An image denoising method based on frequency space joint guided dynamic kernel generation network

Abstract

The application provides an image denoising method based on frequency space joint guided dynamic kernel generation network, which comprises the following steps: inputting a noisy image into an initial convolution layer to obtain a shallow feature representation; adopting an encoder to perform hierarchical down-sampling and context extraction, and recovering spatial details through a jump connection in a decoder; adopting a fixed two-dimensional discrete cosine transform filter bank to perform low-frequency, medium-frequency and high-frequency decomposition on the input feature to construct a frequency prior; fusing spatial context information, channel interaction information and the frequency prior to generate a dynamic convolution kernel; performing frequency space joint guided dynamic feature extraction on multiple receptive field scales; interacting and fusing channel, height and width dimension information through a multi-dimensional adaptive fusion mechanism; finally, completing detail recovery in the decoding stage and outputting a denoised image; and the application alleviates the confusion between high-frequency texture and noise residual, and improves the detail preservation capability and denoising robustness in complex texture, strong noise and scale change scenes.

Description

Technical Field

[0001] This invention relates to the field of image denoising and image restoration technology, and in particular to an image denoising method based on a frequency-space jointly guided dynamic kernel generation network. Background Technology

[0002] Image denoising aims to recover potentially sharp images from noisy observations and is one of the fundamental problems in low-level vision and computational imaging. Under real-world acquisition conditions, image noise originates from complex sources, including sensor thermal noise, quantization noise, and readout noise, as well as factors such as insufficient illumination, limited exposure, imaging chain compression, and transmission interference. The presence of noise significantly weakens edge, texture, and fine-grained structural information in images, further affecting the robustness and accuracy of downstream vision tasks such as object detection, image segmentation, and scene understanding.

[0003] Existing image denoising methods can be broadly categorized into three types: methods based on traditional priors, methods based on convolutional neural networks (CNNs), and methods based on Transformers. Traditional methods typically rely on artificially constructed statistical or structural priors, such as sparse representations, self-similarity, low-rank constraints, or variational models. These methods are effective when noise distributions are simple and the degradation process is known, but they often struggle to balance denoising strength and detail preservation in complex texture regions and realistically noisy scenes. CNN-based methods establish a mapping between noisy and clear images through end-to-end learning, significantly improving denoising performance. Transformer-based methods excel in long-range dependency modeling, enhancing global contextual representation. However, in scenes with strong noise, complex textures, and multi-scale structures, existing methods still generally suffer from insufficient detail recovery, easy confusion between real high-frequency textures and noise residuals, and high computational costs.

[0004] Furthermore, while existing dynamic convolution or dynamic filtering methods enhance the spatial adaptability of models to some extent, their convolution kernel generation process typically relies primarily on spatial or channel features, lacking explicit differentiation between low-frequency, mid-frequency, and high-frequency information. This makes it easy for high-frequency details and noise components to alias, even under severe degradation. Simultaneously, there is usually a lack of a unified and stable collaborative mechanism between frequency domain enhancement, multi-scale feature extraction, and spatial adaptive filtering, making it difficult for models to simultaneously achieve local detail restoration, overall structural consistency, and inference efficiency.

[0005] Therefore, there is an urgent need to propose an image denoising method that can explicitly model information in different frequency bands and directly involve frequency priors in the dynamic convolution kernel generation process, so as to achieve more stable and efficient image restoration in scenarios with complex textures, strong noise and scale changes. Summary of the Invention

[0006] This invention proposes an image denoising method based on a frequency-space jointly guided dynamic kernel generation network to address the problems in existing image denoising methods, such as the easy confusion of high-frequency details with noise residuals, insufficient coupling between frequency modeling and spatial adaptive filtering, limited multi-scale structure recovery capability, and high computational cost of complex models.

[0007] This invention provides an image denoising method based on a frequency-space jointly guided dynamic kernel generation network, the method comprising the following steps:

[0008] Step S1: Obtain the noisy input image and extract shallow features through the initial convolutional layer;

[0009] Step S2: Input the shallow features into a U-shaped encoder-decoder network, which includes an encoder and a decoder. The encoder extracts multi-level context features through downsampling, and the decoder recovers the spatial resolution step by step through upsampling and skip connections to obtain the decoded features.

[0010] Step S3: Use a fixed two-dimensional discrete cosine transform filter bank to perform frequency decomposition on the decoded features to obtain multiple frequency responses. Divide the frequency responses into low-frequency set, mid-frequency set and high-frequency set according to frequency range, and aggregate them to obtain low-frequency structural features, mid-frequency texture features and high-frequency detail features respectively.

[0011] Step S4: Extract the spatial context representation and channel enhancement representation of the decoded features, fuse the spatial context representation, channel enhancement representation, low-frequency structural features, mid-frequency texture features and high-frequency detail features into a unified descriptor, and generate multiple dynamic convolution kernels with preset receptive field scales based on the unified descriptor;

[0012] Step S5: Apply the dynamic convolution kernels at each preset receptive field scale to the input features to obtain the dynamic response at the corresponding scale. Then, convolve and aggregate the dynamic responses at each scale to obtain a multi-scale feature representation.

[0013] Step S6: For the multi-scale feature representation, construct descriptors along the channel dimension, height dimension, and width dimension respectively, obtain fused features based on cross-dimensional attention, and then combine convolution and deformable convolution to obtain enhanced features;

[0014] Step S7: Convolve the enhanced features through the output mapping to obtain the denoised image.

[0015] Furthermore, in step S2, the encoder extracts multi-level contextual representations step by step using stride convolution and convolutional blocks; the decoder restores spatial resolution step by step through upsampling and supplements shallow detail information through jump connections from the encoder to the decoder, thereby enhancing the global semantic expression capability while maintaining the restoration capability of local texture and edge structure.

[0016] Further, the specific method for performing frequency decomposition on the decoded features using a fixed two-dimensional discrete cosine transform filter bank in step S3 to obtain multiple frequency responses, and dividing the frequency responses into low-frequency set, mid-frequency set, and high-frequency set according to frequency range, and aggregating them to obtain low-frequency structural features, mid-frequency texture features, and high-frequency detail features respectively, includes:

[0017] Employing a predefined two-dimensional discrete cosine transform (DCT) filter bank Convolution is performed on the input features to obtain the frequency response:

[0018]

[0019] A single filter kernel in the two-dimensional DCT filter bank can be represented as:

[0020]

[0021] in, and These represent the horizontal frequency index and the vertical frequency index, respectively. and Represents the spatial coordinates in the filter kernel. Indicates the side length of the filter kernel. Represents the normalization coefficient;

[0022] After obtaining the frequency response, it is divided into a low-frequency set according to the frequency range. Mid-frequency collection and high frequency set The low-frequency structural features were obtained by aggregation. Mid-frequency texture features and high-frequency detail features The above polymerization process can be represented as:

[0023]

[0024]

[0025]

[0026] in, Indicates the first The response map obtained after applying DCT filter kernels to the input features , and These represent the sets of low-frequency filter kernels, mid-frequency filter kernels, and high-frequency filter kernels, respectively. , and These represent the number of elements in the corresponding sets. Represents the normalization operator;

[0027] The normalization operator is used to reduce the differences in numerical scale between features of different frequency bands and improve the stability of subsequent fusion and dynamic kernel generation processes.

[0028] Further, the specific method for extracting the spatial context representation and channel enhancement representation of the decoded features in step S4, fusing the spatial context representation, channel enhancement representation, low-frequency structural features, mid-frequency texture features, and high-frequency detail features into a unified descriptor, and generating multiple dynamic convolutional kernels with preset receptive field scales based on the unified descriptor includes:

[0029] Input features for the decoding stage Joint representation modeling is performed, extracting spatial context information and channel interaction information respectively, and combining low-frequency structural features. Mid-frequency texture features and high-frequency detail features Construct a unified descriptor for dynamic convolution kernel generation. The unified descriptor satisfies:

[0030]

[0031] in, Representation of spatial context, This indicates channel enhancement. The unified descriptor represents the fusion mapping function. Spatial structural attributes, channel response attributes, and frequency distribution attributes used to comprehensively characterize input features;

[0032] Based on the unified descriptor For a preset scale set Different scales in Adaptively generate dynamic convolution kernels and perform dynamic convolution on the input features respectively to obtain multi-scale dynamic responses, where the dynamic convolution kernels at the corresponding scales satisfy:

[0033]

[0034] in, This represents a dynamic kernel generation function.

[0035] Furthermore, the decoding stage includes a Frequency Spatial Filtering Block (FSDB), which comprises a frequency-aware extractor, a frequency-space jointly guided dynamic kernel generation module, a dynamic convolutional layer, an activation function, and residual connections. Its output satisfies the following:

[0036]

[0037] Where DynConv(·) represents the dynamic convolution operation based on the normalized dynamic convolution kernel Ŵ, and δ(·) represents the activation function.

[0038] Furthermore, the specific method for applying the dynamic convolution kernels corresponding to each scale to the input features in step S5 to obtain the dynamic response at the corresponding scale, and then concatenating the dynamic responses at each scale and aggregating them through convolution to obtain the multi-scale feature representation includes:

[0039] Apply dynamic convolution kernels at the corresponding scale to the input features. The multi-scale dynamic response is obtained:

[0040]

[0041] The dynamic responses at different scales are then concatenated to obtain joint features:

[0042]

[0043] A unified multi-scale representation is obtained by 1×1 convolution aggregation:

[0044]

[0045] in, Representing scale Dynamic convolution kernels It represents the multi-scale features after aggregation; through multi-scale dynamic convolution response modeling, it achieves joint enhancement of image structure and texture information under different receptive fields.

[0046] Furthermore, the preset scale set S in step S5 is a multi-scale set containing different expansion rates or receptive fields, preferably S = {1, 3, 5}.

[0047] Further, step S6 involves constructing descriptors along the channel, height, and width dimensions for the multi-scale feature representation, obtaining fused features based on cross-dimensional attention, and then combining convolution and deformable convolution to obtain enhanced features. Specific methods for this include:

[0048] Multidimensional adaptive fusion is achieved through a multidimensional adaptive fusion module (MDAF), which processes multi-scale features. Pooling and projection are performed along different dimensions to construct Q, K, and V, satisfying the following:

[0049]

[0050] The attention weight matrix is ​​obtained through cross-dimensional attention calculation:

[0051]

[0052] The fused features are obtained, and the output features are recalibrated by combining convolution and deformable convolution to obtain enhanced features:

[0053]

[0054] in, Indicates fusion characteristics; This indicates a deformable convolution operation.

[0055] Furthermore, the specific method for obtaining the denoised image by convolving the enhanced features with the output mapping in step S7 includes:

[0056]

[0057] in, For denoising the image; This indicates the output mapping convolution operation.

[0058] Furthermore, the decoding stage of the U-shaped codec network includes multiple frequency space filter blocks (FSDB) and at least one multi-scale frequency space filter block (MFSDB), wherein the FSDB is used for local content adaptive reconstruction, and the MFSDB is used for multi-scale feature enhancement and cross-dimensional fusion in scenarios with complex structures and scale changes.

[0059] Compared with the prior art, the present invention has the following advantages:

[0060] 1. This invention performs explicit frequency decomposition on the input features by using a fixed two-dimensional discrete cosine transform filter bank, and constructs low-frequency structural features, mid-frequency texture features and high-frequency detail features respectively, thereby achieving decoupled representation of information in different frequency bands, effectively alleviating the confusion between high-frequency texture and noise residuals, and improving the ability to preserve details in noisy scenes.

[0061] 2. This invention integrates spatial context information, channel interaction information, and frequency prior information into a unified descriptor, and adaptively generates dynamic convolution kernels based on this unified descriptor, thereby realizing integrated modeling of frequency information and spatial adaptive filtering, enabling the model to adaptively adjust filtering behavior for different regional contents.

[0062] 3. This invention uses a multi-scale dynamic feature extraction and multi-dimensional adaptive fusion mechanism to jointly model local texture information and global structural information under different receptive fields. It also enhances the representation ability of complex regions through cross-dimensional interaction between channel dimension, height dimension and width dimension, thereby maintaining low computational complexity while taking into account denoising accuracy and engineering practicality.

[0063] Based on the implementation methods provided in the above aspects, this application can be further combined to provide more implementation methods. Attached Figure Description

[0064] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein:

[0065] Figure 1 This is a schematic diagram of the overall network structure in an embodiment of the present invention;

[0066] Figure 2 This is a schematic diagram of the frequency space joint guidance dynamic kernel generation module in an embodiment of the present invention;

[0067] Figure 3 This is a schematic diagram of the frequency sensing extractor structure in an embodiment of the present invention;

[0068] Figure 4 This is a schematic diagram of the structure of the frequency-space jointly guided multi-scale dynamic feature extraction module and the multi-dimensional adaptive fusion module in an embodiment of the present invention;

[0069] Figure 5 The flowchart provided in this embodiment of the invention provides an image denoising method based on a frequency-space jointly guided dynamic kernel generation network. Detailed Implementation

[0070] The exemplary embodiments disclosed in this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0071] Combination Figure 5 As shown, an optional embodiment of the present invention provides an image denoising method based on a frequency-space jointly guided dynamic kernel generation network, the method comprising the following steps:

[0072] In this embodiment, the input is a size of Noisy RGB image First, through the initial... Convolution maps the input image to a shallow feature space, resulting in a shallow feature representation. Subsequently, the shallow features The input is a U-shaped encoding / decoding network. The U-shaped encoding / decoding network consists of an encoder and a decoder. The encoder uses stride convolution and convolutional blocks to extract multi-level contextual representations step by step. The decoder recovers spatial resolution step by step through upsampling and supplements shallow detail information through skip connections from the encoder to the decoder, thereby enhancing the global semantic expression capability while maintaining the recovery capability of local texture and edge structure.

[0073] To address the challenge of explicitly distinguishing different frequency information using conventional convolution, this embodiment introduces a frequency-aware extractor during the decoding stage. Let the input feature map be... .

[0074] Employing a predefined two-dimensional discrete cosine transform (DCT) filter bank Convolution is performed on the input features to obtain the frequency response:

[0075]

[0076] A single filter kernel in the two-dimensional DCT filter bank can be represented as:

[0077]

[0078] in, and These represent the horizontal frequency index and the vertical frequency index, respectively. and Represents the spatial coordinates in the filter kernel. Indicates the side length of the filter kernel. This represents the normalization coefficient.

[0079] After obtaining the frequency response, it is divided into a low-frequency set according to the frequency range. Mid-frequency collection and high frequency set The low-frequency structural features were obtained by aggregation. Mid-frequency texture features and high-frequency detail features Preferably, the above polymerization process can be represented as follows:

[0080]

[0081]

[0082]

[0083] in, This represents the normalization operation. Through the above processing, we can obtain low-frequency priors reflecting the smooth background structure, mid-frequency priors reflecting the mid-scale texture pattern, and high-frequency priors reflecting the edges and fine-grained details, thereby providing interpretable frequency constraints for subsequent dynamic convolution kernel generation.

[0084] Furthermore, explicit frequency decomposition can alleviate the confusion between high-frequency texture and noise residuals because different frequency components in noisy features have significant differences in statistical properties and spatial distribution. Low-frequency responses mainly correspond to the overall brightness distribution, smooth background, and subject contours in the image; mid-frequency responses mainly correspond to local structural transitions and repetitive texture patterns; while high-frequency responses simultaneously contain real edge details and noise perturbations. If uniform filtering is performed only in the original spatial domain, real high-frequency texture and random high-frequency noise are easily treated as "components to be suppressed," leading to loss of detail or preservation of residual noise. By employing a fixed two-dimensional discrete cosine transform filter bank to explicitly decompose the input features and constructing low-frequency structural features separately... Mid-frequency texture features and high-frequency detail features This approach allows for the decoupling of frequency components that were originally mixed in the same feature channel. Furthermore, low-frequency structural features provide stable structural constraints for the reconstruction process, mid-frequency texture features preserve local repetitive patterns and transition region information, and high-frequency detail features characterize edges and fine-grained texture variations. Because different frequency band features are used differentially in subsequent dynamic kernel generation and feature enhancement, the network can more accurately distinguish between "real high-frequency textures that should be retained" and "random high-frequency noise that should be suppressed," thereby improving its ability to preserve details in complex texture regions and under strong noise conditions.

[0085] After obtaining the frequency prior, this embodiment further constructs a frequency-space jointly guided dynamic kernel generation module. This module extracts the spatial context representation of the input features through lightweight convolution. On the other hand, channel-enhanced representations are obtained through channel interaction branches. and represent the spatial context. Channel enhancement indicates Low-frequency structural characteristics Mid-frequency texture features and high-frequency detail features Through the fusion function Perform joint modeling to form a unified descriptor:

[0086]

[0087] Wherein, the unified descriptor It comprehensively represents spatial context information, channel interaction information, and frequency prior information, and can be used as a conditional input for the generation of dynamic convolutional kernels.

[0088] Furthermore, directly embedding frequency priors into the dynamic convolution kernel generation process is key to achieving integrated modeling of frequency information and spatial adaptive filtering. Traditional dynamic convolution methods typically generate convolution kernels based solely on spatial local responses or channel statistics. While their filtering behavior can change with the input content, they lack explicit awareness of the differences between different frequency components in the input, potentially leading to over-smoothing or noise residue in textured regions. In this embodiment, the unified descriptor not only includes spatial context representation and channel enhancement representation but also explicitly introduces low-frequency structural features. Mid-frequency texture features and high-frequency detail features Therefore, the generated dynamic convolutional kernel can not only reflect the local spatial semantics of the current location, but also perceive the frequency composition features corresponding to that location. For smooth regions, the generated dynamic convolutional kernel tends to enhance noise suppression capabilities; for edge and texture-rich regions, the generated dynamic convolutional kernel tends to preserve structural details and reduce fine texture destruction. Thus, the dynamic convolutional kernel is no longer a simple spatial adaptive filter, but a content-aware filter that is simultaneously modulated by spatial information, channel relationships, and frequency distribution, thereby achieving integrated modeling of frequency information and spatial adaptive filtering.

[0089] After obtaining the unified descriptor Then, for multiple preset scales Using dynamic kernel generation functions Generate dynamic convolution kernels at the corresponding scale:

[0090]

[0091] in, This represents a preset set of scales, preferably a set of scales containing different receptive fields or different dilation rates. The dynamic convolutional kernel can adaptively change its convolutional weights according to the input content, thus allowing the filtering behavior to be simultaneously constrained by spatial semantics, channel relationships, and frequency priors. Preferably, the dynamic kernel generation function... can be The process consists of convolution, parameter mapping, and shape rearrangement; in some implementations, the dynamic convolution kernel may be further normalized to improve numerical stability and inference robustness.

[0092] In this embodiment, the Frequency Spatial Filtering Block (FSDB) serves as the basic reconstruction unit in the decoding stage, consisting of a frequency-aware extractor, a unified descriptor generation module, a dynamic convolutional layer, an activation function, and residual connections. This structure does not rely solely on fixed convolutional kernels for local filtering; instead, it adaptively adjusts the input features using dynamic convolutional kernels relevant to the current content. Therefore, it can better distinguish between real high-frequency details and noise residuals in texture-dense regions, while suppressing unnecessary high-frequency perturbations in relatively smooth regions. Preferably, the output of the FSDB can be expressed as:

[0093]

[0094] in, This represents the activation function. This represents a dynamic convolution operation. Residual connections enhance texture recovery while preserving the fundamental structural information in the input features, thereby suppressing over-smoothing and artifact accumulation.

[0095] To improve the model's adaptability to scenarios with complex structures and scale variations, this embodiment further incorporates a multi-scale frequency spatial filter block (MFSDB). In this module, a preset scale set is used... Medium scale Apply dynamic convolution kernels at the corresponding scale to the input features. The dynamic responses at different scales were obtained:

[0096]

[0097] Then, the dynamic responses at each scale are concatenated to obtain joint features:

[0098]

[0099] After Convolutional aggregation yields a unified multi-scale feature representation:

[0100]

[0101] Through the multi-scale dynamic feature extraction process described above, the model can simultaneously retain fine-grained texture information and large-scale contextual structure information, thereby enhancing its recovery capabilities in regions with repetitive textures, dense edges, and complex geometric structures.

[0102] Furthermore, the reason why multi-scale dynamic feature extraction can enhance the representation ability in complex structures and scale-changing scenes is that different receptive fields correspond to different levels of image information. Smaller-scale dynamic convolutional responses are more conducive to capturing fine-grained edges, local textures, and subtle structural changes, while larger-scale dynamic convolutional responses are more conducive to modeling large-scale contextual relationships, structural continuity, and global correlations between repetitive patterns. By generating dynamic convolutional kernels and extracting dynamic responses at multiple preset scales, and then aggregating the responses at each scale into a unified multi-scale feature representation. This allows the network to retain both local detail information and global structural information in the same layer of features, thereby improving its adaptability to complex textures, scale-varying targets, and irregular structural regions.

[0103] In obtaining multi-scale features Subsequently, this embodiment further introduces a multi-dimensional adaptive fusion module to enhance the collaborative expressive ability between different scales and different dimensions. Specifically, for the multi-scale features... Pooling and projection are performed along different dimensions to construct channel-dimensional descriptors. High-dimensional descriptors and width dimension descriptor :

[0104]

[0105] in, This indicates pooling along the spatial dimension. This indicates pooling along the channel-width dimension. This indicates pooling along the channel-height dimension; , and These represent the corresponding linear mapping parameters.

[0106] Furthermore, the attention weight matrix is ​​calculated using a cross-dimensional attention mechanism:

[0107]

[0108] And the fusion characteristics are obtained:

[0109]

[0110] After obtaining the fusion features Then, the fused features are recalibrated using convolution and deformable convolution to obtain enhanced features:

[0111]

[0112] in, This represents deformable convolution operations. Through the above multidimensional adaptive fusion process, cross-dimensional dependencies between the channel dimension, height dimension, and width dimension can be established, thereby improving the model's ability to represent complex structural regions, non-uniform noise regions, and cross-scale texture patterns.

[0113] Furthermore, the multidimensional adaptive fusion module does not simply superimpose existing features, but remodels the structural relationships in multi-scale features through cross-dimensional interactions between channel, height, and width dimensions. Specifically, channel descriptors characterize the importance of responses between different feature channels, while height and width descriptors depict the structural distribution and spatial dependencies in the vertical and horizontal directions, respectively. The attention weight matrix calculated through the cross-dimensional attention mechanism allows the model to adaptively select feature components that are more conducive to restoring structure and suppressing noise, and then combines convolution and deformable convolution to complete feature recalibration. Since deformable convolution can further adapt to edge offsets, structural curvatures, and local geometric changes, this process can enhance the model's ability to model complex structural regions, non-uniform noise regions, and cross-scale texture patterns while maintaining low computational complexity, thus balancing denoising accuracy and engineering practicality.

[0114] In obtaining enhanced features Then, the final denoising result is obtained by output mapping convolution:

[0115]

[0116] in, This indicates the output mapping convolution operation. It can be directly output as a denoised image, or it can be further mapped to the target image domain according to the application scenario.

[0117] In terms of network structure, multiple FSDBs are preferably placed in the basic reconstruction layer of the decoding stage to perform local content adaptive filtering and frequency-guided detail restoration; at least one MFSDB is preferably placed in a deeper decoding layer to enhance multi-scale feature extraction and cross-dimensional fusion capabilities in complex structures and scale-changing scenarios. This invention does not limit the specific type of convolutional block, number of channels, scale set size, or activation function form. As long as explicit frequency priors are used in dynamic convolution kernel generation, and multi-scale dynamic feature extraction and multi-dimensional adaptive fusion mechanisms are combined, they all fall within the scope of protection of this invention.

[0118] Figure 1 The overall network structure in an embodiment of the present invention is shown. The network adopts a U-shaped encoding / decoding framework, consisting of a shallow feature extraction layer, an encoder, a decoder, and an output mapping layer. The input noisy image is first mapped to a shallow feature space through an initial convolution to obtain shallow features. Subsequently, shallow features are fed into the encoder, where multi-level contextual information is extracted through stepwise downsampling to expand the receptive field and enhance the representation of global structures. In the decoding stage, spatial resolution is restored through stepwise upsampling, and shallow texture and edge information are supplemented by skip connections from the encoder to the decoder, thus balancing global semantic representation and local structure restoration. Unlike conventional U-shaped denoising networks, this invention introduces a frequency-aware extractor, a frequency-space jointly guided dynamic kernel generation module, a frequency-space filter block, and a multi-scale frequency-space filter block in the decoding stage. This allows the network to not only utilize spatial contextual information during detail restoration but also explicitly introduce frequency priors, multi-scale dynamic features, and cross-dimensional fusion information, thereby improving the restoration capability in complex textures and noisy scenes.

[0119] Figure 2 The diagram illustrates the structure of the frequency-space jointly guided dynamic kernel generation module and the frequency-space filtering block in an embodiment of the present invention. In this diagram, the frequency-space jointly guided dynamic kernel generation module, based on the input features from the decoding stage, extracts both the spatial context representation and the channel enhancement representation of the input features. It then combines these with low-frequency structural features, mid-frequency texture features, and high-frequency detail features provided by the frequency-aware extractor to construct a unified descriptor. This unified descriptor comprehensively characterizes the spatial structural attributes, channel response attributes, and frequency distribution attributes of the input features, and further serves as the conditional input for the generation of the dynamic convolutional kernel. Based on this unified descriptor, content-adaptive dynamic convolutional kernels can be generated for different locations and different input contents, thereby allowing the convolutional filtering behavior to be simultaneously constrained by spatial semantic information, channel relationship information, and frequency prior information. Figure 2 The frequency spatial filtering block FSDB in the middle is composed of a frequency-aware extractor, a frequency spatial joint guided dynamic kernel generation module, a dynamic convolutional layer, an activation function, and residual connections. Its function is to perform content-aware frequency spatial joint filtering on the input features at a local scale: prioritizing the enhancement of noise suppression in smooth regions and prioritizing the enhancement of structure preservation in edge and texture regions, thereby suppressing noise while preserving as much real detail as possible.

[0120] Figure 3 The structure of the frequency-aware extractor in an embodiment of the present invention is shown. This module is used to perform explicit frequency decomposition on the input features and is a core component of the present invention for constructing low-frequency structural features, mid-frequency texture features, and high-frequency detail features. Specifically, the frequency-aware extractor uses a fixed two-dimensional discrete cosine transform (DCT) filter bank to convolve the input features, obtaining multiple frequency responses. These frequency responses are then divided into low-frequency, mid-frequency, and high-frequency sets according to their frequency range. Subsequently, the responses in each set are aggregated and normalized to obtain the low-frequency structural features. Mid-frequency texture features and high-frequency detail features The low-frequency structural features primarily characterize the smooth background, subject outline, and overall brightness distribution in the image; the mid-frequency texture features primarily characterize local repetitive patterns and mid-scale structural transitions; and the high-frequency detail features primarily characterize edge, corner, and fine-grained texture variations. This explicit frequency decomposition process decouples different frequency components that were originally mixed in a unified spatial domain representation, thus providing frequency constraints with clear physical meaning and interpretability for subsequent dynamic convolution kernel generation, thereby alleviating the confusion between high-frequency texture and noise residuals.

[0121] Figure 4 The diagram illustrates the structure of the frequency-space jointly guided multi-scale dynamic feature extraction module and the multi-dimensional adaptive fusion module in an embodiment of the present invention. The first half of the diagram corresponds to the multi-scale frequency-space filter block (MFSDB), whose main function is to perform dynamic convolution operations at multiple preset receptive field scales to obtain dynamic responses at different scales, and to concatenate and aggregate the responses at each scale to form a unified multi-scale feature representation. Smaller-scale dynamic responses are more conducive to preserving fine-grained texture and local edge information, while larger-scale dynamic responses are more conducive to modeling the correlation between large-scale contextual structures and repetitive patterns. Therefore, multi-scale dynamic feature extraction can simultaneously take into account both local detail restoration and overall structural consistency. Figure 4 The latter half of the diagram corresponds to the Multidimensional Adaptive Fusion Module (MDAF). This module pools and projects multi-scale features along different dimensions to construct channel-dimensional descriptors, height-dimensional descriptors, and width-dimensional descriptors, respectively. It then establishes correlations between different dimensions through a cross-dimensional attention mechanism to obtain fused features. Furthermore, it combines convolution and deformable convolution to complete feature recalibration, thereby enhancing the collaborative modeling capability across different scales and dimensions. Through this structural design, MDAF improves the model's ability to represent complex structural regions, non-uniform noise regions, and cross-scale texture patterns, enabling the method of this invention to maintain good denoising accuracy and structural restoration quality even in complex scenes.

[0122] Example

[0123] To verify the effectiveness and applicability of the method of the present invention, comparative experiments were conducted in scenarios of synthesizing Gaussian noise into color images, synthesizing Gaussian noise into grayscale images, and denoising real-world images. The experimental results are shown in Tables 1 to 4.

[0124] In the experiment of synthesizing Gaussian noise in color images, CBSD68, Kodak24, and Urban100 were selected as test datasets. CBSD68 was primarily used to verify the method's basic denoising ability on natural color images, Kodak24 was primarily used to verify the method's detail preservation ability on high-quality natural images, and Urban100 was primarily used to verify the method's recovery ability in scenes with complex textures, regular structures, and repetitive patterns. In the experiments, Gaussian noise of different intensities was artificially superimposed on the test images to construct degradation conditions from weak to strong noise. The CBSD68 and Kodak24 datasets were used... The noise settings for the Urban100 dataset are as follows: The noise settings were adjusted. This set of experiments was mainly used to verify the ability of the method of the present invention to recover edge, texture and structural information in color images.

[0125] Table 1 Comparative experiments on the CBSD68 dataset

[0126]

[0127] Table 2 Comparison experiments on Kodak24 and Urban100 datasets.

[0128]

[0129] In the experiment of synthesizing Gaussian noise into grayscale images, Set12, BSD68, and Urban100 were selected as test datasets. Set12 was used to verify the basic recovery ability of the method on typical grayscale images, BSD68 was used to verify the generalization ability of the method on natural grayscale images, and Urban100 grayscale images were used to verify the adaptability of the method in scenes with complex structures and high-frequency details. In the experiment, Gaussian noise of different intensities was superimposed on the test images, and the noise standard deviation was set to [value missing]. This set of experiments is mainly used to verify the stability of the method of the present invention in grayscale image denoising tasks, as well as its ability to preserve image details under different texture complexity conditions.

[0130] Table 3 Comparison experiments on Set12, BSD68, and Urban100 datasets.

[0131]

[0132] In real-world image denoising experiments, SIDD and DND were selected as test datasets. The SIDD dataset primarily reflects the real noise characteristics under mobile terminal shooting conditions and can be used to evaluate the method's recovery ability under the combined effects of sensor noise, exposure perturbation, and imaging chain errors. The DND dataset is used to evaluate the method's generalization ability in various real-world shooting environments. Unlike the synthetic Gaussian noise experiments, this set of experiments does not artificially add noise but directly uses real noisy images provided in publicly available datasets for testing. In terms of evaluation metrics, the real-world image denoising experiments simultaneously use PSNR and SSIM to comprehensively evaluate the recovery results, reflecting the method's performance in both noise suppression and structure preservation.

[0133] Table 4 Comparison experiments on SIDD and DND datasets.

[0134]

[0135] In the comparative experiments, representative traditional denoising methods, convolutional neural network denoising methods, and high-performance image denoising methods in recent years were selected as references, including BM3D, DnCNN, FFDNet, RDN, RIDNet, DeamNet, P3AN, MSANet, ADFNet, MSCNet, MPRNet, and Uformer. By conducting unified comparisons under different datasets, different noise intensities, and different imaging scenarios, the effectiveness of the proposed method can be comprehensively verified.

[0136] Experimental results show that the proposed method demonstrates strong competitiveness in synthesizing Gaussian noise in color images across multiple datasets and noise intensity settings, particularly in images with complex textures and regular structures, where it effectively balances noise suppression and detail preservation. In synthesizing Gaussian noise in grayscale images, the proposed method exhibits stable recovery capabilities on datasets such as Set12, BSD68, and Urban100, indicating that the proposed method is applicable not only to color images but also to grayscale image denoising tasks. In real-world image denoising experiments, the proposed method also achieves good recovery results on the SIDD and DND datasets, demonstrating its good generalization ability and practical application value for real-world complex noise scenes.

[0137] The experiments described above demonstrate that the method of this invention, through explicit frequency decomposition, frequency-space jointly guided dynamic kernel generation, multi-scale dynamic feature extraction, and multi-dimensional adaptive fusion mechanism, achieves stable and effective performance in various denoising tasks involving color images, grayscale images, and real-world images. Especially in complex texture regions, repetitive structure regions, and under conditions of strong noise, the method of this invention better balances noise suppression, edge preservation, and detail recovery, thus verifying the effectiveness and applicability of the proposed solution.

[0138] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An image denoising method based on a frequency-space jointly guided dynamic kernel generation network, characterized in that, The method includes the following steps: Step S1: Obtain the noisy input image and extract shallow features through the initial convolutional layer; Step S2: Input the shallow features into a U-shaped encoder-decoder network, which includes an encoder and a decoder. The encoder extracts multi-level context features through downsampling, and the decoder recovers the spatial resolution step by step through upsampling and skip connections to obtain the decoded features. Step S3: Use a fixed two-dimensional discrete cosine transform filter bank to perform frequency decomposition on the decoded features to obtain multiple frequency responses. Divide the frequency responses into low-frequency set, mid-frequency set and high-frequency set according to frequency range, and aggregate them to obtain low-frequency structural features, mid-frequency texture features and high-frequency detail features respectively. Step S4: Extract the spatial context representation and channel enhancement representation of the decoded features, fuse the spatial context representation, channel enhancement representation, low-frequency structural features, mid-frequency texture features and high-frequency detail features into a unified descriptor, and generate multiple dynamic convolution kernels with preset receptive field scales based on the unified descriptor; Step S5: Apply the dynamic convolution kernels at each preset receptive field scale to the input features to obtain the dynamic response at the corresponding scale. Then, convolve and aggregate the dynamic responses at each scale to obtain a multi-scale feature representation. Step S6: For the multi-scale feature representation, construct descriptors along the channel dimension, height dimension, and width dimension respectively, obtain fused features based on cross-dimensional attention, and then combine convolution and deformable convolution to obtain enhanced features; Step S7: Convolve the enhanced features through the output mapping to obtain the denoised image.

2. The image denoising method based on the frequency space joint guided dynamic kernel generation network according to claim 1, characterized in that, In step S2, the encoder extracts multi-level context representations step by step using stride convolution and convolutional blocks; the decoder restores spatial resolution step by step through upsampling and supplements shallow detail information through jump connections from the encoder to the decoder, thereby enhancing the global semantic expression capability while maintaining the restoration capability of local texture and edge structure.

3. The image denoising method based on the frequency space joint guided dynamic kernel generation network according to claim 1, characterized in that, The specific method for using a fixed two-dimensional discrete cosine transform filter bank to perform frequency decomposition on the decoded features in step S3, obtaining multiple frequency responses, and dividing the frequency responses into low-frequency set, mid-frequency set, and high-frequency set according to frequency range, and aggregating them to obtain low-frequency structural features, mid-frequency texture features, and high-frequency detail features respectively, includes: Using predefined two-dimensional discrete cosine transform (DCT) filter banks convolving the input features to obtain a frequency response: A single filter kernel in the two-dimensional DCT filter bank can be represented as: wherein, and denote the horizontal and vertical frequency indices, respectively, and denote the spatial coordinates in the filter kernel, denotes the filter kernel side length, denotes the normalization coefficient; After obtaining the frequency response, it is divided into a low-frequency set according to the frequency range. Mid-frequency collection and high frequency set The low-frequency structural features were obtained by aggregation. Mid-frequency texture features and high-frequency detail features The above polymerization process can be represented as: in, Indicates the first The response map obtained after applying DCT filter kernels to the input features , and These represent the sets of low-frequency filter kernels, mid-frequency filter kernels, and high-frequency filter kernels, respectively. , and These represent the number of elements in the corresponding sets. Represents the normalization operator; The normalization operator is used to reduce the differences in numerical scale between features of different frequency bands and improve the stability of subsequent fusion and dynamic kernel generation processes.

4. The image denoising method based on a frequency-space jointly guided dynamic kernel generation network according to claim 1, characterized in that, The specific method for extracting the spatial context representation and channel enhancement representation of the decoded features in step S4, fusing the spatial context representation, channel enhancement representation, low-frequency structural features, mid-frequency texture features, and high-frequency detail features into a unified descriptor, and generating multiple dynamic convolutional kernels with preset receptive field scales based on the unified descriptor includes: Input features for the decoding stage Joint representation modeling is performed, extracting spatial context information and channel interaction information respectively, and combining low-frequency structural features. Mid-frequency texture features and high-frequency detail features Construct a unified descriptor for dynamic convolution kernel generation. The unified descriptor satisfies: in, Representation of spatial context, This indicates channel enhancement. The unified descriptor represents the fusion mapping function. Spatial structural attributes, channel response attributes, and frequency distribution attributes used to comprehensively characterize input features; based on the unified description sub , for different scales in a preset scale set Adaptive dynamic convolution kernels are generated, and dynamic convolution is respectively performed on input features to obtain multi-scale dynamic responses, wherein the dynamic convolution kernel under the corresponding scale satisfies:​ wherein, denotes the dynamic kernel generating function.

5. The image denoising method based on a frequency-space jointly guided dynamic kernel generation network according to claim 1, characterized in that, The decoding stage sets up a Frequency Spatial Filtering Block (FSDB), which includes a frequency-aware extractor, a frequency spatial joint guided dynamic kernel generation module, a dynamic convolutional layer, an activation function, and residual connections. Its output satisfies the following: Where DynConv(·) represents dynamic convolution operation based on normalized dynamic convolution kernel Ŵ, and δ(·) represents activation function.

6. The image denoising method based on a frequency-space jointly guided dynamic kernel generation network according to claim 1, characterized in that, The specific method described in step S5, which involves applying dynamic convolutional kernels corresponding to each scale to the input features to obtain dynamic responses at the corresponding scales, and then concatenating and aggregating the dynamic responses at each scale to obtain multi-scale feature representations, includes: applying a dynamic convolution kernel at a corresponding scale to the input features to obtain multi-scale dynamic responses: The dynamic responses at different scales are then concatenated to obtain joint features: A unified multi-scale representation is obtained by 1×1 convolution aggregation: wherein, denotes a scale under a dynamic convolution kernel denotes the aggregated multi-scale features; through multi-scale dynamic convolution response modeling, the joint enhancement of image structure and texture information under different receptive fields is realized.

7. The image denoising method based on the frequency space joint guided dynamic kernel generation network according to claim 6, characterized in that, The preset scale set S in step S5 is a multi-scale set containing different expansion rates or receptive fields, preferably S = {1, 3, 5}.

8. The image denoising method based on the frequency space joint guided dynamic kernel generation network according to claim 1, characterized in that, Step S6 involves constructing descriptors along the channel, height, and width dimensions for the multi-scale feature representation, obtaining fused features based on cross-dimensional attention, and then combining convolution and deformable convolution to obtain enhanced features. Specific methods for this include: The multi-dimensional adaptive fusion is realized by a multi-dimensional adaptive fusion module MDAF, which is realized by performing multi-scale feature The pooling and projection along different dimensions construct Q, K, and V, and satisfy: The attention weight matrix is ​​obtained through cross-dimensional attention calculation: The fused features are obtained, and the output features are recalibrated by combining convolution and deformable convolution to obtain enhanced features: in, , and These are the linear projection parameters corresponding to the query branch, key branch, and value branch, respectively; For query representation AND key representation Feature dimension, Attention scaling term, used for matrix multiplication Scale normalization is performed to enhance the numerical stability of the attention computation process; For cross-dimensional attention weight matrix, This represents the fused features. This indicates a deformable convolution operation.

9. The image denoising method based on a frequency-space jointly guided dynamic kernel generation network according to claim 1, characterized in that, The specific method for obtaining the denoised image by convolving the enhanced features with the output mapping in step S7 includes: in, For denoising the image; This indicates the output mapping convolution operation.

10. The image denoising method based on a frequency-space jointly guided dynamic kernel generation network according to claim 1, characterized in that, The decoding stage of the U-shaped codec network includes multiple frequency space filter blocks (FSDB) and at least one multi-scale frequency space filter block (MFSDB). FSDB is used for local content adaptive reconstruction, while MFSDB is used for multi-scale feature enhancement and cross-dimensional fusion in scenarios with complex structures and scale variations.

Images