High-resolution image moire removing method, system, equipment and medium

Abstract

The invention belongs to the field of image recovery of computer vision, and discloses a high-resolution image moire removing method, system and device and a medium, the method adopts a U-shaped multi-stage encoder-decoder structure, and aims at solving the problem of moire multi-band distribution; according to the method, feature pyramids of the same semantic level are constructed, cavity residual dense modules are integrated on scale branches of all layers of the feature pyramids, so that multi-scale feature extraction and moire texture removal of the same semantic level are achieved, and multi-level residual dense connection and cavity convolution are combined; performing dynamic fusion on the extracted different scale features of the same semantic level through an enhanced attention gate, and enhancing feature fusion by replacing common convolution with grouped convolution and introducing an efficient channel attention module; a joint loss function of pixel loss, perception loss and color loss is constructed to optimize model training, and a depth supervision strategy is introduced to enhance the training stability.

Description

Technical Field

[0001] The present invention relates to the field of image restoration in computer vision, and particularly to a high-resolution image moiré removal method, system, device and medium based on pyramid feature extraction and attention feature fusion. Background Art

[0002] In the context of the rapid development of information technology, images have become important information carriers and are widely used in scenarios such as education and industry. However, when using a digital camera to photograph the electronic screen of a display device, the obtained images often contain some irregular colored stripes, which we call "moiré patterns". Such patterns not only seriously affect the visual quality of the images but also interfere with subsequent high-level computer vision tasks such as image classification and object detection. The moiré patterns in images are jointly affected by multiple factors such as the camera model, shooting distance, angle, and lighting conditions. The complex and variable texture patterns and the difficulty of wide distribution in multiple frequency bands pose great challenges to the task of image moiré removal. The related research on image moiré removal not only helps to improve the image quality and visual experience but also provides more reliable data preprocessing support for subsequent high-level computer vision tasks, having important application value and research significance.

[0003] In the task of image moiré removal, traditional methods include filtering, interpolation, and image decomposition methods. Filtering methods usually regard moiré patterns as noise signals and suppress high-frequency signals through low-pass filters or notch filters. This method is prone to problems such as loss of image details and edge blurring. Interpolation methods reduce colored stripes by reconstructing pixel information, and their effects are limited in areas with complex textures and rich details. Image decomposition methods attempt to separate image content from moiré interference, and they rely relatively heavily on prior conditions and lack flexibility.

[0004] Compared with traditional image moiré removal methods, deep learning-based image moiré removal methods utilize the powerful feature learning ability of neural networks to more effectively identify and remove moiré patterns, further improving the moiré removal effect. However, existing methods are mainly designed based on low-resolution moiré images. When dealing with moiré patterns in high-resolution images, they still face problems such as difficulty in effectively removing large-scale moiré patterns in complex backgrounds and moiré artifacts remaining after removing moiré textures. How to further improve the moiré removal effect remains a research difficulty that urgently needs to be broken through. Summary of the Invention

[0005] The present invention aims to solve the problems that large-scale moiré patterns in complex backgrounds are difficult to effectively remove and moiré artifacts remain after removing moiré textures in existing methods, and provides a high-resolution image moiré removal method, system, device and medium based on pyramid feature extraction and attention feature fusion.

[0006] In a first aspect, the present invention provides a method for removing moiré patterns from high-resolution images, comprising the following steps: Select an image dataset for the task of removing moiré patterns from high-resolution images, the image dataset comprising a set of images with moiré patterns and a set of real images without moiré patterns; Construct a high-resolution image moiré removal model based on pyramid feature extraction and attention feature fusion, the model adopting a U-shaped encoder-decoder architecture, and constructing feature pyramids at the same semantic level in each encoder and decoder for feature extraction, removing moiré textures through a dilated residual dense module, and dynamically fusing different-scale features at the same semantic level through an enhanced attention gate; Construct a joint loss function for the model, the joint loss function comprising a pixel loss function, a perceptual loss function, and a color loss function; Use the image dataset and the joint loss function to train the model so that the model learns the conversion from images with moiré patterns to real images without moiré patterns; Use the trained model to remove moiré patterns from high-resolution images.

[0007] As an optional implementation manner of the first aspect of the present application, the U-shaped encoder-decoder architecture comprises multiple layers of encoders and decoders, the encoder includes a pixel rearrangement downsampling operation, the decoder includes a pixel rearrangement upsampling operation, and the output of the decoder in the middle layer is used to cooperate with a deep supervision strategy to optimize model training.

[0008] As an optional implementation manner of the first aspect of the present application, the encoder and the decoder structures include: feature pyramids at the same semantic level for feature extraction; dilated residual dense modules for removing moiré textures on different-scale branches of the feature pyramids; enhanced attention gates for dynamically fusing different-scale features at the same semantic level extracted by the feature pyramids; and semantic alignment of different-scale features is achieved through bilinear interpolation upsampling.

[0009] As an optional implementation manner of the first aspect of the present application, the dilated residual dense module combines multi-level residual dense connections and dilated convolutions, and the dilated convolutions adopt a dilated convolution array with a dilation rate array.

[0010] As an optional implementation manner of the first aspect of the present application, the enhanced attention gate extracts multi-scale features by replacing the ordinary convolution of the attention gate with a grouped convolution, and suppresses irrelevant channels and enhances key channel features by introducing an efficient channel attention module, so as to achieve dynamic fusion of different-scale features.

[0011] As an alternative implementation of the first aspect of the present application, the combined loss function includes: a pixel loss function, which uses the L1 Charbonnier loss function; a perceptual loss function, which extracts feature maps based on one layer of a pre-trained VGG-19 network and calculates the L2 distance between the real image and the model-predicted image in the feature space; a color loss function, which calculates the L1 norm according to the U and V components of the real image and the model-predicted image in the YUV color space.

[0012] As an alternative implementation of the first aspect of the present application, the combined loss function incorporates a deep supervision strategy, and its calculation formula is: , where represents the output feature of the th layer decoder of the model, represents the corresponding real image, represents the L1 Charbonnier loss between and , represents the perceptual loss between and , represents the color loss between and represents the weighting coefficient of the color loss.

[0013] In a second aspect, an embodiment of the present application provides a high-resolution image demoireing system, including: A dataset selection module for selecting an image dataset for the high-resolution image demoireing task, where the image dataset includes a set of images with moire patterns and a set of real images without moire patterns; A model construction module for constructing a high-resolution image demoireing model based on pyramid feature extraction and attention feature fusion. The model uses a U-shaped encoder-decoder architecture, constructs feature pyramids at the same semantic level in each encoder and decoder for feature extraction, removes moire patterns through a dilated residual dense module, and dynamically fuses different-scale features at the same semantic level through an enhanced attention gate; A loss function construction module for constructing a combined loss function for the model, where the combined loss function includes a pixel loss function, a perceptual loss function, and a color loss function; A model training module for training the model using the image dataset and the combined loss function, enabling the model to learn the conversion from images with moire patterns to real images without moire patterns; A demoireing module for removing moire patterns from high-resolution images using the trained model.

[0014] In a third aspect, an embodiment of the present application provides an electronic device, which includes a processor, a memory, and a program or instruction stored on the memory and executable on the processor. When the program or instruction is executed by the processor, the steps of the method described in the first aspect are implemented.

[0015] In a fourth aspect, an embodiment of the present application provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the method described in the first aspect are implemented.

[0016] Compared with the prior art, the present invention proposes a high-resolution image demoireing method based on pyramid feature extraction and attention feature fusion for removing moire patterns in high-resolution images. The method uses a U-shaped multi-level encoder-decoder structure to solve the problem of multi-band distribution of moire patterns; through the constructed feature pyramid at the same semantic level and the dilated residual dense modules on the three scale branches of the feature pyramid, multi-scale feature extraction at the same semantic level and moire texture removal are realized to solve the problem that large-scale moire patterns are difficult to effectively remove in complex backgrounds; through the enhanced attention gate to dynamically fuse different scale features at the same semantic level extracted, the detail information recovery and color restoration effect in the moire artifact area are effectively improved to solve the problem of remaining moire artifacts after removing moire patterns; through the proposed joint loss function to optimize model training, thereby enhancing the stability of training. The method effectively solves the key problems existing in the current high-resolution image demoireing task, such as difficult to accurately distinguish moire patterns in complex backgrounds, poor removal effect of large-scale moire patterns, loss of detail information in demoireed images, and color distortion, showing excellent demoireing effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flowchart of a high-resolution image demoireing method according to an embodiment of the present invention; Figure 2 is a structural diagram of a high-resolution image demoireing model based on pyramid feature extraction and attention feature fusion; Figure 3 is a structural diagram of an encoder / decoder according to an embodiment of the present invention; Figure 4 is a structural diagram of a dilated residual dense module (DRDB) according to an embodiment of the present invention; Figure 5 is a structural diagram of an enhanced attention gate (EAG) according to an embodiment of the present invention; Figure 6 is a schematic structural diagram of a high-resolution image demoireing system provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] Next, in combination with the accompanying drawings in the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0019] The terms "first", "second", etc. in the description and claims of the present application are used to distinguish similar objects, rather than to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein. In addition, "and / or" in the description and claims means at least one of the connected objects. The character " / " generally represents an "or" relationship between the associated objects before and after. In the description of the present invention, "a plurality of" means two or more, unless otherwise specifically defined.

[0020] Embodiment 1 Please refer to Figure 1 , which is a flowchart of a high-resolution image demoireing method provided by an embodiment of the present invention. The method may include the following steps: S1: Select an image dataset for the high-resolution image demoireing task. The image dataset includes a set of images with moiré and a set of real images without moiré.

[0021] It can be understood that a suitable moiré image dataset is of great significance for promoting the research of the image demoireing task. The current publicly available high-resolution moiré image datasets are the FHDMi and UHDM datasets, which can be used for model training. Exemplarily, the FHDMi dataset contains 12,000 pairs of high-resolution images obtained by real-scene shooting, and the resolution of each pair is , covering scenarios such as wallpapers, sports video frames, movie clips, documents, etc.; the UHDM dataset contains 5,000 pairs of 4K high-definition images, and its resolution is , covering scenarios such as landscapes, sports, video frames, and documents. In the image dataset, represents the set of images with moiré, and represents the set of real images without moiré. Then each pair of paired image data can be expressed as , where , .

[0022] S2: Construct a high-resolution image demoireing model based on pyramid feature extraction and attention feature fusion. The model adopts a U-shaped encoder-decoder architecture, and feature pyramids at the same semantic level are constructed in each encoder and decoder for feature extraction. The moire texture is removed through the dilated residual dense block, and different-scale features at the same semantic level are dynamically fused through the enhanced attention gate.

[0023] The high-resolution image demoireing model based on pyramid feature extraction and attention feature fusion uses a convolutional neural network with a U-shaped encoder-decoder architecture as the backbone network. First, to solve the problem of difficultly effectively removing large-scale moire in complex backgrounds, this model constructs feature pyramids at the same semantic level in each encoder and decoder for feature extraction, and the moire texture is removed through the DRDB module on each layer scale branch of the feature pyramid. Subsequently, for the moire artifacts remaining after removing the moire texture, this model dynamically fuses different-scale features at the same semantic level through the EAG to achieve information complementarity at the same semantic level, thereby improving the detail information restoration and color restoration effects in the moire artifact area.

[0024] Refer to Figure 2 As shown in, the overall structure of the high-resolution image demoireing model based on pyramid feature extraction and attention feature fusion is U-shaped, including 3 layers of encoders and decoders. The input image is first downsampled through pixel rearrangement (PixelUnshuffle) and then input to the encoder E1. The encoding result is not only passed to the decoder D1 at the same layer, but also downsampled through bilinear interpolation and then input to the next-level encoder E2. The encoder E2 is similar to E1, and the encoding result of the encoder E3 is directly passed to the decoder D3 at the same layer. The decoding results of each decoder are output after upsampling through pixel rearrangement (Pixel Shuffle). At the same time, the decoding results of the decoder D3 and the decoder D2 are also upsampled through bilinear interpolation and then passed to the upper-layer decoder. The decoding results output by the middle-layer decoder are used to introduce a deep supervision strategy to provide multi-level supervision signals.

[0025] Refer to Figure 3, the decoder structure in the model is the same as the encoder. The core modules include a feature pyramid at the same semantic level for feature extraction, a DRDB module for removing moiré textures, and an EAG module for removing moiré artifacts. First, feature extraction is performed by constructing a feature pyramid at the same semantic level. The DRDB module removes moiré textures at different scales of the feature pyramid. To ensure semantic consistency of features during subsequent fusion, bilinear interpolation upsampling is used to achieve semantic alignment for subsequent feature fusion operations. By improving the Attention Gate (AG), EAG can dynamically fuse features at different scales, achieving information complementarity between different-scale features at the same semantic level.

[0026] Refer to Figure 4 , the DRDB module combines multi-level residual dense connections and dilated convolutions, aiming to enhance the model's ability to remove large-scale moiré textures. The module expands the receptive field during feature extraction without increasing the computational cost through a dilated convolution array with a zigzag dilation rate array, effectively capturing large-scale moiré features and avoiding the "grid effect" of dilated convolutions. At the same time, through the combination of the dilated convolution array and residual dense connections, not only is the problem of vanishing gradients caused by network deepening alleviated, but also the detailed information of the image can be retained during the transmission of deep-layer information.

[0027] Refer to Figure 5 , the EAG module is an improvement of AG. The main function of AG is to enhance the attention to important regions by adaptively focusing on the feature information of important regions. However, AG requires a high level of semantic consistency between input features to accurately capture key features. When the correlation between input features is weak, AG will not be able to effectively extract important information. To solve the above problems, EAG replaces the ordinary convolution of AG with grouped convolution, thereby extracting richer multi-scale features without significantly increasing the computational burden. At the same time, an Efficient Channel Attention (ECA) module is introduced to suppress irrelevant channels and enhance the features of key channels, thereby weakening the negative impact when the correlation between high-level and low-level features is weak. EAG can dynamically fuse features at different scales at the same semantic level through the improved attention gate mechanism, achieving information complementarity between different-scale features at the same semantic level.

[0028] S3: Construct a joint loss function for the model. The joint loss function includes a pixel loss function, a perceptual loss function, and a color loss function.

[0029] For the model proposed by this method, a joint loss function needs to be constructed during training. This joint loss function includes a pixel loss function, a perceptual loss function, and a color loss function. By combining the above three loss functions, the final joint loss function is obtained. Using to represent the set of moiré images and to represent the set of real images, the training process of the model can be expressed as follows: Given the paired image data and , use the proposed model and the joint loss function to learn the conversion from moiré images to real moiré-free images.

[0030] The pixel loss function is usually used to represent the pixel-level difference between the real image and the moiré-removed image output by the model. Here, the pixel-level L1 Charbonnier loss function is selected. The pixel loss function is: where and represent the real image and the moiré-removed image output by the model respectively, is the smoothing term, and is taken during the model training process.

[0031] To make up for the deficiency of the pixel-level loss and make the moiré-removed image output by the model closer to the subjective perception of people, thus better measuring the similarity between the moiré-removed image and the real image, a perceptual loss function is further introduced. The perceptual loss function is: where and represent the real image and the moiré-removed image output by the model respectively, represents the feature map of the pre-trained VGG-19 network at the th layer. During the model training process , represents the square of the L2 norm value, represents the number of channels of the feature map extracted by the pre-trained network at the th layer, and represent the height and width of the feature map extracted by the pre-trained VGG-19 network at the th layer respectively, represents the perceptual loss between the real image and the moiré-removed image output by the model.

[0032] In the task of image demoireing, the color distortion problem of the image will further cause the loss of image details and color information, seriously affecting the quality and visual effect of the demoireed image. To improve the color restoration effect of the demoireed image of the model, a color loss function is introduced. By constraining the color information of the model in the YUV color space, the accuracy of color restoration is improved. The color loss function is: Where, and respectively represent the U component of the demoireed image output by the model and the real image in the YUV color space, and respectively represent the V component of the demoireed image and the real image in the YUV color space, represents the L1 norm.

[0033] Combined with the deep supervision strategy, the joint loss function of the model can be expressed as: Where, represents the output feature of the th layer decoder of the model, represents the corresponding real image, represents and the L1 Charbonnier loss between them, represents and the perceptual loss between them, represents and the color loss between them, represents the weighting coefficient of the color loss.

[0034] S4: Use the image dataset and the joint loss function to train the model so that the model learns the conversion from the moire image to the real moire-free image.

[0035] After selecting the required dataset and completing the construction of the model and the loss function, the model needs to be trained. In multiple rounds of training, the model gradually learns the conversion from the moire image to the real moire-free image. After the training is completed, the moire removal ability of the model is tested. When a high-resolution image containing moire is input, the model can output a high-resolution moire-free image.

[0036] S5: Use the trained model to remove the moire from the high-resolution image.

[0037] After the model training and verification are completed, the moiré removal ability of the model can be used. The high-resolution moiré image is used as the input of the model, and the trained model is used to implement the moiré removal task for the image.

[0038] In summary, in this embodiment, the present invention proposes a high-resolution image moiré removal method based on pyramid feature extraction and attention feature fusion. The method adopts a U-shaped multi-level encoder-decoder structure, aiming to address the problem of multi-band distribution of moiré. By constructing feature pyramids at the same semantic level in each encoder and decoder, and applying dilated residual dense blocks (DRDB) on three scale branches of the pyramid, multi-scale feature extraction at the same semantic level and moiré texture removal are achieved. Further, through the enhanced attention gate (EAG), dynamic fusion of different scale features at the same semantic level is performed to enhance the restoration of detail information and color reproduction effect in the moiré artifact area. The method optimizes the model training through the proposed joint loss function (including pixel loss, perceptual loss, and color loss, and introducing a deep supervision strategy), thereby enhancing the stability of training and finally realizing moiré removal for high-resolution images.

[0039] Embodiment 2 Please refer to Figure 6 , which shows a schematic structural diagram of a high-resolution image moiré removal system proposed in the second embodiment of the present application. The system includes the following key modules: The dataset selection module 100 is used to select an image dataset for the high-resolution image moiré removal task. The image dataset includes a set of images with moiré and a set of real images without moiré; The model construction module 200 is used to construct a high-resolution image moiré removal model based on pyramid feature extraction and attention feature fusion. The model adopts a U-shaped encoder-decoder architecture, and feature pyramids at the same semantic level are constructed in each encoder and decoder for feature extraction. Moiré textures are removed through dilated residual dense blocks, and dynamic fusion of different scale features at the same semantic level is performed through the enhanced attention gate; The loss function construction module 300 is used to construct a joint loss function for the model. The joint loss function includes a pixel loss function, a perceptual loss function, and a color loss function; The model training module 400 is used to train the model using the image dataset and the joint loss function, so that the model learns the conversion from moiré images to real images without moiré; The moiré removal module 500 is used to remove moiré in high-resolution images using the trained model.

[0040] A high-resolution image moiré removal system in an embodiment of the present application may be a device, or a component, an integrated circuit, or a chip in a terminal. The device may be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device may be a mobile phone, a tablet computer, a laptop computer, a handheld computer, a vehicle-mounted electronic device, a wearable device, an Ultra-mobile Personal Computer (UMPC), a netbook, or a Personal Digital Assistant (PDA), etc. The non-mobile electronic device may be a server, a Network Attached Storage (NAS), a Personal Computer (PC), etc. The embodiments of the present application do not make specific limitations.

[0041] A high-resolution image moiré removal system in an embodiment of the present application may be a device with an operating system. The operating system may be an Android operating system, an IOS operating system, or other possible operating systems. The embodiments of the present application do not make specific limitations.

[0042] A high-resolution image moiré removal system provided by the embodiments of the present application can implement Figure 1 each process implemented by a high-resolution image moiré removal method in the method embodiment. To avoid repetition, it will not be elaborated here.

[0043] Optionally, an embodiment of the present application further provides an electronic device, including a processor, a memory, and a program or instruction stored on the memory and executable on the processor. When the program or instruction is executed by the processor, it implements each process of the above-mentioned high-resolution image moiré removal method embodiment and can achieve the same technical effect. To avoid repetition, it will not be elaborated here.

[0044] An embodiment of the present application further provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by the processor, it implements each process of the above-mentioned high-resolution image moiré removal method embodiment and can achieve the same technical effect. To avoid repetition, it will not be elaborated here.

[0045] Wherein, the processor is the processor in the electronic device described in the above embodiment. The readable storage medium includes a computer-readable storage medium, such as a computer Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, or an optical disc, etc.

[0046] It should be noted that in this text, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article or device. Without more limitations, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including that element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in the reverse order according to the functions involved. For example, the described methods may be performed in an order different from that described, and various steps may be added, omitted, or combined. Additionally, the features described with reference to certain examples may be combined in other examples.

[0047] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-described example methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions for causing a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in various embodiments of the present application.

[0048] The embodiments of the present application have been described above in conjunction with the accompanying drawings. However, the present application is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Those of ordinary skill in the art, under the inspiration of the present application and without departing from the purpose of the present application and the scope protected by the claims, can still make many forms, all of which fall within the protection scope of the present application.

Claims

1. A high-resolution image demoireing method, characterized in that Including the following steps: Select an image dataset for the task of moiré removal from high-resolution images, where the image dataset includes a set of images with moiré and a set of real images without moiré; Construct a high-resolution image moiré removal model based on pyramid feature extraction and attention feature fusion. The model adopts a U-shaped encoder-decoder architecture, and constructs feature pyramids at the same semantic level in each encoder and decoder for feature extraction. Remove moiré textures through a dilated residual dense module, and dynamically fuse different-scale features at the same semantic level through an enhanced attention gate; Construct a joint loss function for the model, where the joint loss function includes a pixel loss function, a perceptual loss function, and a color loss function; Use the image dataset and the joint loss function to train the model so that the model learns the conversion from images with moiré to real images without moiré; Use the trained model to remove moiré from high-resolution images.

2. The method according to claim 1, wherein The U-shaped encoder-decoder architecture includes multiple layers of encoders and decoders. The encoder includes a pixel rearrangement downsampling operation, and the decoder includes a pixel rearrangement upsampling operation. The output of the decoder in the middle layer is used to cooperate with the deep supervision strategy to optimize model training.

3. The method according to claim 1, characterized in that, The encoder and the decoder structure include: Feature pyramids at the same semantic level for feature extraction; A dilated residual dense module for removing moiré textures on different-scale branches of the feature pyramid; An enhanced attention gate for dynamically fusing different-scale features at the same semantic level extracted by the feature pyramid; And semantic alignment of different-scale features is achieved through bilinear interpolation upsampling.

4. The method according to claim 1 or 3, characterized in that, The dilated residual dense module combines multi-level residual dense connections and dilated convolutions, and the dilated convolution adopts a dilated convolution array with a zigzag dilation rate array.

5. The method according to claim 1 or 3, characterized in that The enhanced attention gate extracts multi-scale features by replacing the ordinary convolution of the attention gate with grouped convolution, and suppresses irrelevant channels and enhances key channel features by introducing an efficient channel attention module to achieve dynamic fusion of different-scale features.

6. The method according to claim 1, characterized in that, The joint loss function includes: A pixel loss function, adopting an L1 Charbonnier loss function; A perceptual loss function, extracting feature maps based on a layer of a pre-trained VGG-19 network, and calculating the L2 distance between the real image and the model-predicted image in the feature space; A color loss function, calculating the L1 norm according to the U component and V component of the real image and the model-predicted image in the YUV color space.

7. The method according to claim 6, characterized in that The joint loss function combines a deep supervision strategy, and its calculation formula is: ， Among them, represents the output features of the -th layer decoder of the model, represents the corresponding real image, represents the L1 Charbonnier loss between and ; represents the perceptual loss between and ; represents the color loss between represents the weighting coefficient of the color loss.

8. A high-resolution image moiré removal system, characterized in that, Including: A dataset selection module for selecting an image dataset for the task of moiré removal from high-resolution images, where the image dataset includes a set of images with moiré and a set of real images without moiré; A model construction module, configured to construct a high-resolution image demoireing model based on pyramid feature extraction and attention feature fusion. The model adopts a U-shaped encoder-decoder architecture, and feature pyramids at the same semantic level are constructed in each encoder and decoder for feature extraction. Moire patterns are removed through a dilated residual dense module, and different-scale features at the same semantic level are dynamically fused through an enhanced attention gate; A loss function construction module, configured to construct a joint loss function for the model. The joint loss function includes a pixel loss function, a perceptual loss function, and a color loss function; A model training module, configured to use the image dataset and the joint loss function to train the model, enabling the model to learn the conversion from moire-patterned images to real moire-free images; A demoireing module, configured to use the trained model to remove moire patterns from high-resolution images.

9. An electronic device, characterized in that, Comprising a processor, a memory, and a program or instructions stored on the memory and executable on the processor. When the program or instructions are executed by the processor, the steps of a high-resolution image demoireing method according to any one of claims 1-7 are implemented.

10. A readable storage medium, characterized in that, Programs or instructions are stored on the readable storage medium. When the programs or instructions are executed by a processor, the steps of a high-resolution image demoireing method according to any one of claims 1-7 are implemented.

Images