An image denoising method, device, computer equipment and storage medium

Abstract

The application discloses an image denoising method and device, computer equipment and a storage medium, and belongs to the technical field of artificial intelligence. The application improves the U-Net network structure, uses a hollow convolution instead of a normal convolution in the network, and introduces an attention mechanism expansion feature weighting. A plurality of input pictures are obtained by performing a downsampling operation on a picture to be denoised. Each input picture is input into a corresponding hollow convolution layer to obtain a plurality of first feature maps. Attention weights of the first feature maps are calculated based on an attention layer. The first feature maps are weighted using the attention weights to obtain a plurality of weighted feature maps. Each weighted feature map is input into a corresponding deconvolution layer to obtain a plurality of second feature maps. The plurality of second feature maps are fused to obtain a denoised picture. The application also relates to the technical field of block chains. The denoised picture is stored in a block chain node. The application improves the efficiency and quality of image denoising processing.

Description

Technical Field

[0001] The present application belongs to the field of artificial intelligence technology, and specifically relates to an image denoising method, apparatus, computer equipment, and storage medium. Background Art

[0002] In the AI ​​drawing tool Stable Diffusion, the U-Net module is used as a key module for image denoising. U-Net receives a noisy image as input and outputs a denoised image. The U-Net module includes a contraction path and an expansion path. The contraction path consists of a series of convolutional layers and pooling layers, which are used to gradually reduce the spatial resolution of the image and extract important features about the image structure. Then, the expansion path gradually restores these features to the size of the original image through upsampling and the reverse operation of the convolution layer, and generates the final denoising result. The key feature of the U-Net module is the design of skip connections, which are direct connections established between earlier layers and later layers for transmitting information. This connection breaks the limitation in traditional neural networks that information can only be transmitted layer by layer, allowing information to flow directly between different layers of the network.

[0003] In general, the U-Net module is used in Stable Diffusion to extract image features and perform image denoising. It extracts high-level features through contraction paths, reconstructs them through expansion paths, and utilizes skip connections to provide better information transfer. However, the U-Net module also has some shortcomings, such as a lack of global contextual information, which leads to loss of detailed information in the generated images. This results in poor quality of lower-resolution images when using Stable Diffusion to generate business-required images. Summary of the Invention

[0004] The purpose of the embodiments of the present application is to propose an image denoising method, apparatus, computer device and storage medium to solve the technical problem that the U-Net module lacks global context information, resulting in the loss of detailed information of the generated image.

[0005] In order to solve the above technical problems, the present invention provides an image denoising method, which adopts the following technical solutions:

[0006] An image denoising method is applied to an image processing platform. The image processing platform is built based on a U-Net network architecture. The image processing platform includes an encoder and a decoder. The encoder's network layer structure is a plurality of dilated convolutional layers of different specifications, and an attention layer is provided between two adjacent dilated convolutional layers. The decoder's network layer structure is a deconvolution layer corresponding to the plurality of dilated convolutional layers of different specifications. The image denoising method includes:

[0007] Get the image to be denoised and perform downsampling on it to obtain several input images, where each input image corresponds to the input of a dilated convolutional layer;

[0008] Input each input image into the corresponding dilated convolutional layer, obtain the feature map output by each dilated convolutional layer, and obtain several first feature maps;

[0009] Calculate the attention weight of the first feature map based on the attention layer;

[0010] Use the attention weight to perform feature weighting on the first feature map to obtain several weighted feature maps;

[0011] Input each weighted feature map into the corresponding deconvolution layer to obtain several second feature maps;

[0012] Perform feature fusion on several second feature maps to obtain a denoised image.

[0013] Furthermore, the steps of obtaining the image to be denoised and performing a downsampling operation on the image to be denoised to obtain several input images specifically include:

[0014] Traverse the encoder and identify the input image specifications of each hollow convolution layer;

[0015] Obtain the image to be denoised and preprocess it;

[0016] Based on the input image specifications of each dilated convolutional layer, the preprocessed denoised image is sequentially downsampled to obtain several input images, where the resolution of each input image matches the input image specifications of a dilated convolutional layer.

[0017] Furthermore, each dilated convolutional layer is connected to a fully connected layer, each input image is input into the corresponding dilated convolutional layer, and the feature map output by each dilated convolutional layer is obtained to obtain several first feature maps. Specifically, the steps include:

[0018] Traverse the encoder and identify the dilated convolution kernels of each dilated convolution layer;

[0019] In each dilated convolution layer, the dilated convolution kernel is used to perform convolution operations on the input image to obtain the image convolution features;

[0020] Input the image convolution features into the corresponding fully connected layer to obtain the feature map;

[0021] The feature maps output by each fully connected layer are obtained to obtain several first feature maps.

[0022] Furthermore, each fully connected layer is connected to a corresponding pooling layer, and the attention layer includes a channel attention unit. The step of calculating the attention weight of the first feature map based on the attention layer specifically includes:

[0023] Configuring a corresponding first initial weight for each first feature map;

[0024] Use the pooling layer to perform an average pooling operation on each first feature map to obtain the average signal strength of each hole convolution layer channel;

[0025] According to the average signal strength of each hole convolution layer channel, the first initial weight of each first feature map is updated to obtain the first attention weight of each first feature map.

[0026] Furthermore, the attention layer includes a spatial attention unit, and the step of calculating the attention weight of the first feature map based on the attention layer specifically includes:

[0027] Configuring a corresponding second initial weight for each first feature map;

[0028] Obtain the spatial position coordinates of each dilated convolutional layer channel and identify the spatial features between each dilated convolutional layer channel;

[0029] Based on the spatial features between the channels of each dilated convolution layer, the search path of each dilated convolution kernel is determined;

[0030] Based on the search path of each dilated convolution kernel, the second initial weight of each first feature map is updated to obtain the second attention weight of each first feature map.

[0031] Furthermore, the step of using the attention weight to perform feature weighting on the first feature map to obtain a plurality of weighted feature maps specifically includes:

[0032] Multiply the first attention weight of each first feature map by the second attention weight of each first feature map to obtain the attention weight of each first feature map;

[0033] The attention weight of each first feature map is used to perform feature weighting on the corresponding first feature map to obtain several weighted feature maps.

[0034] Furthermore, each weighted feature map is input into the corresponding deconvolution layer to obtain several second feature maps, specifically including:

[0035] Get the required denoising resolution and set the size of the transposed convolution kernel and the step size of the deconvolution operation in the deconvolution layer according to the required denoising resolution;

[0036] In the deconvolution layer, the transposed convolution kernel is used to perform a deconvolution operation on the weighted feature map according to the deconvolution operation step size to obtain the feature map output by the deconvolution layer;

[0037] Get the feature map output by each deconvolution layer and obtain several second feature maps.

[0038] In order to solve the above technical problems, the present application also provides an image denoising device, which adopts the following technical solution:

[0039] An image denoising device is used to run an image processing platform. The image processing platform is built based on a U-Net network architecture. The image processing platform includes an encoder and a decoder. The encoder's network layer structure is a plurality of hole convolution layers of different specifications, and an attention layer is provided between two adjacent hole convolution layers. The decoder's network layer structure is a deconvolution layer corresponding to the plurality of hole convolution layers of different specifications. The image denoising device includes:

[0040] The downsampling operation module is used to obtain the image to be denoised and perform a downsampling operation on the denoised image to obtain several input images, where each input image corresponds to the input of a hollow convolution layer;

[0041] The dilated convolution module is used to input each input image into the corresponding dilated convolution layer, obtain the feature map output by each dilated convolution layer, and obtain several first feature maps;

[0042] An attention weight module, configured to calculate the attention weight of the first feature map based on the attention layer;

[0043] A feature weighting module, configured to perform feature weighting on the first feature map using the attention weight to obtain a plurality of weighted feature maps;

[0044] The deconvolution module is used to input each weighted feature map into the corresponding deconvolution layer to obtain several second feature maps;

[0045] The feature map fusion module is used to perform feature fusion on several second feature maps to obtain a denoised image.

[0046] In order to solve the above technical problems, the embodiment of the present application further provides a computer device, which adopts the following technical solution:

[0047] A computer device includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of any one of the above-mentioned image denoising methods when executing the computer-readable instructions.

[0048] In order to solve the above technical problems, the embodiment of the present application further provides a computer-readable storage medium, which adopts the following technical solution:

[0049] A computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of any one of the above-mentioned image denoising methods.

[0050] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0051] The present application discloses an image denoising method, device, computer equipment and storage medium, which belongs to the field of artificial intelligence technology. The present application improves the U-Net network structure, uses hole convolution to replace ordinary convolution in the network, and introduces an attention mechanism to expand feature weighting. By obtaining the image to be denoised, the image to be denoised is downsampled to obtain several input images; each input image is input into the corresponding hole convolution layer to obtain several first feature maps; the attention weight of the first feature map is calculated based on the attention layer; the first feature map is feature-weighted using the attention weight to obtain several weighted feature maps; each weighted feature map is input into the corresponding deconvolution layer to obtain several second feature maps; the several second feature maps are feature-fused to obtain a denoised image. The present application improves the receptive field range of the model by using hole convolution and introducing an attention mechanism, so that the model obtains more comprehensive global context information, improves the quality of image denoising processing, and effectively reduces the running cost and training cost, improves the cost performance of the model, and improves the model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the solutions in this application, a brief introduction will be given below to the drawings required for use in the description of the embodiments of this application. Obviously, the drawings described below are some embodiments of this application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0053] Figure 1 shows an exemplary system architecture diagram in which the present application can be applied;

[0054] Figure 2 A flowchart of an embodiment of an image denoising method according to the present application is shown;

[0055] Figure 3 A schematic structural diagram of an embodiment of an image denoising device according to the present application is shown;

[0056] Figure 4 A schematic structural diagram of an embodiment of a computer device according to the present application is shown. DETAILED DESCRIPTION

[0057] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of the application are for the purpose of describing specific embodiments only and are not intended to limit this application. The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0058] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0059] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0060] like Figure 1 As shown, system architecture 100 may include a terminal device 101, a network 102, and a server 103. Terminal device 101 may be a laptop computer 1011, a tablet computer 1012, or a mobile phone 1013. Network 102 is a medium for providing a communication link between terminal device 101 and server 103. Network 102 may include various connection types, such as wired or wireless communication links or fiber optic cables.

[0061] The user can use the terminal device 101 to interact with the server 103 via the network 102 to receive or send messages, etc. Various communication client applications can be installed on the terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0062] The terminal device 101 can be various electronic devices with a display screen and supporting web browsing. In addition to the laptop computer 1011, tablet computer 1012 or mobile phone 1013, the terminal device 101 can also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 (Moving Picture Experts Group Audio Layer IV) player, a laptop computer and a desktop computer, etc.

[0063] The server 103 may be a server that provides various services, such as a background server that provides support for web pages displayed on the terminal device 101 .

[0064] It should be noted that the image denoising method provided in the embodiments of the present application is generally executed by a server / terminal device, and accordingly, the image denoising device is generally provided in the server / terminal device.

[0065] It should be understood that Figure 1 The numbers of terminal devices, networks and servers in the embodiment are merely illustrative. The above system may have any number of terminal devices, networks and servers according to implementation requirements.

[0066] An image denoising method is applied to an image processing platform. The image processing platform is built based on a U-Net network architecture. The image processing platform includes an encoder and a decoder. The encoder's network layer structure is composed of several dilated convolutional layers of different specifications, and an attention layer is provided between two adjacent dilated convolutional layers. The decoder's network layer structure is composed of deconvolution layers corresponding to the several dilated convolutional layers of different specifications.

[0067] The U-Net network is a convolutional neural network (CNN) architecture widely used in deep learning. Its main feature is its U-shaped symmetrical structure, consisting of an "encoder" (contracting path) and a "decoder" (expanding path). It is primarily used for image processing tasks. The encoder is responsible for feature extraction, gradually reducing the spatial resolution of the image through convolutional layers and maximum pooling layers while increasing the number of feature channels. Each downsampling increases the number of features while reducing the spatial dimension, enabling the network to learn features from local to global perspectives. The decoder gradually restores the spatial resolution and details of the image through an upsampling process. These upsampling layers are typically implemented by transposed convolutional layers. After each upsampling step, the output feature map is merged with the feature map of the corresponding encoder layer (via skip connections) to restore the lost spatial information.

[0068] It should be noted that a key feature of U-Net is its skip connection, which connects the feature maps in the encoder to the feature maps of the corresponding layer in the decoder. This helps to recover detailed information during upsampling and allows the network to learn more accurate outputs.

[0069] Dilated Convolution, also known as expanded convolution or dilated convolution, is a convolution method that adds spaces (usually zeros) between the convolution kernel elements to expand the convolution kernel. Dilated convolution introduces the concept of intervals (or holes) in standard convolution. By inserting interval holes between the convolution kernels, the convolution kernel sees more pixels, thereby achieving the purpose of expanding the receptive field. In standard convolution, the convolution kernel slides with a stride of 1 during the convolution operation, while in dilated convolution, the convolution kernel can slide in the form of larger intervals or holes. The number of intervals in the dilated convolution determines the dilation rate of the convolution. The higher the dilation rate, the larger the receptive field.

[0070] In a specific embodiment of the present application, spacing holes are inserted between the convolution kernels so that the convolution kernels see more pixels, effectively expanding the receptive field. Using hole convolution instead of ordinary convolution in the U-Net network can obtain a wider range of image context information and richer semantic information; and reduce the number of parameters. At the same time, by replacing ordinary convolution in the U-Net network with hole convolution, the number of parameters of the model can be reduced with the same amount of computation and memory usage, thereby improving the efficiency and generalization ability of the model; improving the cost-effectiveness of the model, and using hole convolution to replace ordinary convolution in the U-Net network can effectively reduce the operating cost and training cost while still ensuring the accuracy and stability of the model.

[0071] In addition, this application also uses the CBAM (Convolutional Block Attention Module) attention mechanism to implement feature weighting. CBAM consists of two parts: the Channel Attention Module and the Spatial Attention Module. The CBAM attention mechanism can detect the spatial and channel information in the feature map, enhance the representation ability of useful information in the feature map, and improve network performance and accuracy; improve the receptive field. Within the U-Net network, the CBAM mechanism can adaptively control the size of the receptive field, allowing the network to learn more and broader contextual information and have stronger perception capabilities; and improve image quality. CBAM focuses on increasing the correlation between spatial and channel information, which can further improve image quality while ensuring that the accuracy rate does not decrease, making the image processing results more accurate and natural. At the same time, it can also reduce the jaggedness, glitches and other phenomena that appear in the prediction.

[0072] Continue to refer Figure 2 , shows a flow chart of an embodiment of the image denoising method according to the present application. The image denoising method comprises the following steps:

[0073] S201: Obtain a picture to be denoised, perform a downsampling operation on the picture to be denoised, and obtain several input pictures, wherein each input picture corresponds to an input of a dilated convolutional layer.

[0074] Specifically, first, the image to be denoised is obtained from an external data source, that is, the image to be denoised. Typically, the image to be denoised contains various forms of noise, such as Gaussian noise, salt and pepper noise, etc., which may come from sensor noise, transmission errors, or compression losses. In order to input the image layer by layer into each dilated convolution layer in the encoder, the image to be denoised needs to be downsampled. The purpose of downsampling is to reduce redundant information by reducing the image size, so that subsequent convolution operations can focus on extracting more important features, while also reducing computational complexity. Specific downsampling methods can include max pooling or average pooling, which can gradually reduce the original image. After downsampling, each reduced image is passed as input to the corresponding dilated convolution layer in the encoder part of the U-Net.

[0075] S202: Input each input image into the corresponding dilated convolutional layer, obtain the feature map output by each dilated convolutional layer, and obtain several first feature maps.

[0076] Specifically, the input images are fed one by one into the encoder portion of the U-Net network architecture for feature extraction. The encoder consists of several dilated convolutional layers, each with a different dilation rate. These convolutional layers have different receptive fields and can capture image features at different levels. Dilated convolutions expand the receptive field without increasing the amount of computation by inserting holes (i.e., skipping several pixels) during the convolution operation. Dilated convolutions can capture both global and local features in the image. After dilated convolution processing, the input image will generate a set of feature maps, namely the first feature maps, which retain the key information of the input image and enhance the ability to express multi-scale features.

[0077] S203: Calculate the attention weight of the first feature map based on the attention layer.

[0078] Specifically, the attention mechanism is introduced into the U-Net architecture to enhance focus on important features. The core idea of ​​this attention mechanism is that different features have different degrees of importance in image denoising, so it is necessary to use the attention layer to calculate the attention weights for each position on each feature map. These weights reflect the importance of different regions in the feature map. For example, the CBAM attention mechanism can be used. The attention layer automatically weights different parts of the feature map by learning different channel and spatial weights. For example, the attention mechanism may assign higher weights to edges or texture details in the image because these features are particularly important for denoising and image reconstruction. The weights can be calculated by multiplying them with the features output by the encoder to obtain the final attention weights.

[0079] S204: Use the attention weight to perform feature weighting on the first feature map to obtain several weighted feature maps.

[0080] Specifically, the attention weights calculated in S203 are used to weight the first feature map output by each hole convolution layer. The specific weighting process is usually to multiply the eigenvalues ​​of each position of each feature map by the corresponding attention weight. The purpose of this operation is to enhance those features that are more important for the denoising task, while suppressing those irrelevant or unimportant features. The weighted feature maps are called weighted feature maps. These feature maps not only contain the structure and detail information in the original image, but also combine the attention mechanism, so that the model can pay more attention to the image areas that are helpful for the denoising process. For example, for a noisy image with complex textures, the weighted feature map may highlight the edge details more, thereby better reconstructing a clean image in the subsequent decoder.

[0081] S205: Input each weighted feature map into the corresponding deconvolution layer to obtain several second feature maps.

[0082] Specifically, this step is performed in the decoder stage. Each weighted feature map is input into the corresponding deconvolution layer one by one. The deconvolution layer (also called the transposed convolution layer) is used to perform upsampling operations, the purpose of which is to gradually restore the spatial resolution of the feature map. Corresponding to the downsampling operation in the encoder, deconvolution amplifies the low-resolution weighted feature map, while extracting features through the convolution kernel to help reconstruct image details. Different deconvolution layers can have different convolution kernel sizes and step sizes to ensure that the feature map maintains important feature information while being upsampled layer by layer. After this process, the deconvolution layer outputs several second feature maps, which have higher resolution than the previous weighted feature maps and retain richer spatial detail information.

[0083] S206: Perform feature fusion on several second feature maps to obtain a denoised image.

[0084] Specifically, in this step, the second feature maps from different deconvolution layers are fused together to generate the final denoised image. There are many ways to fuse features. The most common method is to add or concatenate feature maps of different resolutions and channels element by element to combine the key information in each feature map. The purpose of fusion is to integrate features at different levels to ensure that the denoised image not only restores spatial details but also retains semantic consistency. Through this feature fusion, low-level detail features and high-level semantic features can be effectively combined, and the generated denoised image should be as close to the real, noise-free image as possible. The final denoised image will maintain the original resolution of the image while eliminating noise as much as possible to restore a clear visual effect.

[0085] In the above embodiments, the present application improves the receptive field range of the model by using void convolution and introducing the attention mechanism, so that the model obtains more comprehensive global context information, improves the quality of image denoising processing, and effectively reduces the operating cost and training cost, improves the cost performance of the model, and improves the model performance.

[0086] Furthermore, the steps of obtaining the image to be denoised and performing a downsampling operation on the image to be denoised to obtain several input images specifically include:

[0087] Traverse the encoder and identify the input image specifications of each hollow convolution layer;

[0088] Obtain the image to be denoised and preprocess it;

[0089] Based on the input image specifications of each dilated convolutional layer, the preprocessed denoised image is sequentially downsampled to obtain several input images, where the resolution of each input image matches the input image specifications of a dilated convolutional layer.

[0090] In this embodiment, the encoder is first traversed to identify the input image specifications required by each dilated convolutional layer. Different dilated convolutional layers have different receptive fields and convolution kernel sizes, and therefore require input images of different resolutions. The purpose of this step is to ensure that each convolutional layer can receive an image size that matches its configuration. Next, the image to be denoised is obtained and subjected to necessary preprocessing. These preprocessing steps typically include image normalization, outlier removal, or other data augmentation methods to ensure that the image has a good initial state before entering the network. Afterwards, based on the previously identified dilated convolutional layer input specifications, the preprocessed image is downsampled layer by layer. The purpose of downsampling is to gradually reduce the image resolution to meet the input requirements of each dilated convolutional layer. The downsampling method can be a technique such as maximum pooling, mean pooling, or interpolation, so that the input image of each layer not only retains important feature information but also meets the resolution requirements of the corresponding convolutional layer. Finally, after downsampling, the multiple input images generated will be used as the input of each dilated convolutional layer in turn, ready for further convolution operations.

[0091] Through the above steps, it is possible to ensure that each dilated convolutional layer in the encoder receives an input image that matches its network configuration, avoiding the problem of inaccurate feature extraction caused by mismatch in input image resolution.

[0092] Furthermore, each dilated convolutional layer is connected to a fully connected layer, each input image is input into the corresponding dilated convolutional layer, and the feature map output by each dilated convolutional layer is obtained to obtain several first feature maps. Specifically, the steps include:

[0093] Traverse the encoder and identify the dilated convolution kernels of each dilated convolution layer;

[0094] In each dilated convolution layer, the dilated convolution kernel is used to perform convolution operations on the input image to obtain the image convolution features;

[0095] Input the image convolution features into the corresponding fully connected layer to obtain the feature map;

[0096] The feature maps output by each fully connected layer are obtained to obtain several first feature maps.

[0097] In this embodiment, the encoder is first traversed to identify the dilated convolution kernel used in each dilated convolution layer. The characteristic of the dilated convolution kernel is that it expands the receptive field of the convolution without increasing the parameters by introducing a dilation rate in the convolution operation, allowing the convolution layer to capture more contextual information. Different dilated convolution layers may use different dilation rates to facilitate the extraction of image features at multiple scales. After identifying the dilated convolution kernel of each convolution layer, the input image is then convolved. The input image is processed by the dilated convolution kernel to generate convolution features. These convolution features contain important information such as edges and textures in the image and can capture the details and overall structure of the image at different scales. After the convolution feature extraction is completed, the features of each layer are input into the corresponding fully connected layer. The function of the fully connected layer is to further compress and integrate the convolution features, making them more compact and more expressive. The fully connected layer generates feature maps for each dilated convolution layer by mapping the convolution features to a specific feature space. Finally, the outputs of each fully connected layer are aggregated to obtain multiple first feature maps.

[0098] Through the above steps, the dilated convolution kernel is used to expand the receptive field, improving the network's ability to capture features of different scales, especially when dealing with complex noise. At the same time, the fully connected layer further refines the feature map, effectively compressing redundant information and improving denoising accuracy.

[0099] Furthermore, each fully connected layer is connected to a corresponding pooling layer, and the attention layer includes a channel attention unit. The step of calculating the attention weight of the first feature map based on the attention layer specifically includes:

[0100] Configuring a corresponding first initial weight for each first feature map;

[0101] Use the pooling layer to perform an average pooling operation on each first feature map to obtain the average signal strength of each hole convolution layer channel;

[0102] According to the average signal strength of each hole convolution layer channel, the first initial weight of each first feature map is updated to obtain the first attention weight of each first feature map.

[0103] In this embodiment, each first feature map generated by the dilated convolution layer is first configured with corresponding initial weights, namely, first initial weights. These first initial weights can be randomly initialized or based on weights obtained from a pre-trained model. The role of the initial weights is to provide a basis for weight updates for subsequent channel attention calculations and determine the importance of each feature map in its channel. Next, a pooling layer is used to perform an average pooling operation on each first feature map. The role of average pooling is to reduce the dimensionality of the feature map and summarize the information of each feature channel by taking the average value of each channel. Here, average pooling calculates the average signal strength of each dilated convolution layer channel. The average signal strength of the channel reflects the response of the channel in the current input image, that is, the importance of the channel to the input image. Next, the initial weights of each first feature map are updated based on these average signal strengths. The weight update method can be based on weighting or multiplication, with the aim of giving feature maps with high signal strengths higher weights and feature maps with low signal strengths lower weights.

[0104] Through the above steps, the signal strength of each channel can be effectively extracted through the average pooling operation. Combined with the attention weight update mechanism, the model pays more attention to the feature areas in the image that are helpful for denoising, thereby improving the accuracy and robustness of the denoising effect.

[0105] Furthermore, the attention layer includes a spatial attention unit, and the step of calculating the attention weight of the first feature map based on the attention layer specifically includes:

[0106] Configuring a corresponding second initial weight for each first feature map;

[0107] Obtain the spatial position coordinates of each dilated convolutional layer channel and identify the spatial features between each dilated convolutional layer channel;

[0108] Based on the spatial features between the channels of each dilated convolution layer, the search path of each dilated convolution kernel is determined;

[0109] Based on the search path of each dilated convolution kernel, the second initial weight of each first feature map is updated to obtain the second attention weight of each first feature map.

[0110] In this embodiment, a corresponding second initial weight is first configured for each first feature map. This initial weight is randomly generated based on the spatial position and channel information of the feature map, and may also refer to the existing weight distribution in the pre-trained model. Next, the system obtains the spatial position coordinates of the channel from each hole convolution layer. These coordinates describe the specific position of each channel in the feature map and identify the spatial features between the channels. The spatial feature recognition process is mainly to observe the distribution pattern and mutual relationship of each channel in the feature map to determine which areas or channels have strong correlations. For example, the performance of structured features such as edges and textures in space usually has a strong correlation. Based on these spatial features, the search path of the hole convolution kernel can be determined. The characteristic of the hole convolution is to capture the connection between distant pixels at different scales through a jump convolution operation. Therefore, it is necessary to design the search path of the convolution kernel to maximize the extraction of spatial features. Specifically, the search path guides how the convolution kernel is sampled on the image to ensure that the convolution operation not only covers important local areas, but also captures a wider range of global features. Then, based on the search path of each dilated convolution kernel, the second initial weight of the corresponding feature map is updated. This update process is a further weighted operation on the spatial features, so that the convolution kernel has a higher response weight at important spatial positions, and finally the second attention weight of each first feature map is obtained.

[0111] Through the above steps, by optimizing the search path of the dilated convolution kernel, it is possible to more accurately identify and enhance the important spatial areas in the feature map, and by updating the second initial weights, the model can more effectively focus on the valuable image feature areas, thereby improving the globality and accuracy of the denoising effect.

[0112] Furthermore, the step of using the attention weight to perform feature weighting on the first feature map to obtain a plurality of weighted feature maps specifically includes:

[0113] Multiply the first attention weight of each first feature map by the second attention weight of each first feature map to obtain the attention weight of each first feature map;

[0114] The attention weight of each first feature map is used to perform feature weighting on the corresponding first feature map to obtain several weighted feature maps.

[0115] In this embodiment, the first attention weight and the second attention weight of each first feature map are firstly multiplied element-by-element to generate a comprehensive attention weight. The first attention weight is mainly calculated based on the signal strength of each hole convolution layer channel, reflecting the importance of the contribution of the channel in the entire denoising process; the second attention weight is weighted by the search path of the hole convolution kernel combined with the spatial features, emphasizing the importance of features at different spatial positions in the image. Multiplying these two weights can comprehensively consider the information of the feature map in signal strength and spatial distribution, ensuring that when denoising the image, both the key channel information is paid attention to, and the local and global features in the image space are not ignored, so as to achieve global interaction and integration of channel and spatial information in the feature map. Then, the obtained comprehensive attention weight is used to perform a weighted operation on each first feature map to generate a weighted feature map.

[0116] Through the above steps, the attention mechanisms of channel and spatial dimensions are effectively combined. The multi-weighted strategy enables the model to focus more on important features in the image, thereby improving the accuracy and robustness of denoising and ensuring that the key details of the image are preserved.

[0117] Furthermore, each weighted feature map is input into the corresponding deconvolution layer to obtain several second feature maps, specifically including:

[0118] Get the required denoising resolution and set the size of the transposed convolution kernel and the step size of the deconvolution operation in the deconvolution layer according to the required denoising resolution;

[0119] In the deconvolution layer, the transposed convolution kernel is used to perform a deconvolution operation on the weighted feature map according to the deconvolution operation step size to obtain the feature map output by the deconvolution layer;

[0120] Get the feature map output by each deconvolution layer and obtain several second feature maps.

[0121] In this embodiment, the resolution of the target image is first determined based on the denoising requirements of the user or system. The denoising requirements usually depend on the actual application scenario, such as high-definition image restoration or lower-resolution denoising tasks. After determining the required denoising resolution, the key parameters in the deconvolution layer are configured according to the resolution, including the size of the transposed convolution kernel and the step size of the deconvolution operation. The size of the transposed convolution kernel determines the spatial expansion capability of the image when it is restored from low resolution to high resolution, while the step size controls the speed of the deconvolution operation in the spatial dimension. A step size that is too large may cause the restored image to be not smooth enough, while a step size that is too small may increase the computational overhead. Therefore, the configuration of these parameters is to enable the decoder to efficiently generate images that meet the required resolution. Next, the deconvolution layer uses the configured transposed convolution kernel to perform a deconvolution operation on the weighted feature map. Deconvolution is an upsampling process that gradually restores the low-dimensional feature map to high-dimensional image features through the transposed convolution kernel. In this process, the deconvolution operation focuses on recovering important information in the image based on the high-weight areas in the weighted feature map while suppressing the influence of noise. Each layer of deconvolution outputs a feature map, which is passed to the next layer in turn, ultimately obtaining several second feature maps. These second feature maps carry the key information of the image, and in the layer-by-layer recovery process, noise is effectively removed while image details are preserved.

[0122] Through the above steps, the parameters of the deconvolution layer can be flexibly adjusted according to the denoising requirements, making it more suitable for image denoising tasks of different resolutions. Through precise deconvolution operations, the image can be gradually restored and noise removed while maintaining the main features and details of the image, ensuring that the final output denoised image has high quality.

[0123] In the above embodiment, the present application discloses an image denoising method, which belongs to the field of artificial intelligence technology. The present application improves the U-Net network structure, uses hole convolution to replace ordinary convolution in the network, and introduces an attention mechanism to expand feature weighting. By obtaining the image to be denoised, the image to be denoised is downsampled to obtain several input images; each input image is input into the corresponding hole convolution layer to obtain several first feature maps; the attention weight of the first feature map is calculated based on the attention layer; the first feature map is feature-weighted using the attention weight to obtain several weighted feature maps; each weighted feature map is input into the corresponding deconvolution layer to obtain several second feature maps; the several second feature maps are feature-fused to obtain a denoised image. The present application improves the receptive field range of the model by using hole convolution and introducing an attention mechanism, so that the model obtains more comprehensive global context information, improves the quality of image denoising processing, and effectively reduces the running cost and training cost, improves the cost performance of the model, and improves the model performance.

[0124] In this embodiment, the electronic device (eg, Figure 1 The server shown in the figure) can receive instructions or obtain data through a wired connection or a wireless connection. It should be noted that the above-mentioned wireless connection method may include but is not limited to 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other wireless connection methods currently known or to be developed in the future.

[0125] It should be emphasized that in order to further ensure the privacy and security of the above-mentioned denoised image information, the above-mentioned denoised image information can also be stored in a node of a blockchain.

[0126] The blockchain referred to in this application is a new application model for computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Blockchain is essentially a decentralized database, a series of data blocks generated using cryptographic methods. Each data block contains information about a batch of network transactions, which is used to verify the validity of this information (to prevent counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, the platform product service layer, and the application service layer.

[0127] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial Intelligence (AI) is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to achieve optimal results.

[0128] Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0129] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware via computer-readable instructions. The computer-readable instructions can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes in the above-described method embodiments. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0130] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0131] Further references Figure 3 , as a response to the above Figure 2 The present application provides an embodiment of an image denoising device, which is similar to the embodiment of the present invention. Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0132] like Figure 3 As shown, the image denoising device 300 described in this embodiment is used to run an image processing platform. The image processing platform is built based on the U-Net network architecture. The image processing platform includes an encoder and a decoder. The network layer structure of the encoder is a plurality of hole convolution layers of different specifications, and an attention layer is provided between two adjacent hole convolution layers. The network layer structure of the decoder is a deconvolution layer corresponding to the plurality of hole convolution layers of different specifications. The image denoising device 300 includes:

[0133] A downsampling operation module 301 is used to obtain a picture to be denoised and perform a downsampling operation on the picture to be denoised to obtain a plurality of input pictures, wherein each input picture corresponds to an input of a dilated convolutional layer;

[0134] The dilated convolution module 302 is used to input each input image into the corresponding dilated convolution layer, obtain the feature map output by each dilated convolution layer, and obtain a plurality of first feature maps;

[0135] An attention weight module 303, configured to calculate an attention weight of the first feature map based on the attention layer;

[0136] A feature weighting module 304 is configured to perform feature weighting on the first feature map using the attention weights to obtain a plurality of weighted feature maps;

[0137] A deconvolution module 305 is used to input each weighted feature map into a corresponding deconvolution layer to obtain a plurality of second feature maps;

[0138] The feature map fusion module 306 is used to perform feature fusion on a plurality of second feature maps to obtain a denoised image.

[0139] Furthermore, the downsampling operation module 301 is specifically configured to:

[0140] Traverse the encoder and identify the input image specifications of each dilated convolutional layer; obtain the image to be denoised and preprocess it; based on the input image specifications of each dilated convolutional layer, downsample the preprocessed image to be denoised in sequence to obtain several input images, where the resolution of each input image matches the input image specifications of a dilated convolutional layer.

[0141] Furthermore, each dilated convolutional layer is connected to a fully connected layer, and the dilated convolutional module 302 is specifically used to:

[0142] Traverse the encoder and identify the dilated convolution kernels of each dilated convolution layer. In each dilated convolution layer, use the dilated convolution kernel to perform a convolution operation on the input image to obtain the image convolution features. Input the image convolution features into the corresponding fully connected layer to obtain a feature map. Obtain the feature maps output by each fully connected layer to obtain several first feature maps.

[0143] Furthermore, each fully connected layer is connected to a corresponding pooling layer, the attention layer includes a channel attention unit, and the attention weight module 303 is specifically used to:

[0144] Configure a corresponding first initial weight for each first feature map; use the pooling layer to perform an average pooling operation on each first feature map to obtain the average signal strength of each hole convolution layer channel; according to the average signal strength of each hole convolution layer channel, update the first initial weight of each first feature map to obtain the first attention weight of each first feature map.

[0145] Furthermore, the attention layer includes a spatial attention unit, and the attention weight module 303 is further used to:

[0146] Configure the corresponding second initial weight for each first feature map; obtain the spatial position coordinates of each dilated convolution layer channel and identify the spatial features between each dilated convolution layer channel; determine the search path of each dilated convolution kernel based on the spatial features between each dilated convolution layer channel; based on the search path of each dilated convolution kernel, update the second initial weight of each first feature map to obtain the second attention weight of each first feature map.

[0147] Furthermore, the feature weighting module 304 is specifically configured to:

[0148] The first attention weight of each first feature map is multiplied by the second attention weight of each first feature map to obtain the attention weight of each first feature map; the attention weight of each first feature map is used to perform feature weighting on the corresponding first feature map to obtain several weighted feature maps.

[0149] Furthermore, the deconvolution module 305 is specifically configured to:

[0150] Get the required denoising resolution and set the size of the transposed convolution kernel and the step size of the deconvolution operation in the deconvolution layer according to the required denoising resolution;

[0151] In the deconvolution layer, the transposed convolution kernel is used to perform a deconvolution operation on the weighted feature map according to the deconvolution operation step size to obtain the feature map output by the deconvolution layer;

[0152] Get the feature map output by each deconvolution layer and obtain several second feature maps.

[0153] In the above embodiment, the present application discloses an image denoising device, which belongs to the field of artificial intelligence technology. The present application improves the U-Net network structure, uses hole convolution to replace ordinary convolution in the network, and introduces an attention mechanism to expand feature weighting. By obtaining the image to be denoised, the image to be denoised is downsampled to obtain several input images; each input image is input into the corresponding hole convolution layer to obtain several first feature maps; the attention weight of the first feature map is calculated based on the attention layer; the first feature map is feature-weighted using the attention weight to obtain several weighted feature maps; each weighted feature map is input into the corresponding deconvolution layer to obtain several second feature maps; the several second feature maps are feature-fused to obtain a denoised image. The present application improves the receptive field range of the model by using hole convolution and introducing an attention mechanism, so that the model obtains more comprehensive global context information, improves the quality of image denoising processing, and effectively reduces the operating cost and training cost, improves the cost performance of the model, and improves the model performance.

[0154] To solve the above technical problems, the present application also provides a computer device. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0155] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected through a system bus. It should be noted that the figure only shows a computer device 4 with a memory 41, a processor 42, and a network interface 43, but it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to a microprocessor, an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a digital signal processor (DSP), an embedded device, etc.

[0156] The computer device may be a desktop computer, notebook computer, PDA, cloud server, etc. The computer device may interact with the user via a keyboard, mouse, remote control, touchpad, or voice control device.

[0157] The memory 41 includes at least one type of readable storage medium, including flash memory, a hard disk, a multimedia card, a card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, a magnetic disk, an optical disk, etc. In some embodiments, the memory 41 may be an internal storage unit of the computer device 4, such as the hard disk or memory of the computer device 4. In other embodiments, the memory 41 may also be an external storage device of the computer device 4, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash memory card, etc. equipped on the computer device 4. Of course, the memory 41 may also include both the internal storage unit of the computer device 4 and its external storage device. In this embodiment, the memory 41 is generally used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions for the image denoising method. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or are to be output.

[0158] In some embodiments, the processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 42 is generally used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to execute computer-readable instructions stored in the memory 41 or process data, such as computer-readable instructions for executing the image denoising method.

[0159] The network interface 43 may include a wireless network interface or a wired network interface. The network interface 43 is generally used to establish a communication connection between the computer device 4 and other electronic devices.

[0160] In the above embodiment, the present application discloses a computer device, which belongs to the field of artificial intelligence technology. The present application improves the U-Net network structure, uses hole convolution to replace ordinary convolution in the network, and introduces an attention mechanism to expand feature weighting. By obtaining the image to be denoised, the image to be denoised is downsampled to obtain several input images; each input image is input into the corresponding hole convolution layer to obtain several first feature maps; the attention weight of the first feature map is calculated based on the attention layer; the first feature map is feature-weighted using the attention weight to obtain several weighted feature maps; each weighted feature map is input into the corresponding deconvolution layer to obtain several second feature maps; the several second feature maps are feature-fused to obtain a denoised image. The present application improves the receptive field range of the model by using hole convolution and introducing an attention mechanism, so that the model obtains more comprehensive global context information, improves the quality of image denoising processing, and effectively reduces the running cost and training cost, improves the cost performance of the model, and improves the model performance.

[0161] The present application also provides another embodiment, namely, providing a computer-readable storage medium, wherein the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor to enable the at least one processor to perform the steps of the image denoising method as described above.

[0162] In the above embodiment, the application discloses a computer readable storage medium, which belongs to the technical field of artificial intelligence. The U-Net network structure is improved, the ordinary convolution in the network is replaced with a hollow convolution, and an attention mechanism is introduced for feature weighting. A to-be-de-noised picture is obtained, and a down-sampling operation is performed on the to-be-de-noised picture to obtain a plurality of input pictures. Each input picture is input into a corresponding hollow convolution layer to obtain a plurality of first feature maps. The attention weights of the first feature maps are calculated based on an attention layer. The first feature maps are weighted using the attention weights to obtain a plurality of weighted feature maps. Each weighted feature map is input into a corresponding deconvolution layer to obtain a plurality of second feature maps. The plurality of second feature maps are fused to obtain a de-noised picture. By using the hollow convolution and introducing the attention mechanism, the receptive field range of the model is improved, the model obtains more comprehensive global context information, the quality of image de-noising processing is improved, the running cost and the training cost are effectively reduced, the performance-price ratio of the model is improved, and the performance of the model is improved.

[0163] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment method can be realized by means of software and a general hardware platform, of course, it can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a plurality of instructions for making a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) execute the methods described in various embodiments of the application.

[0164] The application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, small computers, mainframe computers, distributed computing environments including any of the above systems or devices. The application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The application can also be practiced in a distributed computing environment, in which tasks are performed by remote processing devices connected by a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0165] Obviously, the embodiments described above are only some of the embodiments of the present application, rather than all of the embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions described in the aforementioned specific embodiments, or to make equivalent replacements for some of the technical features therein. Any equivalent structure made using the contents of the present application specification and the accompanying drawings, directly or indirectly used in other related technical fields, is also within the scope of patent protection of the present application.

Claims

1. An image denoising method, characterized in that: The image denoising method is applied to an image processing platform, which is built based on a U-Net network architecture. The image processing platform includes an encoder and a decoder. The encoder's network layer structure is a plurality of dilated convolutional layers of different specifications, and an attention layer is provided between two adjacent dilated convolutional layers. The decoder's network layer structure is a deconvolution layer corresponding to the plurality of dilated convolutional layers of different specifications. The image denoising method includes: Obtaining a picture to be denoised, and performing a downsampling operation on the picture to be denoised to obtain a plurality of input pictures, wherein each of the input pictures corresponds to an input of the dilated convolutional layer; Input each of the input images into the corresponding dilated convolutional layer, obtain a feature map output by each dilated convolutional layer, and obtain a plurality of first feature maps; Calculating an attention weight of the first feature map based on the attention layer; Performing feature weighting on the first feature map using the attention weight to obtain a plurality of weighted feature maps; Inputting each of the weighted feature maps into a corresponding deconvolution layer to obtain a plurality of second feature maps; Feature fusion is performed on several of the second feature maps to obtain a denoised image.

2. The image denoising method according to claim 1, wherein: The step of obtaining the image to be denoised and performing a downsampling operation on the image to be denoised to obtain a plurality of input images specifically includes: Traversing the encoder to identify the input image specifications of each of the dilated convolutional layers; Acquire the image to be denoised, and preprocess the image to be denoised; Based on the input image specifications of each of the dilated convolutional layers, the preprocessed images to be denoised are sequentially downsampled to obtain a plurality of input images, wherein the resolution of each input image matches the input image specifications of one of the dilated convolutional layers.

3. The image denoising method according to claim 1, wherein: Each of the dilated convolutional layers is connected to a fully connected layer, and the step of inputting each of the input images into the corresponding dilated convolutional layer, obtaining the feature map output by each dilated convolutional layer, and obtaining a plurality of first feature maps specifically includes: Traversing the encoder to identify the dilated convolution kernels of each dilated convolution layer; In each of the dilated convolution layers, the dilated convolution kernel is used to perform a convolution operation on the input image to obtain image convolution features; Input the image convolution features into the corresponding fully connected layer to obtain the feature map; The feature maps output by each of the fully connected layers are obtained to obtain a plurality of first feature maps.

4. The image denoising method according to claim 3, wherein: Each of the fully connected layers is connected to a corresponding pooling layer, the attention layer includes a channel attention unit, and the step of calculating the attention weight of the first feature map based on the attention layer specifically includes: Configuring a corresponding first initial weight for each of the first feature maps; Use the pooling layer to perform an average pooling operation on each of the first feature maps to obtain the average signal strength of each hole convolution layer channel; According to the average signal strength of each of the dilated convolutional layer channels, the first initial weight of each of the first feature maps is updated to obtain a first attention weight of each of the first feature maps.

5. The image denoising method according to claim 4, wherein: The attention layer includes a spatial attention unit, and the step of calculating the attention weight of the first feature map based on the attention layer specifically includes: Configuring a corresponding second initial weight for each of the first feature maps; Obtaining the spatial position coordinates of each of the dilated convolutional layer channels, and identifying the spatial features between each of the dilated convolutional layer channels; Determining the search path of each of the dilated convolution kernels based on the spatial features between the channels of each of the dilated convolution layers; Based on the search path of each of the dilated convolution kernels, the second initial weight of each of the first feature maps is updated to obtain the second attention weight of each of the first feature maps.

6. The image denoising method according to claim 5, wherein: The step of using the attention weight to perform feature weighting on the first feature map to obtain a plurality of weighted feature maps specifically includes: Multiplying the first attention weight of each of the first feature maps by the second attention weight of each of the first feature maps to obtain an attention weight of each of the first feature maps; The attention weight of each of the first feature maps is used to perform feature weighting on the corresponding first feature map to obtain a plurality of weighted feature maps.

7. The image denoising method according to claim 1, wherein: The step of inputting each weighted feature map into a corresponding deconvolution layer to obtain a plurality of second feature maps specifically includes: Obtaining a required denoising resolution, and setting a size of a transposed convolution kernel and a step size of a deconvolution operation in the deconvolution layer according to the required denoising resolution; In the deconvolution layer, using the transposed convolution kernel, a deconvolution operation is performed on the weighted feature map according to the deconvolution operation step size to obtain a feature map output by the deconvolution layer; Obtain a feature map output by each of the deconvolution layers to obtain a plurality of second feature maps.

8. An image denoising device, characterized in that: The image denoising device is used to run an image processing platform, which is built based on the U-Net network architecture. The image processing platform includes an encoder and a decoder. The network layer structure of the encoder is a plurality of hole convolution layers of different specifications, and an attention layer is provided between two adjacent hole convolution layers. The network layer structure of the decoder is a deconvolution layer corresponding to the plurality of hole convolution layers of different specifications. The image denoising device includes: A downsampling operation module is used to obtain a picture to be denoised, perform a downsampling operation on the picture to be denoised, and obtain a plurality of input pictures, wherein each input picture corresponds to an input of the dilated convolution layer; A dilated convolution module is configured to input each of the input images into the corresponding dilated convolution layer, obtain a feature map output by each dilated convolution layer, and obtain a plurality of first feature maps; an attention weight module, configured to calculate an attention weight of the first feature map based on the attention layer; A feature weighting module, configured to perform feature weighting on the first feature map using the attention weight to obtain a plurality of weighted feature maps; A deconvolution module, configured to input each of the weighted feature maps into a corresponding deconvolution layer to obtain a plurality of second feature maps; The feature map fusion module is used to perform feature fusion on a plurality of the second feature maps to obtain a denoised image.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of the image denoising method according to any one of claims 1 to 7 when executing the computer-readable instructions.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the image denoising method according to any one of claims 1 to 7.

Images