifically used for the model for moiré noise removal. Images with moiré noise and corresponding clear images are selected from the image dataset, and image area blocks of size r×r at the same position are cropped and then scaled according to specific circumstances.

[0024] In this embodiment, alternative datasets for de-moiré noise removal can be UHDM, TIP2018, and LCDMoire. Specifically, UHDM and TIP2018 are real datasets, and may experience misalignment between moiré-stained images and corresponding clear images. LCDMoire, on the other hand, is a synthetic dataset and does not exhibit this misalignment. Real datasets tend to have richer and more complex data distributions and content, closer to reality. When UHDM is selected, since UHDM images have a 4K resolution and no border around them for image positioning, an image region of r = 768 is randomly cropped and no scaling is performed. When TIP2018 is selected, since the TIP2018 image has a pure white / black border around it for image positioning, the central two-thirds of the image is cropped and scaled to a 256×256 image. When LCDMoire is selected, an image region of r = 512 is randomly cropped and no scaling is performed.

[0025] S2, respectively constructs the residual dense block, the scale feature extraction fusion module and the main bypass dynamic coding module, fuses the residual dense block and the scale feature extraction fusion module into a combined module, and adopts jump connection for the combined module. The main bypass dynamic coding module structure adopts residual connection. The combined module and the main bypass dynamic coding module are coupled in a preset manner to form a deep convolutional neural network.

[0026] In this embodiment, the residual dense block extracts image features and fully utilizes the hierarchical features of all convolutional layers, while skip connections (residuals) address the vanishing / exploding gradient problem in gradient propagation. The reason for combining it with the scale feature extraction and fusion module is that the residual dense block can extract rich image feature information, while the scale feature extraction and fusion module can further obtain and fuse the content of different scales in this feature information.

[0027] Specifically, S2 can be split into S21-S24:

[0028] S21, construct a residual dense block: the residual dense block includes m layers of convolutional layers, the convolutional layers of the first m-1 layers include filters and activation functions corresponding to the number of scale channels, and the convolutional layers of the mth layer include filters corresponding to the number of scale channels, wherein the input of the convolutional layer of the first layer is an image pair, and the input of the convolutional layers of other layers is the output of the previous convolutional layer and the concatenation of the image pairs along the scale channel dimension, and the first output value of the output of the mth convolutional layer is input to the input end of the scale feature extraction and fusion module.

[0029] See Figure 2 , the residual dense block is as follows Figure 2The dotted box on the left shows the structure of four convolutional layers. In one example, for the dense residual block in the first scale, the first layer is a convolution layer with 3×3×48 filters and an activation function PReLU corresponding to the number of scale channels, and its input is the module input; the second layer is a convolution layer with 3×3×96 filters and an activation function PReLU corresponding to the number of scale channels, and its input is the concatenation of the output of the first convolution layer and the module input along the channel dimension; the third layer is a convolution layer with 3×3×192 filters and an activation function PReLU corresponding to the number of scale channels, and its input is the concatenation of the outputs of the first and second convolution layers and the module input along the channel dimension; the fourth layer is a convolution layer, and its input is the concatenation of the outputs of the first three layers and the module input along the channel dimension; the output of the fourth convolution layer will be input into the scale feature extraction fusion module, and then its output is added to the input of the residual dense block to obtain the final output. The number of scale channels of the first scale is 48, the number of scale channels of the second scale is 96, and the number of scale channels of the third scale is 192.

[0030] S22, construct a scale feature extraction and fusion module: The scale feature extraction and fusion module includes n parameter-sharing dilated convolution layers. Each dilated convolution layer uses the same convolution parameters and different dilated coefficients. The dilated convolution layer uses convolution operation to capture scale feature information of different scales but similar shapes, and then concatenates the scale feature information with the first output value along the channel dimension. The scale feature information of different scales is preliminarily fused using 1×1 convolution and an activation function. The channel attention mechanism is then used for the preliminarily fused scale feature information. Finally, the 1×1 convolution and activation function are used again to complete the final information fusion and output the second output value.

[0031] Continue reading Figure 2 , scale feature extraction and fusion module such as Figure 2 As shown in the dashed box on the right. In one example, the scale feature extraction and fusion module uses parameter-sharing dilated convolution to capture feature information of different scales but similar shapes through convolution operations with the same set of convolution parameters but different dilation coefficients. This feature information is then concatenated with the input along the channel dimension. In this example, the dilated convolution coefficients are selected as 1, 6, 12, 18, and 24. A 1×1 convolution is used to initially fuse the feature information of different scales, and the activation function PReLU is used. The channel attention mechanism (CA) is then applied to the preliminary fused information. Finally, 1×1 convolution and the activation function PReLU are used again to complete the final information fusion. The multi-scale feature extraction and fusion module makes more effective use of feature information of different spatial scales and reduces the requirements for model parameters and computational performance by sharing parameters.

[0032] In summary, the operation process of the scale feature extraction and fusion module is as follows:

[0033] Di =Conv i (I)

[0034] O=Conv(P(CA(Conv(P([I,D i1 ,D i2 ,…])))))

[0035] Where I represents the input of the scale feature extraction and fusion module, O represents the output of the scale feature extraction and fusion module, Conv i Indicates the use of dilated convolution with dilated coefficient i, D i Indicates the use of dilated convolution Conv with a dilated coefficient of i i The output of , Conv represents 1×1 convolution, P represents PReLU activation function, CA represents channel attention mechanism, and [...] represents the concatenation of features in the brackets along the channel dimension.

[0036] S23, combines the residual dense block and the scale feature extraction fusion module and uses a jump connection for the whole: the second output value output by the scale feature extraction fusion module is added to the image pair to obtain the final output, and the final output is input to the main bypass dynamic encoding module.

[0037] S24, construct a main-bypass dynamic coding module: the main-bypass dynamic coding module includes S dynamic coding modules stacked in sequence, and convolution layers respectively coupled to the front and rear stages of the dynamic coding modules. Each dynamic coding module includes a main path and a bypass. For the first S-1 dynamic coding modules, each main path includes a main path convolution layer and a main path DDF module coupled after the main path convolution layer. Each bypass includes a bypass convolution layer and a bypass DDF module coupled after the bypass convolution layer. The main path of the Sth dynamic coding module includes a main path convolution layer and a main path DDF module coupled after the main path convolution layer. The bypass of the Sth dynamic coding module includes only one convolution layer. The main path DDF module and the bypass DDF module of the front stage are respectively coupled with the main path convolution layer and the bypass convolution layer of the rear stage.

[0038] See Figure 3 The main bypass dynamic coding module adopts residual connection as a whole, and its residual processing is convolution plus PReLU activation function, several dynamic coding modules (such as Figure 3 The structure is shown in the dotted box), convolution plus PReLU activation function, and stacked in sequence. In this example, three dynamic encoding modules are used to participate in the stacking. The dynamic encoding module is a dual-input dual-output module, with inputs as the main input and bypass input, and outputs as the main output and bypass output. The internal structure consists of two parts: the main path and the bypass. The main path structure is a convolution layer followed by a modified DDF module, as shown in the following figure. Figure 4As shown. The bypass structure is a convolution layer followed by a DDF module. In addition to being input into the DDF module behind the bypass, the output of the bypass convolution layer is also input into the improved DDF module of the main path to generate a dynamic filter to process the output features of the convolution layer of the main path, thereby encoding the dynamic texture of the moiré noise. It should be noted that for the first dynamic encoding module, its dual inputs are the same. For the last dynamic encoding module, since only the main path output is taken, there is no need to calculate the bypass DDF module. That is, for any of the first S-1 dynamic encoding modules, the bypass convolution layer of the same dynamic encoding module is also coupled with the main path DDF module.

[0039] In summary, the operation process of the dynamic encoding module is:

[0040] O m =DDF(P(Conv m (I m )),P(Conv b (I b )))

[0041] O b =DDF(P(Conv b (I b )),P(Conv b (I b )))

[0042] Where, O m Indicates the main output of the dynamic encoding module, O b Indicates the bypass output of the dynamic encoding module, I m Indicates the main input of the dynamic encoding module, I b Represents the bypass input of the dynamic encoding module, Conv m Represents the convolution layer of the main path, Conv b represents the bypass convolutional layer, P represents the PReLU activation function, and DDF represents the DDF module.

[0043] Compared with the simple encoding and mapping of image features using AdaIn, the present invention proposes a main-bypass dynamic encoding module with better dynamic encoding capabilities. It adopts a main-bypass dual-branch structure. The main path is responsible for capturing and processing image features, and the bypass is responsible for generating dynamic convolution kernel features. Due to the use of dynamic convolution, the module can generate a dynamic convolution kernel suitable for the specific position of the sample according to the input, and has sample adaptability and local adaptability.

[0044] S25, construct a deep convolutional neural network, which includes a residual dense block, a scale feature extraction and fusion module, and a main bypass dynamic encoding module. The input is an image block with moiré noise. After a series of linear and nonlinear operations such as modules, convolutions, and activations in the network, the image block output with moiré noise removed is obtained.

[0045] The deep convolutional neural network includes t layers of scales, and the combination module includes a first combination module and a second combination module. Each layer of scale includes a first combination module, a main bypass dynamic encoding module, a second combination module and a separate convolution layer that are coupled in sequence. The scale feature information output by the first combination module of the previous layer is input into the first combination module of the next layer through bilinear interpolation down-sampling. The main bypass dynamic encoding module of each layer outputs the processed scale feature information of the current layer, and the scale feature information is respectively input into the corresponding second combination module. The scale feature information output by the second combination module of the next layer is input into the second combination module of the previous layer through bilinear interpolation up-sampling. The scale feature information of the current layer and the scale feature information output by the second combination module of the next layer are spliced ​​along the channel dimension and input into the second combination module of each layer to obtain the scale feature information output by the second combination module, which is input into the separate convolution layer of the current scale and outputs the corresponding de-moiré noise image.

[0046] S3, inputs the moiré-noised image in the image pair into a deep convolutional neural network and outputs a moiré-noised image.

[0047] See Figure 5In one example, the deep convolutional neural network (DCNN) structure is as follows: First, an image of W×H×c (c is 3 in color images) is input. Preliminary image processing is performed using a pixel unshuffle downsampling algorithm, a convolutional layer consisting of 48 3×3×3 filters, and a PReLU activation function. The resulting image is then fed into a residual dense block and a scale feature extraction and fusion module to produce the first-layer scale feature information, which has 48 channels. The first-layer scale feature information is then downsampled using bilinear interpolation and fed into a residual dense block and a scale feature extraction and fusion module to produce the second-layer scale feature information, which has 96 channels. Finally, the second-layer scale feature information is downsampled using bilinear interpolation and fed into a residual dense block and a scale feature extraction and fusion module to produce the third-layer scale feature information, which has 192 channels. Each layer of scale feature information is then passed through its own main bypass dynamic encoding module to produce the corresponding processed feature information. The processed third-layer scale feature information is then fed into a residual dense block and a scale feature extraction and fusion module to produce the third-layer scale output, which has 192 channels. The second-level scale features of the third-level scale output, which are processed by bilinear interpolation and upsampling, are then concatenated along the channel dimension and fed into a residual dense block and a scale feature extraction fusion module to produce the second-level scale output, which has 96 channels. The first-level scale features of the second-level scale output, which are processed by bilinear interpolation and upsampling, are then concatenated along the channel dimension and fed into a residual dense block and a scale feature extraction fusion module to produce the first-level scale output, which has 48 channels. Finally, each of the three to one-level scale outputs is processed through a convolutional layer consisting of 192 3×3×12 filters, a convolutional layer consisting of 96 3×3×12 filters, and a convolutional layer consisting of 48 3×3×12 filters, followed by a pixel shuffle upsampling algorithm. This results in three output images of different sizes. The largest output image represents a de-moiré-denoised image of the same size as the input image. The remaining two images are de-moiré-denoised images of smaller spatial dimensions and are used for loss function calculation.

[0048] S4, calculates the loss function between the de-moiré noise image and the clear image in the image pair, and trains the deep convolutional neural network.

[0049] Specifically, the loss function and gradient are calculated based on the three output images and the corresponding clear images, and the Adam back-propagation algorithm is used for training to train the network parameters of each module and layer. The loss function calculation formula is as follows:

[0050]

[0051] Loss total (O1,O2,O3,G)=Loss(O1,G)+Loss(O2,G↓ )+Loss(O3,G ↓↓ )

[0052] In the formula, φ(·) represents the deep convolutional neural network VGG16, · represents multiplication, ||·|| represents the Euclidean norm, G ↓ Indicates that bilinear interpolation downsampling of G is performed to make G ↓ The size is consistent with O2, G ↓↓ Indicates that bilinear interpolation downsampling of G is performed to make G ↓↓ The size is consistent with O3, Loss total is the final loss function.

[0053] S5, select a test set from the de-moiré noise dataset, input the moiré noise images in the test set into the trained deep convolutional neural network, and output the corresponding de-moiré noise images.

[0054] Specifically, the moiré noise image in the test dataset is input into the trained deep convolutional neural network, and three corresponding moiré noise-removed images of different sizes are output. The largest one is taken as the moiré noise removal result.

[0055] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made based on the essence of the present invention are intended to be covered by the scope of protection of the present invention.

Claims

1. A method for removing image moiré noise based on multi-scale dynamic feature aggregation, characterized in that: The following steps are involved: Selecting a de-moiré noise dataset, and generating image pairs for training according to properties of the de-moiré noise dataset; Separately constructing a residual dense block, a scale feature extraction and fusion module, and a main bypass dynamic encoding module, fusing the residual dense block and the scale feature extraction and fusion module into a combined module, and adopting a skip connection for the combined module. The main bypass dynamic encoding module structure adopts a residual connection, and the combined module and the main bypass dynamic encoding module are coupled in a preset manner to form a deep convolutional neural network; Inputting the moiré-noised image in the image pair into the deep convolutional neural network, and outputting a moiré-noised image; Calculating a loss function between the de-moiré noised image and the clear image in the image pair, and training the deep convolutional neural network; Selecting a test set from the de-moiré noise dataset, inputting the moiré noise images in the test set into the trained deep convolutional neural network, and outputting corresponding de-moiré noise images; The residual dense block includes m convolutional layers, the first m-1 convolutional layers include filters and activation functions corresponding to the number of scale channels, and the m-th convolutional layer includes filters corresponding to the number of scale channels, wherein the input of the first convolutional layer is the image pair, and the input of the convolutional layers of other layers is the output of the previous convolutional layer and the concatenation of the image pair along the scale channel dimension, and the first output value of the m-th convolutional layer is input to the input end of the scale feature extraction and fusion module; The scale feature extraction and fusion module includes n parameter-sharing dilated convolution layers, each of which uses the same convolution parameters and different dilated coefficients. The dilated convolution layer uses a convolution operation to capture scale feature information of different scales but similar shapes, then concatenates the scale feature information with the first output value along the channel dimension, uses 1×1 convolution to preliminarily fuse the scale feature information of different scales and uses an activation function, then uses a channel attention mechanism for the preliminarily fused scale feature information, and finally uses 1×1 convolution and activation function again to complete the final information fusion and output a second output value; The second output value output by the scale feature extraction and fusion module is added to the image pair to obtain a final output, and the final output is input to the main bypass dynamic encoding module; The main-bypass dynamic encoding module includes S dynamic encoding modules stacked in sequence, and convolution layers respectively coupled to the front and rear stages of the dynamic encoding modules. Each of the dynamic encoding modules includes a main path and a bypass path. For the first S-1 dynamic encoding modules, each main path includes a main path convolution layer and a main path DDF module coupled after the main path convolution layer. Each bypass path includes a bypass convolution layer and a bypass DDF module coupled after the bypass convolution layer. The main path of the Sth dynamic encoding module includes a main path convolution layer and a main path DDF module coupled after the main path convolution layer. The bypass of the Sth dynamic encoding module includes only one convolution layer. The main path DDF module and the bypass DDF module of the front stage are respectively coupled with the main path convolution layer and the bypass convolution layer of the rear stage. For any of the first S-1 dynamic encoding modules, the bypass convolution layer of the same dynamic encoding module is also coupled with the main DDF module; The deep convolutional neural network includes t scale layers. The combination module includes a first combination module and a second combination module. Each scale layer includes a first combination module, a main bypass dynamic encoding module, a second combination module, and a separate convolution layer, which are coupled in sequence. The scale feature information output by the first combination module of the previous layer is down-sampled by bilinear interpolation and input into the first combination module of the next layer. The main bypass dynamic encoding module of each layer outputs the processed scale feature information of the current layer, and the scale feature information is respectively input into the corresponding second combination module. The scale feature information output by the second combination module of the next layer is up-sampled by bilinear interpolation and input into the second combination module of the previous layer. The scale feature information of the current layer and the scale feature information output by the second combination module of the next layer are spliced ​​along the channel dimension and input into the second combination module of each layer to obtain the scale feature information output by the second combination module. The scale feature information is input into the separate convolution layer of the current scale, and the corresponding de-moiré noise image is output.

2. The method for removing image moiré noise based on multi-scale dynamic feature aggregation according to claim 1, wherein: The operation process of the scale feature extraction and fusion module is as follows: D i =Conv i (I) Where I represents the input of the scale feature extraction and fusion module, O represents the output of the scale feature extraction and fusion module, Conv i Indicates the use of dilated convolution with dilated coefficient i, D i Indicates the use of dilated convolution Conv with a dilated coefficient of i i The output of , Conv represents 1×1 convolution, P represents PReLU activation function, CA represents channel attention mechanism, and [...] represents the concatenation of features in the brackets along the channel dimension.

3. The method for removing image moiré noise based on multi-scale dynamic feature aggregation according to claim 1, wherein: The calculation process of the dynamic encoding module is: O m =DDF(P(Conv m (I m )),P(Conv b (I b ))) O b =DDF(P(Conv b (I b )),P(Conv b (I b ))) Where, O m Indicates the main output of the dynamic encoding module, O b Indicates the bypass output of the dynamic encoding module, I m Indicates the main input of the dynamic encoding module, I b Represents the bypass input of the dynamic encoding module, Conv m Represents the convolution layer of the main path, Conv b represents the bypass convolutional layer, P represents the PReLU activation function, and DDF represents the DDF module.

4. The method for removing image moiré noise based on multi-scale dynamic feature aggregation according to claim 1, wherein: The calculation formula of the loss function is as follows: Loss total (O1,O2,O3,G)=Loss(O1,G)+Loss(O2,G ↓ )+Loss(O3,G ↓↓ ) In the formula, φ(·) represents the deep convolutional neural network VGG16, · represents multiplication, ||·|| represents the Euclidean norm, G ↓ Indicates that bilinear interpolation downsampling of G is performed to make G ↓ The size is consistent with O2, G ↓↓ Indicates that bilinear interpolation downsampling of G is performed to make G ↓↓ The size is consistent with O3, Loss total is the final loss function.

Images