Multi-degraded image restoration method based on degradation feature decoupling and guiding

Abstract

The invention relates to a multi-degraded image restoration method based on degradation feature decoupling and guiding. Comprising the following steps: preprocessing an input image, including image pairing, cutting and data enhancement processing; designing a multi-degraded image restoration network; designing a learning framework and a loss function for training the designed network, and training the network by using the image training data set; and inputting a to-be-detected image into the obtained model to generate a restored image. According to the invention, reversible feature transformation and a degradation feature progressive decoupling mechanism realized by each degradation processing module are utilized, and a scale-by-scale degradation feature progressive guiding module based on degradation perception is introduced into a multi-scale feature space, so that the multi-scale feature space can be obtained on the premise of not changing a backbone network structure; different degradation factors such as noise, blurring and low illumination are effectively distinguished and removed in a targeted mode, so that image content details are stably recovered under the complex multi-degradation condition, the overall visual quality is improved, and good practical application value and model migration capacity are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing and computer vision, and particularly to a multi-degradation image restoration method based on decoupling and guiding of degraded features. BACKGROUND

[0002] With the popularity of imaging devices and the continuous improvement of image quality requirements in the fields of security monitoring, autonomous driving, mobile shooting, intelligent medical treatment, etc., multi-degradation image restoration technology has gradually become an important research direction of image processing. In the real world, images are often affected by multiple factors such as noise, blur, compression artifacts, and low-light degradation. Due to the limitations of sensor performance, lighting conditions, exposure strategies, and compression encoding in the image formation process, these degradations often exhibit mixed superposition, uneven spatial distribution, and significant intensity differences, resulting in image brightness abnormalities, texture blurring, detail loss, and noise enhancement, which seriously affect the stability and accuracy of downstream detection, recognition, and reconstruction tasks.

[0003] Unlike traditional single-degradation restoration tasks, the core challenge of multi-degradation image restoration is not only to restore the spatial details of the image, but also to simultaneously distinguish different degradation factors and avoid interference and coupling between different degradations. Existing methods mainly include traditional prior model-based restoration algorithms and deep learning-based end-to-end restoration methods: the former usually relies on hand-designed degradation models or regularization terms such as sparse representation, total variation, and non-local self-similarity, but it is difficult to accurately model complex multi-degradation situations in real-world scenarios and has high computational overhead; the latter has strong restoration capabilities using deep networks, but most methods mix all degradations in a unified feature space for processing, which can easily be dominated by strong noise or severe blur, causing content and degradation features to be coupled, resulting in problems such as over-denoising, detail loss, or residual degradation.

[0004] The multi-degradation restoration task also has the following difficulties: first, there are obvious distribution differences and mutual interference between different degradation types, for example, noise is randomly distributed in fine-grained, while compression artifacts usually have a blocky structure. This difference makes it difficult for a single network to simultaneously consider all types of degradation factors; second, the combination of degradation factors has high randomness and lacks a clear prior order, making the restoration model prone to failure when facing unseen degradation combinations; third, existing deep models usually lack an interpretable degradation separation mechanism, making it difficult to explicitly model degradation information, thereby limiting the model's generalization ability in complex degradation scenarios.

[0005] Although some attempts have been made to handle mixed degradation by introducing multi-branch structure, degenerated label embedding or blind denoising framework, these methods either rely on artificial definition of degradation type, are difficult to adapt to real scenarios, or cannot achieve true decoupling of degradation and content at the feature level, still making it difficult to fundamentally solve the problem of multi-degradation coupling. In addition, existing research generally lacks modeling of the causal relationship and action path between degradation factors, making it difficult for the network to distinguish between "content changes" and "degradation changes", ultimately affecting the stability and consistency of the restoration. Therefore, how to achieve explicit separation, causal order modeling and scale-by-scale guided recovery of multiple degradation factors without relying on degradation labels, and maintain the robustness and generalization ability of the model under complex degradation combination, has become a key problem that needs to be solved in current multi-degradation image restoration technology. SUMMARY

[0006] Therefore, the purpose of the present application is to provide a multi-degradation image restoration method based on degradation feature decoupling and guidance, which gradually decouples the degradation features by designing a degradation feature gradual decoupling module (step B2), and uses reversible group normalization units, reversible feature projection units, degradation feature standardization modules, and spatially adaptive degradation feature scaling modules to gradually disassemble the mixed degradation into multiple degradation subspaces with clear semantics, thereby effectively solving the problem of mixed degradation types, content and degradation features interfering with each other in existing methods. On this basis, the present application further constructs a degradation feature gradual guidance module (step B3) for each scale, which realizes the degradation perception enhancement of different scale content features through degradation feature upsampling, spatial modulation and channel offset, so that the network can realize targeted restoration according to the degradation type and spatial distribution, and avoid over-enhancement or degradation residue. The above two modules realize lightweight adaptation without changing the architecture of the main network, and combine multiple constraints such as reconstruction loss, deep semantic guidance loss and degradation feature decoupling loss (step C) to improve the restoration effect from multiple dimensions such as content consistency, semantic stability and degradation orthogonality. Through the synergistic effect of gradual degradation separation and scale degradation guidance, the present application can stably recover image details and structures under complex multi-degradation conditions, improve the stability, interpretability and generalization ability of the restoration result, and provides an efficient, robust and engineering feasible solution for actual multi-degradation image restoration.

[0007] To achieve the above purpose, the present application adopts the following technical solution: a multi-degradation image restoration method based on degradation feature decoupling and guidance, comprising the following steps:

[0008] Step A, pre-processing the input image, including image pairing, cropping and data enhancement processing to obtain an image training data set;

[0009] Step B, design a multi-degradation image restoration network, including a multi-scale image encoder, a degradation feature progressive decoupling module, a degradation feature progressive guiding module and a multi-scale image decoder;

[0010] Step C, design a learning framework and loss function for training the multi-degradation image restoration network designed in step B, and train the multi-degradation image restoration network in step B using the image training data set described in step A to obtain a trained multi-degradation image restoration model;

[0011] Step D, input the image to be tested into the multi-degradation image restoration model obtained in step C to generate a restored image.

[0012] In a preferred embodiment, step A is implemented as follows:

[0013] Step A1, pair the non-degradation normal image with the multi-degradation image, wherein the non-degradation normal image is used as a label image;

[0014] Step A2, randomly crop each multi-degradation image with a size of to an image with a size of , and use the same random cropping method for its corresponding non-degradation normal image to ensure that they have the same size and position, wherein and represent the height and width of the multi-degradation image and the non-degradation normal image, respectively, and represent the height and width of the cropped image, respectively;

[0015] Step A3, for each training paired image, randomly apply one of the following 8 data augmentation methods: keep the original image, vertical flip, rotate 90 degrees, rotate 90 degrees and then vertical flip, rotate 180 degrees, rotate 180 degrees and then vertical flip, rotate 270 degrees, rotate 270 degrees and then vertical flip.

[0016] In a preferred embodiment, step B is implemented as follows:

[0017] Step B1, design a multi-scale image encoder, use the encoder of the URWKV backbone network as the multi-scale image encoder, and perform feature extraction on the input multi-degradation image to obtain multi-scale intermediate features , then perform bilinear two times down-sampling on the scale intermediate features to obtain backbone features ; wherein , , and 3 are the height, width and RGB channel number of the multi-degradation image, respectively, This represents the number of intermediate features across multiple scales, which varies depending on the scale of the multi-scale image encoder. The Middle Scale intermediate features are , , , and These are the first intermediate features in the multi-scale model. The height, width, and number of channels of the intermediate features at scale. , and These are the height, width, and number of channels of the backbone feature, respectively;

[0018] Step B2: Design a progressive decoupling module for degradation features, by The degenerate feature decoupling modules are cascaded together. Each module includes a reversible group normalization unit, a reversible feature projection unit, a degenerate feature standardization module, a spatially adaptive degenerate feature scaling module, and a degenerate feature purification and separation operation. The reversible group normalization unit includes an 8-group normalization layer, containing forward and reverse group normalization operations. The reversible feature projection unit includes a convolutional kernel with a size of [missing information - likely a value]. A non-biased convolutional layer with a stride of 1, padding of 0, and a group number of 1, including forward and backward convolution operations; the progressive decoupling module for degradation features is used to extract degradation features sequentially and stepwise, decomposing the mixed degradation into multiple degradation subspaces with clear semantics, while removing degradation from the backbone features; specifically: for the backbone features extracted in step B1 After the degradation feature progressive decoupling module, the following is obtained: Degenerative features of the same size and degenerative skeletal features ;in Indicates the number of degradation features. The degradation characteristics are , , The number of channels representing the degradation feature; the process is expressed by the formula:

[0019]

[0020] in, This indicates a decoupling module for degradation features. This indicates the stacking process of decoupling modules for degradation features;

[0021] Step B3: Design progressive guidance modules for degradation features at each scale. Each progressive guidance module for degradation features at each scale consists of... The degradation feature guidance modules are cascaded sequentially. Each degradation feature guidance module includes a degradation feature upsampling module, a reversible group normalization unit, a reversible feature projection unit, a spatially adaptive degradation feature modulation module, and a degradation feature guidance separation operation. The reversible group normalization unit includes a group normalization layer with 8 groups, containing forward group normalization and reverse group normalization operations. The reversible feature projection unit includes a convolutional kernel with a size of [missing information]. A non-biased convolutional layer with a stride of 1, padding of 0, and 1 group, containing forward and backward convolution operations, wherein the backward convolution operation is implemented using the inverse operator obtained from LU decomposition; the degradation feature progressive guidance module removes degradation from the corresponding scale features based on the degradation features; for the 1st The scale, the first one extracted in step B1 Scale intermediate features The degradation features extracted in step B2 are used as direct input. As an auxiliary input, after the first After processing by the progressively guided module for the degradation features of the scale, the first... Scale de-degradation intermediate features The process can be expressed by the following formula:

[0022]

[0023] in, Indicates the first Gradual guidance module for scale degradation features. Indicates the first The degradation characteristics of scale are progressively guided by the stacking of modules;

[0024] Completed After each scale, collect the dedegradable intermediate features at each scale. Multi-scale de-degradation intermediate features are obtained. ;

[0025] Step B4: Design a multi-scale image decoder. Use the decoder of the URWKV backbone network as the multi-scale image decoder, and process the de-degenerate backbone features obtained in step B2. and the multi-scale de-degradation intermediate features obtained in step B3 As input to the multi-scale image decoder; specifically, the de-degradation backbone features are used. and the Scale de-degradation intermediate features As the first The input to the scale image decoder, then the... The output of the scale image decoder and As the first The input to the scale image decoder is thus gradually processed. After decoding the image at each scale, the output is the restored image. .

[0026] In a preferred embodiment, step B2 is specifically implemented as follows:

[0027] Step B21: Extract the backbone features obtained in step B1. The first degenerate feature decoupling module uses its input features as input features, and then the output features of the first degenerate feature decoupling module use its input features as input features for the second degenerate feature decoupling module. This is how the degenerate feature decoupling modules are implemented. This stacking process yields the output of the progressive decoupling module for degenerate features, i.e., the degenerate backbone features. ,at the same time The decoupling module for each degradation feature outputs during intermediate execution. Degenerative features of the same size For the first The decoupling module for the degradation feature will be derived from the previous one, i.e., the first one. The output features of each degradation feature decoupling module are denoted as follows: If this is the first degenerate feature decoupling module, then Recorded as , First, a forward group normalization operation is performed through a reversible group normalization unit, and then a forward convolution operation is performed through a reversible feature projection unit to obtain the backbone projection features. Then, the skeleton projection features Based on the channel dimension, it is split into two features to obtain the core content features. and skeletal degeneration characteristics ; characteristics of skeletal degeneration Input degradation feature standardization module to obtain standardized backbone degradation features Standardized backbone degeneration characteristics As direct input, core content features As auxiliary input, after passing through the spatial adaptive degradation feature scaling module, the features to be de-degraded are obtained. ; characteristics of skeletal degeneration As direct input, features to be degraded As auxiliary input, after degradation feature purification and separation operations, degradation features are obtained. and de-degeneration characteristics De-degenerate features and core content characteristics The features are merged into a single feature based on the channel dimension. The merged feature is then input into a reversible feature projection unit to perform a backward convolution operation, and then input into a reversible group normalization unit to perform a backward group normalization operation, yielding the output feature. When stacking to the last degenerate feature decoupling module, the output feature obtained is the output feature of the degenerate feature progressive decoupling module, i.e., the de-degenerate backbone feature. ,Then The degradation feature decoupling module generates during the intermediate execution process Each degradation feature consists of ;

[0028] Step B22: Design a degradation feature standardization module, which includes, in sequence, group normalization units, Depthwise convolutional units and group normalization units; wherein the group normalization unit includes a number of groups. The group normalization layer, Depthwise convolutional units include convolutional kernels of size 1. Step size is 1, fill size is 1, number of groups is The unbiased convolutional layer; the degradation feature standardization module is used to perform grouping and standardization operations on the input degradation features, converting complex degradation features into standard degradation features; specifically, it performs standardization operations on the backbone degradation features. sequentially through group normalization units, After depthwise convolutional units and group normalized units, standardized backbone degradation features are obtained. ;in , and These represent the height, width, and number of channels of the standardized backbone degradation feature; the process is expressed by the formula:

[0029]

[0030] in, Represents the group normalized unit, express Depthwise convolutional units;

[0031] Step B23: Design a spatial adaptive degradation feature scaling module, which includes, in sequence: Adaptive pooling layer with bias Pointwise convolutional units and sigmoid operations; with bias Pointwise convolutional units include convolutional kernels of size 1. Step size is 1, padding is 0, number of groups is The biased convolutional layer; the spatial adaptive degradation feature scaling module is used to adaptively adjust the intensity of degradation features according to the spatial distribution of content features; specifically, it scales the backbone content features... Passing through in sequence Adaptive pooling layer with bias After the point-wise convolution unit and the Sigmoid operation unit, the channel attention weight is obtained, which is multiplied with the normalized backbone degradation feature to obtain the to-be-degradation-removed feature ; the process is expressed by a formula as follows:

[0032]

[0033] wherein, represents an adaptive pooling layer unit, represents a bias a point-wise convolution unit, represents a Sigmoid operation, represents a channel-wise multiplication operation.

[0034] Step B24, a degradation feature purification separation operation is designed, which includes a GELU activation function and a subtraction operation, and is used for separating the degradation feature and the to-be-degradation-removed feature; specifically, the to-be-degradation-removed feature is subjected to the GELU activation function to obtain the degradation-removed feature , and the backbone degradation feature is subtracted from the degradation-removed feature to obtain the degradation feature ; the process is expressed by a formula as follows:

[0035]

[0036]

[0037] wherein, represents a GELU activation function operation.

[0038] In a preferred embodiment, the step B3 is specifically implemented as follows:

[0039] Step B31, the first scale intermediate feature extracted in the step B1 is taken as the input feature of the first scale first degradation feature guide module, and then the output feature of the first scale first degradation feature guide module is taken as the input feature of the first scale second degradation feature guide module, so that the scale degradation feature guide module is stacked times to obtain the output of the first scale degradation feature progressive guide module, i.e., the first scale degradation-removed backbone feature ; for the second scale degradation feature guide module, the steps B31 and B32 are repeated a degradation feature guiding module, the degradation feature extracted in step B2 is inputted into the degradation feature guiding module as a second input feature, after the degradation feature upsampling module, the degradation feature of the first scale is obtained the intermediate feature of the second scale the degradation feature of the second scale the degradation feature of the second scale ; at the same time, the output feature from the previous, i.e., the first degradation feature guiding module, is denoted as if it is the first degradation feature guiding module at this time, it is denoted as , the first scale projection feature is obtained by inputting the reversible group normalization unit first to perform the forward group normalization operation, and then inputting the reversible feature projection unit to perform the forward convolution operation the first scale projection feature then the first scale projection feature is split into two features according to the channel dimension to obtain the first scale content feature the first scale content feature and the first scale degradation feature the first scale content feature is inputted directly, and the first scale degradation guiding feature is inputted as auxiliary input, after the spatial adaptive degradation feature modulation module, the first scale content enhancement feature is obtained the first scale content enhancement feature the first scale degradation feature is inputted directly, and the first scale degradation guiding feature is inputted as auxiliary input, after the degradation feature guiding separation operation, the first scale de-degradation feature is obtained the first scale de-degradation feature the first scale de-degradation feature and the first scale content recombination feature the first scale de-degradation feature is inputted directly, and the first scale degradation guiding feature is inputted as auxiliary input, after the degradation feature guiding separation operation, the first scale de-degradation feature is obtained the first scale de-degradation feature the first scale de-degradation feature and the first scale content recombination feature the first scale de-degradation feature is inputted directly, and the first scale degradation guiding feature is inputted as auxiliary input, after the degradation feature guiding separation operation, the first scale de-degradation feature is obtained the first scale de-degradation feature the first scale de-degradation feature the first scale de-degradation feature the first scale de-degradation feature the first scale de-degradation feature​​ ;

[0040] Step B32, design a degradation feature up-sampling module, including a group normalization unit, a point-wise convolution unit and a spatial channel shuffle unit; wherein the group normalization unit includes a group normalization layer with a group number of ; the point-wise convolution unit includes a bias-free convolution layer with a convolution kernel size of , a stride of 1 and a padding of 0, and a group number of ; the spatial channel shuffle unit includes a pixel shuffle layer with an up-sampling factor of ; the degradation feature up-sampling module is used for spatial reconstruction of low-resolution degradation features; specifically, the degradation feature is sequentially input into the group normalization unit, the point-wise convolution unit and the spatial channel shuffle unit to obtain the degradation guide feature of the first scale ; the process is expressed by the formula:

[0041]

[0042] wherein, denotes the group normalization unit, denotes the point-wise convolution unit, denotes the spatial channel shuffle unit;

[0043] Step B33, design a spatial adaptive degradation feature modulation module, including a degradation feature guided adaptive spatial scaling module and a degradation feature guided adaptive channel offset module; the spatial adaptive degradation feature modulation module is used for degradation perception from two scales of local and global and for enhancement of content features; specifically, the degradation guide feature of the first scale is input into the degradation feature guided adaptive spatial scaling module to obtain the content-wise enhancement weight of the first scale ; the degradation guide feature of the first scale is directly input, and the content feature of the first scale is auxiliary input, and the degradation feature guided adaptive channel offset module is used to obtain the channel-wise enhancement weight of the first scale ; the content feature of the first scale is multiplied by the content-wise enhancement weight of the first scale element by element, and then multiplied by the channel-wise enhancement weight of the Scale per-channel enhanced weight element-wise addition operation to obtain the scale content enhanced feature ; the process is represented by the formula:

[0044]

[0045]

[0046]

[0047] wherein, represents a degraded feature guided adaptive spatial scaling module, represents a degraded feature guided adaptive channel shift module, represents an element-wise multiplication operation, represents an element-wise addition operation;

[0048] Step B34, designing a degraded feature guided separation operation, including a subtraction operation, for separating the scale degraded guide feature from the scale degraded feature; specifically, subtracting the scale degraded guide feature from the scale degraded feature to obtain the scale de-degraded feature ; the process is represented by the formula:

[0049]

[0050] In a preferred embodiment, the step B33 is specifically implemented as follows:

[0051] Step B331, designing a degraded feature guided adaptive spatial scaling module, including a group convolution unit, a group normalization unit, a GELU activation function, a bias point-wise convolution unit, a Sigmoid operation and an element-wise addition operation; wherein the group convolution unit includes a bias-free convolution layer with a convolution kernel size of , a stride of 1, a padding of 1, and a group number of , the group normalization unit includes a group normalization layer with a group number of , and the bias point-wise convolution unit includes a bias convolution layer with a convolution kernel size of , a stride of 1, a padding of 0, and a group number of ; the degraded feature guided adaptive spatial scaling module utilizes the Scale degradation guides the spatial distribution of features, generating pixel-wise spatial weights for the first pixel. Local enhancement or suppression of scale content features; specifically: for the first obtained in step B32... Scale degradation guided features , in turn Group convolutional units, group normalized units, GELU activation function, biased activation function After pointwise convolution units and a sigmoid operation, a unit bias is added to obtain the first... Scale-based content-based weighting The process can be expressed by the following formula:

[0052]

[0053] in, express Group convolutional units, Represents the group normalized unit, This represents the GELU activation function. Indicates bias Pointwise convolutional unit, This represents the Sigmoid operation. This indicates an element-wise multiplication operation. This represents an element-wise addition operation;

[0054] Step B332: Design a degradation feature-guided adaptive channel offset module, including... Adaptive pooling layer with bias Pointwise convolutional units, GELU activation function, biased Pointwise convolutional units with bias Pointwise convolutional units, Tanh activation functions, and channel-wise multiplication operations; the degradation feature-guided adaptive channel offset module is used to perform global conditional control on content features and adjust the overall de-degradation intensity of degradation features; specifically, for the first obtained in step B32... Scale degradation guided features , in turn Adaptive pooling layer with bias Pointwise convolutional units, GELU activation function, biased Pointwise convolutional units with bias After pointwise convolutional units and Tanh activation functions, it is then combined with the first... Scale content features Perform a channel-by-channel multiplication operation to obtain the first... Scale-based channel-wise weight enhancement The process can be expressed by the following formula:

[0055]

[0056] wherein, represents an adaptive pooling layer, with bias a point-wise convolution unit, represents a GELU activation function, represents a Tanh activation function, represents a channel-wise multiplication operation.

[0057] In a preferred embodiment, the step C is specifically implemented as follows:

[0058] Step C1, constructing a total loss containing a reconstruction loss , a deep semantic guidance loss , and a degraded feature decoupling loss ; the expression is:

[0059]

[0060] wherein, , and are balance parameters;

[0061] Step C2, randomly dividing the image training data set obtained in step A into several batches, each batch containing several pairs of images;

[0062] Step C3, calculating the gradient of the parameters in the network according to the loss using the back propagation method, and updating the network parameters using the Adam optimization method;

[0063] Step C4, repeating steps C2 to C3 until the model converges, and finally obtaining a trained multi-degraded image restoration model.

[0064] In a preferred embodiment, the step C1 is specifically implemented as follows:

[0065] Step C11, designing a reconstruction loss including an L1 loss , a structure loss and a VGG perception loss ; the expression is:

[0066]

[0067]

[0068]

[0069] ​

[0070] wherein, , and are balance parameters; is a restored image, is a label image, represents an absolute error value calculation operation; and represent the mean value obtained by local statistics using a Gaussian window of ; and represent the variance obtained by local statistics using a Gaussian window of ; and are two constant values to prevent the denominator from being 0; represents a mean square error value calculation operation, represents an extraction operation of the 3rd layer, the 8th layer and the 15th layer features using a VGG-16 classification model pre-trained on an ImageNet dataset;

[0071] Step C12, designing a deep semantic guidance loss , including an intra-feature similarity loss and an inter-feature similarity loss ; the deep semantic guidance loss is used to constrain the deep features of the model and enhance the semantic representation ability of the model; specifically: using a visual base model represented by DINOv3 to extract semantic features of the label image , then inputting the backbone features extracted in step B1 into a point-by-point convolution unit to obtain backbone projected semantic features , then calculating the intra-feature similarity loss and the inter-feature similarity loss of the semantic features and the backbone projected semantic features , and performing weighted summation on the intra-feature similarity loss and the inter-feature similarity loss to obtain the deep semantic guidance loss ; wherein, the point-by-point convolution unit includes an unbiased convolution layer with a convolution kernel size of , a step of 1, padding of 0, and a group number of ; the process is represented by a formula as follows:

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078]

[0079] wherein, and denote balance parameters, denotes a visual base model, denotes a point-wise convolution unit, denotes a bilinear size alignment operation that aligns a feature size to a feature size, denotes a feature normalization operation, denotes a cosine similarity calculation operation, denotes an absolute error value calculation operation.

[0080] Step C13, designing a degradation feature decoupling loss including a non-degradation guiding loss , a multi-degradation contrast loss and a multi-degradation orthogonal loss ; the degradation feature decoupling loss is used to guide the degradation features of normal images to tend to be non-degradation, and guide the degradation features of degradation images to tend to be distinguishable and decoupled; the expression thereof is:

[0081]

[0082] wherein, , and are balance parameters.

[0083] In a preferred embodiment, the step C13 is specifically implemented as follows:

[0084] Step C131, designing a non-degradation guiding loss , the non-degradation guiding loss is used to guide the degradation features of normal images to tend to be non-degradation; specifically, for a label image , input it into the multi-scale image encoder designed in step B1 to obtain the extracted multi-scale non-degradation intermediate features and backbone features , and denoted as the non-degraded image backbone feature , inputting the designed degradation feature progressive decoupling module of step B2, obtaining a plurality of degradation features with the same size , denoted as the non-degraded image degradation feature , calculating the absolute error value of the non-degraded image degradation feature and the all-zero feature, obtaining the non-degradation guiding loss ; wherein , and are the height, width and channel number of the non-degraded image backbone feature, denotes the number of degradation features, and the th non-degraded image degradation feature is , , is the channel number of the degradation feature; the process is represented by the formula:

[0085]

[0086] wherein denotes the absolute error value calculation operation;

[0087] Step C132, designing a multi-degradation contrast loss and a multi-degradation orthogonal loss , the multi-degradation contrast loss is used to constrain the degradation features of different degradation types to tend to be distinguishable in the feature space, and the multi-degradation orthogonal loss is used to constrain the degradation features of different degradation types to maintain pairwise orthogonality in the feature space, so as to promote the mutual decoupling of each degradation factor; specifically: for the a plurality of degradation features with the same size obtained in step B2, normalizing each degradation feature , then calculating the matrix product between the stacked degradation features to obtain a degradation feature similarity matrix, and then calculating the temperature information noise contrast estimation loss of the degradation feature similarity matrix to obtain the multi-degradation contrast loss , calculating the absolute error value of the degradation feature similarity matrix and the unit matrix to obtain the multi-degradation orthogonal loss ; the process is represented by the formula:

[0088]

[0089]

[0090]

[0091]

[0092] wherein, represents a feature normalization operation, represents a stacking operation of multiple features, represents a temperature coefficient, taking a value of 0.07, represents an information noise contrast estimation loss calculation function, represents an identity matrix, represents an absolute error value calculation operation.

[0093] In a preferred embodiment, the step D is specifically implemented as follows:

[0094] Step D1, performing a boundary padding operation on the multiple-degradation to-be-tested image with a size of to pad its height and width to and respectively using all-zero pixels, to obtain an image with a size of , wherein , are the height and width of the multiple-degradation to-be-tested image respectively, , are the height and width of the padded multiple-degradation to-be-tested image respectively;

[0095] Step D2, inputting the padded multiple-degradation to-be-tested image to the trained multiple-degradation image restoration model in step C to obtain a restored image with a size of , and recording as an inference restored image ;

[0096] Step D3, performing a de-boundary padding operation on the inference restored image with a size of to remove the boundary padding area, to obtain a final inference restored image with the same size as before padding.

[0097] Compared with existing technologies, the present invention has the following advantages: First, the degradation feature progressive decoupling module (step B2) proposed in this invention, through structures such as reversible group normalization units, reversible feature projection units, and degradation feature standardization modules, sequentially separates mixed degradation into multiple semantically clear degradation subspaces, effectively solving the problem of restoration instability caused by degradation type coupling and content feature interference in traditional restoration methods; this module can achieve explicit differentiation between degradation and content while ensuring information integrity, making the degradation modeling process more interpretable and controllable. Second, the degradation feature progressive guidance module (step B3) designed in this invention can perform degradation-aware spatial modulation and channel offset on content features of different resolutions according to the spatial scale differences of image features, thereby achieving targeted local de-degradation; through degradation feature upsampling, spatial scaling, and channel offset mechanisms, the present invention can adaptively enhance texture details, suppress local noise and blur, effectively improve the structural consistency and detail fidelity of the restored image, which is superior to the traditional method of blindly processing all degradation. Finally, this invention constructs a joint optimization framework (step C1) consisting of reconstruction loss, deep semantic guidance loss, and degradation feature decoupling loss. This framework constrains the learning process from multiple dimensions, including content accuracy, semantic consistency, and degradation orthogonality, enabling the model to simultaneously improve restoration quality and generalization ability in multi-degradation scenarios. Compared to existing methods that require training from scratch or additional design of multi-degradation labels, this invention achieves lightweight adaptation while maintaining the core structure, effectively reducing training and deployment costs. It possesses portability and practical application value, making it particularly suitable for multi-degradation image restoration tasks in complex real-world scenarios. Attached Figure Description

[0098] Figure 1 This is a structural diagram of the multi-degraded image restoration network (see step B) in a preferred embodiment of the present invention.

[0099] Figure 2 This is a structural diagram of the degradation feature progressive decoupling module (see step B2) in a preferred embodiment of the present invention.

[0100] Figure 3 This is a structural diagram of the progressive guidance module for degradation features at each scale in a preferred embodiment of the present invention (see step B3).

[0101] Figure 4 This is a structural diagram of the degradation feature-guided adaptive spatial scaling module (see step B331) and the degradation feature-guided adaptive channel offset module (see step B332) in a preferred embodiment of the present invention. Detailed Implementation

[0102] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0103] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0104] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application; as used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise, it will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, devices, components and / or combinations thereof, but do not preclude the presence or addition of one or more other features, steps, operations, devices, components and / or combinations thereof.

[0105] Multi-degradation image restoration method based on degradation feature decoupling and guidance, reference Figures 1-4 , comprising the following steps:

[0106] Step A, pre-processing the input image, including image pairing, cropping and data enhancement processing, to obtain an image training data set;

[0107] Step B, designing a multi-degradation image restoration network, including a multi-scale image encoder, a degradation feature progressive decoupling module, a degradation feature progressive guidance module for each scale, and a multi-scale image decoder;

[0108] Step C, designing a learning framework and loss function for training the network designed in step B, and training the network in step B using the image training data set described in step A to obtain a trained multi-degradation image restoration model;

[0109] Step D, inputting the image to be tested into the model obtained in step C to generate a restored image.

[0110] Further, the step A is implemented as follows:

[0111] Step A1, pairing a non-degraded normal image with a multi-degradation image, wherein the non-degraded normal image is used as a label image;

[0112] Step A2, randomly cropping each multi-degradation image with a size of to an image with a size of , while using the same random cropping method for its corresponding non-degraded normal image to ensure that they have the same size and position, wherein, and represent the height and width of the multi-degradation image and the non-degraded normal image, respectively, and represent the height and width of the cropped image, respectively;

[0113] Step A3, for each training paired image, randomly apply one of the following 8 data augmentation methods: keep the original image, vertical flip, rotate 90 degrees, rotate 90 degrees and then vertical flip, rotate 180 degrees, rotate 180 degrees and then vertical flip, rotate 270 degrees, rotate 270 degrees and then vertical flip.

[0114] Further, the step B is specifically implemented as follows:

[0115] Step B1, design a multi-scale image encoder, use the encoder of the URWKV backbone network as the multi-scale image encoder, perform feature extraction on the input multi-degradation image to obtain multi-scale intermediate features , then the first scale intermediate feature is obtained after bilinear two times down-sampling to obtain the backbone feature ; wherein , , 3 are the height, width and RGB channel number of the multi-degradation image, , the number of multi-scale intermediate features changes according to the scale change of the multi-scale image encoder, and the value is , the first scale intermediate feature in the multi-scale feature is , , , and are the height, width and channel number of the first scale intermediate feature in the multi-scale intermediate feature, , and are the height, width and channel number of the backbone feature;

[0116] Step B2, design a degradation feature progressive decoupling module, which is composed of degradation feature decoupling modules in turn, each degradation feature decoupling module mainly consists of a reversible group normalization unit, a reversible feature projection unit, a degradation feature standardization module, a spatial adaptive degradation feature scaling module and a degradation feature purification separation operation; wherein the reversible group normalization unit is mainly composed of a group normalization layer with a group number of 8, which contains forward group normalization operation and reverse group normalization operation, the reversible feature projection unit is mainly composed of a convolution kernel with a size of The system consists of unbiased convolutional layers with a stride of 1, padding of 0, and a group number of 1, including both forward and backward convolution operations. The progressive decoupling module for degradation features is used to extract degradation features sequentially, decomposing the mixed degradation into multiple degradation subspaces with clear semantics, while simultaneously removing degradation from the backbone features. Specifically, for the backbone features extracted in step B1... After the degradation feature progressive decoupling module, the following is obtained: Degenerative features of the same size and degenerative skeletal features ;in This represents the number of degenerate features, with values ​​ranging from 1 to 2. , No. The degradation characteristics are , , The number of channels representing the degradation features; this process is expressed by the formula:

[0117]

[0118] in, This indicates a decoupling module for degradation features. This indicates the stacking process of decoupling modules for degradation features;

[0119] Step B3: Design progressive guidance modules for degradation features at each scale. Each progressive guidance module for degradation features at each scale consists of... The degradation feature guidance modules are cascaded sequentially. Each degradation feature guidance module mainly consists of a degradation feature upsampling module, a reversible group normalization unit, a reversible feature projection unit, a spatially adaptive degradation feature modulation module, and a degradation feature guidance separation operation. The reversible group normalization unit mainly consists of a group normalization layer with 8 groups, including forward group normalization and reverse group normalization operations. The reversible feature projection unit mainly consists of a convolutional kernel with a size of... The system consists of a single, unbiased convolutional layer with a stride of 1, padding of 0, and 1 group. It includes forward and backward convolution operations, with the backward convolution operation implemented using the inverse operator obtained from LU decomposition to ensure feature dimensionality and information integrity. The progressive degenerate feature guidance module removes degeneracy from the corresponding scale features based on the degenerate features. For the first... The scale, the first one extracted in step B1 Scale intermediate features The degradation features extracted in step B2 are used as direct input. As an auxiliary input, after the first After processing by the progressively guided module for the degradation features of the scale, the first... Scale de-degradation intermediate features The process can be expressed by the following formula:

[0120]

[0121] in, Indicates the first Gradual guidance module for scale degradation features. Indicates the first The degradation characteristics of scale are progressively guided by the stacking of modules;

[0122] Completed After each scale, collect the dedegradable intermediate features at each scale. Multi-scale de-degradation intermediate features are obtained. ;

[0123] Step B4: Design a multi-scale image decoder. Use the decoder of the URWKV backbone network as the multi-scale image decoder, and process the de-degenerate backbone features obtained in step B2. and the multi-scale de-degradation intermediate features obtained in step B3 As input to the multi-scale image decoder; specifically, the de-degradation backbone features are used. and the Scale de-degradation intermediate features As the first The input to the scale image decoder, then the... The output of the scale image decoder and As the first The input to the scale image decoder is thus gradually processed. After decoding the image at each scale, the output is the restored image. .

[0124] Furthermore, step B2 is specifically implemented as follows:

[0125] Step B21: Extract the backbone features obtained in step B1. The first degenerate feature decoupling module uses its input features as input features, and then the output features of the first degenerate feature decoupling module use its input features as input features for the second degenerate feature decoupling module. This is how the degenerate feature decoupling modules are implemented. This stacking process yields the output of the progressive decoupling module for degenerate features, i.e., the degenerate backbone features. ,at the same time The decoupling module for each degradation feature outputs during intermediate execution. Degenerative features of the same size For the first The decoupling module for the degradation feature will take the feature from the previous one (i.e., the first one) The output features of the degradation feature decoupling module are denoted as follows: (If this is the first degenerate feature decoupling module, then...) Recorded as ), First, a forward group normalization operation is performed through a reversible group normalization unit, and then a forward convolution operation is performed through a reversible feature projection unit to obtain the backbone projection features. Then, the skeleton projection features Based on the channel dimension, it is split into two features to obtain the core content features. and skeletal degeneration characteristics ; characteristics of skeletal degeneration Input degradation feature standardization module to obtain standardized backbone degradation features Standardized backbone degeneration characteristics As direct input, core content features As auxiliary input, after passing through the spatial adaptive degradation feature scaling module, the features to be de-degraded are obtained. ; characteristics of skeletal degeneration As direct input, features to be degraded As auxiliary input, after degradation feature purification and separation operations, degradation features are obtained. and de-degeneration characteristics De-degenerate features and core content characteristics The features are merged into a single feature based on the channel dimension. The merged feature is then input into a reversible feature projection unit to perform a backward convolution operation, and then input into a reversible group normalization unit to perform a backward group normalization operation, yielding the output feature. When stacking to the last degenerate feature decoupling module, the output feature obtained is the output feature of the degenerate feature progressive decoupling module, i.e., the de-degenerate backbone feature. ,Then The degradation feature decoupling module generates during the intermediate execution process Each degradation feature consists of ;

[0126] Step B22: Design a degradation feature standardization module, consisting of group normalization units, ... It consists of depthwise convolutional units and group normalization units; the group normalization unit mainly consists of groups of... The group normalization layer composition, Depthwise convolutional units mainly consist of convolutional kernels with a size of Step size is 1, fill size is 1, number of groups is The system consists of unbiased convolutional layers; the degradation feature standardization module is used to perform grouping and standardization operations on the input degradation features, converting complex degradation features into standard degradation features; specifically, it standardizes the backbone degradation features. sequentially through group normalization units, After the depth-wise convolution unit and the group normalization unit, the standardized backbone degradation feature is obtained ; wherein , and are the height, width and channel number of the standardized backbone degradation feature respectively; the process is expressed by the formula:

[0127]

[0128] wherein, the group normalization unit is denoted by , and the depth-wise convolution unit is denoted by

[0129] Step B23, designing a space adaptive degradation feature scaling module, which is composed of an adaptive pooling layer, a bias point-wise convolution unit and a Sigmoid operation unit in sequence; wherein the bias point-wise convolution unit is mainly composed of a bias convolution layer with a convolution kernel size of , a step of 1, a padding of 0 and a group number of ; the space adaptive degradation feature scaling module is used to adaptively adjust the intensity of the degradation feature according to the spatial distribution of the content feature, so as to improve the decoupling ability of the degradation feature and the content feature; specifically: after the backbone content feature passes through the adaptive pooling layer, the bias point-wise convolution unit and the Sigmoid operation unit in sequence, the channel attention weight is obtained, and the channel-wise multiplication operation is performed between the channel attention weight and the standardized backbone degradation feature to obtain the to-be-degraded feature ; the process is expressed by the formula:

[0130]

[0131] wherein, the adaptive pooling layer unit is denoted by , the bias point-wise convolution unit is denoted by , the Sigmoid operation is denoted by , and the channel-wise multiplication operation is denoted by .

[0132] Step B24, designing a degradation feature purification separation operation, which is mainly composed of a GELU activation function and a subtraction operation, and is used to separate the degradation feature and the degrading feature; specifically: after the to-be-degraded feature passes through the GELU activation function, the degrading feature is obtained, and the backbone degradation feature​​ With de-degeneration characteristics Subtraction yields degenerative characteristics. This process can be expressed by the following formula:

[0133]

[0134]

[0135] in, This represents the GELU activation function operation.

[0136] Furthermore, step B3 is specifically implemented as follows:

[0137] Step B31: Extract the first... Scale intermediate features As the first The first degenerate feature guides the input features of the module, then the second... The output feature of the first degradation feature guidance module is used as the first... The second degradation feature guides the input features of the scale module, thus performing the second degradation feature. Scale degradation feature guidance module The stacking process yields the... The output of the scale degradation feature progressive guidance module, i.e., the first... Scale de-degeneration skeletal features For the first Scale No. A degradation feature guidance module will extract the degradation features obtained in step B2. As the second input feature, After the degradation feature upsampling module, the result is the same as the first... Scale intermediate features The width and height are the same. Scale degradation guided features At the same time, it will come from the previous one (i.e., the first one). The output features of the degradation feature guidance module are denoted as ( ). (If this is the first degenerate feature guidance module, then...) Recorded as ), First, input the invertible group normalization unit to perform a forward group normalization operation, then input the invertible feature projection unit to perform a forward convolution operation, and obtain the first... Scale projection features Then the first Scale projection features Decomposed into two features according to the channel dimension, resulting in the first... Scale content features and the Scale degradation features ; will the first Scale content features As direct input, the Scale degradation guided features As an auxiliary input, after passing through the spatial adaptive degradation feature modulation module, the first... Scale content enhancement features ; will the first Scale degradation features As direct input, the Scale degradation guided features As auxiliary input, after the degradation feature-guided separation operation, the first... Scale de-degradation features ; will the first Scale de-degradation features and the Scale content reorganization features The features are merged into a single feature based on the channel dimension. The merged feature is then input into a reversible feature projection unit to perform a backward convolution operation, and then input into a reversible group normalization unit to perform a backward group normalization operation, yielding the output feature. When stacked to the ; When the last degradation feature guides the module at the scale, the output feature obtained is the first... The output feature of the scale degradation feature progressively guided module, i.e., the first... Scale de-degradation intermediate features ;

[0138] Step B32: Design a degradation feature upsampling module, mainly consisting of group normalization units, It consists of pointwise convolutional units and spatial channel shuffling units; the group normalization unit mainly consists of groups with a number of groups. The group normalization layer composition, Pointwise convolutional units mainly consist of a convolutional kernel size of Step size is 1, padding is 0, number of groups is The spatial channel shuffling units are composed of unbiased convolutional layers with an upsampling factor of 1. The pixel shuffling layer is composed of; the degradation feature upsampling module is used to spatially reconstruct low-resolution degradation features; specifically, it reconstructs the degradation features... sequentially through group normalization units, After pointwise convolution units and spatial channel shuffling units, we obtain the first... Scale degradation guided features The process can be expressed by the following formula:

[0139]

[0140] wherein, denotes a group normalization unit, denotes a point-wise convolution unit, denotes a spatial channel shuffle unit;

[0141] Step B33, a spatially adaptive degradation feature modulation module is designed, mainly composed of a degradation feature guided adaptive spatial scaling module and a degradation feature guided adaptive channel shift module; the spatially adaptive degradation feature modulation module is used for degradation perception from two scales of local and global and enhancement of content features; specifically: the first scale degradation guided feature is input into the degradation feature guided adaptive spatial scaling module, and the first scale content enhancement weight is obtained. The first scale degradation guided feature is directly input, and the first scale content feature is input as auxiliary input, and after passing through the degradation feature guided adaptive channel shift module, the first scale channel enhancement weight is obtained. The first scale content feature is multiplied by the first scale content enhancement weight element by element, and then added with the first scale channel enhancement weight element by element to obtain the first scale content enhancement feature ; this process can be represented by the formula:

[0142]

[0143]

[0144]

[0145] wherein, denotes a degradation feature guided adaptive spatial scaling module, denotes a degradation feature guided adaptive channel shift module, denotes an element by element multiplication operation, denotes an element by element addition operation.

[0146] Step B34, a degradation feature guided separation operation is designed, mainly composed of a subtraction operation, used for separating the first scale degradation guided feature from the first separating the scale-degraded feature; specifically, the first scale-degraded feature and the second scale-degraded guiding feature subtraction to obtain the first scale-degraded feature ; the process can be represented by the formula:

[0147]

[0148] Further, the step B33 specifically implements the steps as follows:

[0149] Step B331, designing a degradation feature guiding adaptive spatial scaling module, mainly composed of group convolution unit, group normalization unit, GELU activation function, bias pointwise convolution unit and Sigmoid operation and element-wise addition operation; wherein The group convolution unit is composed of a bias-free convolution layer with a convolution kernel size of , a step of 1, a padding of 1, and a group number of The group normalization unit is mainly composed of a group normalization layer with a group number of The bias pointwise convolution unit is mainly composed of a bias convolution layer with a convolution kernel size of , a step of 1, a padding of 0, and a group number of The degradation feature guiding adaptive spatial scaling module utilizes the spatial distribution of the first scale-degraded guiding feature to generate pixel-by-pixel spatial weights to locally enhance or suppress the first scale content feature; specifically: for the first scale-degraded guiding feature obtained in step B32, it is sequentially subjected to group convolution unit, group normalization unit, GELU activation function, bias pointwise convolution unit and Sigmoid operation, and then a unit bias is added to obtain the first scale content-enhancing weight ; the process can be represented by the formula:

[0150]

[0151] wherein, represents group convolution unit, group normalization unit, GELU activation function, bias point-wise convolution unit, denotes a Sigmoid operation, denotes an element-wise multiplication operation, denotes an element-wise addition operation;

[0152] Step B332, designing a degradation feature guided adaptive channel shift module, mainly composed of an adaptive pooling layer with bias a point-wise convolution unit, a GELU activation function, and a bias a point-wise convolution unit with bias a point-wise convolution unit, a Tanh activation function, and a channel-wise multiplication operation; the degradation feature guided adaptive channel shift module is used for global conditional control of the content feature and adjustment of the overall de-degradation strength of the degradation feature; specifically, for the first scale degradation guided feature , sequentially passing through an adaptive pooling layer with bias a point-wise convolution unit, a GELU activation function, and a bias a point-wise convolution unit with bias a point-wise convolution unit, a Tanh activation function, and a channel-wise multiplication operation, the first scale content feature is obtained by performing a channel-wise multiplication operation, and the first scale channel-wise enhancement weight is obtained; this process can be represented by the formula:

[0153]

[0154] wherein, denotes an adaptive pooling layer, with bias a point-wise convolution unit, denotes a GELU activation function, denotes a Tanh activation function, denotes a channel-wise multiplication operation.

[0155] Further, the step C is specifically implemented as follows:

[0156] Step C1, constructing a total loss containing a reconstruction loss , a deep semantic guided loss , and a degradation feature decoupling loss ; the expression is:

[0157] ​

[0158] wherein, , and are balance parameters;

[0159] Step C2, randomly divide the image training data set obtained in step A into several batches, each batch containing several pairs of images;

[0160] Step C3, calculate the gradient of the parameters in the network using the back propagation method according to the loss, and update the network parameters using the Adam optimization method;

[0161] Step C4, repeat steps C2 to C3 until the model converges, and finally obtain the trained multi-degradation image restoration model.

[0162] Further, the step C1 is specifically implemented as follows:

[0163] Step C11, design a reconstruction loss mainly composed of L1 loss , structure loss and VGG perception loss ; Its expression is:

[0164]

[0165]

[0166]

[0167]

[0168] wherein, , and are balance parameters; is a restored image, is a label image, represents an absolute error value calculation operation; and represent the mean value obtained by local statistics using a Gaussian window of , and represent the variance obtained by local statistics using a Gaussian window of , and are two constant values to prevent the denominator from being 0; represents a mean square error value calculation operation, represents the extraction operation of the 3rd, 8th and 15th layer features using the VGG-16 classification model pre-trained on the ImageNet dataset;

[0169] Step C12: Design deep semantic guidance loss Mainly due to intra-feature similarity loss and feature similarity loss Composition; the deep semantic guidance loss This is used to constrain the deep features of the model and enhance its semantic representation capabilities; specifically, it involves using a visual base model, such as DINOv3, to extract label images. semantic features Then, the skeletal features extracted in step B1 are... enter In the pointwise convolutional unit, the backbone projection semantic features are obtained. Next, semantic features and backbone projection semantic features Calculate the intra-feature similarity loss and feature similarity loss Intra-feature similarity loss and feature similarity loss By performing a weighted summation, we obtain the deep semantic guidance loss. ;in, Pointwise convolutional units mainly consist of a convolutional kernel with a size of Step size is 1, padding is 0, number of groups is It consists of unbiased convolutional layers; this process can be expressed by the formula:

[0170]

[0171]

[0172]

[0173]

[0174]

[0175]

[0176]

[0177] in, and Represents the balance parameters. Represents the basic visual model. express Pointwise convolutional unit, Indicates features Size aligned to feature Bilinear dimension alignment operation. denotes a characteristic normalization operation, denotes a cosine similarity calculation operation, denotes an absolute error value calculation operation;

[0178] Step C13, design degradation characteristic decoupling loss , mainly composed of non-degradation guided loss , multi-degradation contrast loss and multi-degradation orthogonal loss ; the degradation characteristic decoupling loss is used to guide the degradation characteristics of normal images to tend to non-degradation, and guide the degradation characteristics of degraded images to tend to be distinguishable and decoupled; its expression is:

[0179]

[0180] wherein, , and are balance parameters.

[0181] Further, the step C13 is specifically implemented as follows:

[0182] Step C131, design non-degradation guided loss , the non-degradation guided loss is used to guide the degradation characteristics of normal images to tend to non-degradation; specifically: for the label image , input it into the multi-scale image encoder designed in step B1 to obtain the extracted multi-scale non-degradation intermediate features and backbone features , let be the non-degradation image backbone feature , input into the degradation characteristic progressive decoupling module designed in step B2 to obtain degradation characteristics of the same size , let be the non-degradation image degradation characteristic , calculate the absolute error value of the non-degradation image degradation characteristic and the all-zero feature to obtain the non-degradation guided loss ; wherein , and are the height, width and channel number of the non-degradation image backbone feature, denotes the number of degradation characteristics, and the value is 6, the th non-degradation image degradation characteristic is , , is the channel number of the degradation characteristic; this process is represented by the formula:

[0183]

[0184] in, This indicates the operation for calculating the absolute error value.

[0185] Step C132: Design multi-degradation contrast loss and multiple degenerate orthogonal losses The multi-degradation contrast loss The multi-degeneracy orthogonal loss is used to constrain the degenerate features of different degeneracy types to tend towards distinguishability in the feature space. This is used to constrain the degradation features of different degradation types to remain pairwise orthogonal in the feature space, so as to promote the decoupling of each degradation factor; specifically: for the features obtained in step B2 Degenerative features of the same size For each degradation feature After normalization, a stacking operation is performed, and then the matrix product between the stacked degenerate features is calculated to obtain the degenerate feature similarity matrix. Next, the temperature-inclusive information noise contrast estimation loss is calculated on the degenerate feature similarity matrix to obtain the multi-degenerate contrast loss. The absolute error values ​​of the degenerate feature similarity matrix and the identity matrix are calculated to obtain the multi-degenerate orthogonal loss. The process can be expressed by the following formula:

[0186]

[0187]

[0188]

[0189]

[0190] in, This indicates a feature normalization operation. This represents a stacking operation of multiple features. This represents the temperature coefficient, with a value of 0.07. This represents the function for calculating the information-noise contrast estimation loss. Represents the identity matrix. This indicates the operation for calculating the absolute error value.

[0191] Furthermore, step D is specifically implemented as follows:

[0192] Step D1, change the size to Multiple degraded test images Perform a boundary padding operation, using all zero pixels to fill its height and width to the required values. and The resulting size is Image ,in , These represent the height and width of the multi-degraded test image, respectively. , These represent the height and width of the multi-degraded test image after filling;

[0193] Step D2: Fill the multi-degraded test image with the filler. Input the data into the multi-degradation image restoration model trained in step C to obtain... Size restoration image ,Will Let it be the image for reasoning and reconstruction. ;

[0194] Step D3, change the size to Reasoning and reconstruction image Perform a boundary removal and padding operation to remove the boundary padding area, resulting in a final inference and reconstruction image of the same size as before padding. .

Claims

1. A method for restoring multiple degraded images based on degradation feature decoupling and guidance, characterized in that, Includes the following steps: Step A: Preprocess the input images, including image pairing, cropping, and data augmentation, to obtain the image training dataset; Step B: Design a multi-degraded image restoration network, including a multi-scale image encoder, a degradation feature progressive decoupling module, a scale-by-scale degradation feature progressive guidance module, and a multi-scale image decoder; Step C: Design a learning framework and loss function for training the multi-degradation image restoration network designed in Step B, and use the image training dataset described in Step A to train the multi-degradation image restoration network in Step B to obtain the trained multi-degradation image restoration model. Step D: Input the image to be tested into the multi-degraded image restoration model obtained in step C to generate the restored image.

2. The method for restoring multiple degraded images based on degradation feature decoupling and guidance according to claim 1, characterized in that, The specific implementation steps of step A are as follows: Step A1: Pair the non-degradable normal image with the multi-degradable image, where the non-degradable normal image is used as the label image; Step A2: Divide each sheet into sheets of size [size missing]. Multiple degraded images are randomly cropped into Images of the same size are cropped using the same random cropping method on their corresponding non-degradable normal counterparts to ensure they have the same size and position. and These represent the height and width of the degraded image and the normal image without degradation, respectively. and These represent the height and width of the cropped image, respectively. Step A3: For each training paired image, randomly apply one of the following 8 data augmentation methods: keep the original image, flip vertically, rotate 90 degrees, rotate 90 degrees and then flip vertically, rotate 180 degrees, rotate 180 degrees and then flip vertically, rotate 270 degrees, rotate 270 degrees and then flip vertically.

3. The method for restoring multiple degraded images based on degradation feature decoupling and guidance according to claim 1, characterized in that, The specific implementation steps of step B are as follows: Step B1: Design a multi-scale image encoder, using the URWKV backbone network encoder as the multi-scale image encoder for the input multi-degraded images. Feature extraction is performed to obtain multi-scale intermediate features. Then the first Scale intermediate features Bone features were obtained after bilinear downsampling by a factor of 2. ;in , 3 represents the height, width, and number of RGB channels of the multi-degraded image, respectively. This represents the number of intermediate features across multiple scales, which varies depending on the scale of the multi-scale image encoder. The Middle Scale intermediate features are , , , and These are the first intermediate features in the multi-scale model. The height, width, and number of channels of the intermediate features at scale. , and These are the height, width, and number of channels of the backbone feature, respectively; Step B2: Design a progressive decoupling module for degradation features, by The degenerate feature decoupling modules are cascaded together. Each module includes a reversible group normalization unit, a reversible feature projection unit, a degenerate feature standardization module, a spatially adaptive degenerate feature scaling module, and a degenerate feature purification and separation operation. The reversible group normalization unit includes an 8-group normalization layer, containing forward and reverse group normalization operations. The reversible feature projection unit includes a convolutional kernel with a size of [missing information - likely a value]. A non-biased convolutional layer with a stride of 1, padding of 0, and number of groups of 1, including forward convolution and backward convolution operations; the progressive decoupling module for degradation features is used to extract degradation features step by step in sequence, decompose the mixed degradation into multiple degradation subspaces with clear semantics, and remove degradation from the backbone features at the same time. Specifically, for the backbone features extracted in step B1... After the degradation feature progressive decoupling module, the following is obtained: Degenerative features of the same size and degenerative skeletal features ;in Indicates the number of degradation features, the first The degradation characteristics are , , The number of channels representing the degradation feature; the process is expressed by the formula: in, This indicates a decoupling module for degradation features. This indicates the stacking process of decoupling modules for degradation features; Step B3: Design progressive guidance modules for degradation features at each scale. Each progressive guidance module for degradation features at each scale consists of... The degradation feature guidance modules are cascaded sequentially. Each degradation feature guidance module includes a degradation feature upsampling module, a reversible group normalization unit, a reversible feature projection unit, a spatially adaptive degradation feature modulation module, and a degradation feature guidance separation operation. The reversible group normalization unit includes a group normalization layer with 8 groups, containing forward group normalization and reverse group normalization operations. The reversible feature projection unit includes a convolutional kernel with a size of [missing information]. A non-biased convolutional layer with a stride of 1, padding of 0, and 1 group, containing forward and backward convolution operations, wherein the backward convolution operation is implemented using the inverse operator obtained from LU decomposition; the degradation feature progressive guidance module removes degradation from the corresponding scale features based on the degradation features; for the 1st The scale, the first one extracted in step B1 Scale intermediate features The degradation features extracted in step B2 are used as direct input. As an auxiliary input, after the first After processing by the progressively guided module for the degradation features of the scale, the first... Scale de-degradation intermediate features The process can be expressed by the following formula: in, Indicates the first Gradual guidance module for scale degradation features. Indicates the first The degradation characteristics of scale are progressively guided by the stacking of modules; Completed After each scale, collect the dedegradable intermediate features at each scale. Multi-scale de-degradation intermediate features are obtained. ; Step B4: Design a multi-scale image decoder. Use the decoder of the URWKV backbone network as the multi-scale image decoder, and process the de-degenerate backbone features obtained in step B2. and the multi-scale de-degradation intermediate features obtained in step B3 As input to the multi-scale image decoder; specifically, the de-degradation backbone features are used. and the Scale de-degradation intermediate features As the first The input to the scale image decoder, then the... The output of the scale image decoder and As the first The input to the scale image decoder is thus gradually processed. After decoding the image at each scale, the output is the restored image. .

4. The multi-degraded image restoration method based on degradation feature decoupling and guidance according to claim 3, characterized in that, The specific implementation steps of step B2 are as follows: Step B21: Extract the backbone features obtained in step B1. The first degenerate feature decoupling module uses its input features as input features, and then the output features of the first degenerate feature decoupling module use its input features as input features for the second degenerate feature decoupling module. This is how the degenerate feature decoupling modules are implemented. This stacking process yields the output of the progressive decoupling module for degenerate features, i.e., the degenerate backbone features. ,at the same time The decoupling module for each degradation feature outputs during intermediate execution. Degenerative features of the same size For the first The decoupling module for the degradation feature will be derived from the previous one, i.e., the first one. The output features of each degradation feature decoupling module are denoted as follows: If this is the first degenerate feature decoupling module, then Recorded as , First, a forward group normalization operation is performed through a reversible group normalization unit, and then a forward convolution operation is performed through a reversible feature projection unit to obtain the backbone projection features. Then, the skeleton projection features Based on the channel dimension, it is split into two features to obtain the core content features. and skeletal degeneration characteristics ; characteristics of skeletal degeneration Input degradation feature standardization module to obtain standardized backbone degradation features Standardize the characteristics of backbone degeneration As direct input, core content features As auxiliary input, after passing through the spatial adaptive degradation feature scaling module, the features to be de-degraded are obtained. ; characteristics of skeletal degeneration As direct input, features to be degraded As auxiliary input, after degradation feature purification and separation operations, degradation features are obtained. and de-degeneration characteristics De-degenerate features and core content characteristics The features are merged into a single feature based on the channel dimension. The merged feature is then input into a reversible feature projection unit to perform a backward convolution operation, and then input into a reversible group normalization unit to perform a backward group normalization operation, yielding the output feature. When stacking to the last degenerate feature decoupling module, the output feature obtained is the output feature of the degenerate feature progressive decoupling module, i.e., the de-degenerate backbone feature. ,Then The degradation feature decoupling module generates during the intermediate execution process Each degradation feature consists of ; Step B22: Design a degradation feature standardization module, which includes, in sequence, group normalization units, Depthwise convolutional units and group normalization units; The group normalization unit includes a number of groups. The group normalization layer, Depthwise convolutional units include convolutional kernels of size 1. Step size is 1, fill size is 1, number of groups is The unbiased convolutional layer; the degradation feature standardization module is used to perform grouping and standardization operations on the input degradation features, converting complex degradation features into standard degradation features; specifically, it performs standardization operations on the backbone degradation features. sequentially through group normalization units, After depthwise convolutional units and group normalized units, standardized backbone degradation features are obtained. ;in , and These represent the height, width, and number of channels of the standardized backbone degradation feature; the process is expressed by the formula: in, Represents the group normalized unit, express Depthwise convolutional units; Step B23: Design a spatial adaptive degradation feature scaling module, which includes, in sequence,... Adaptive pooling layer with bias Pointwise convolutional units and sigmoid operations; with bias Pointwise convolutional units include convolutional kernels of size 1. Step size is 1, padding is 0, number of groups is The biased convolutional layer; the spatial adaptive degradation feature scaling module is used to adaptively adjust the intensity of degradation features according to the spatial distribution of content features; specifically, it scales the backbone content features... Passing through in sequence Adaptive pooling layer with bias After pointwise convolutional units and sigmoid operation units, channel attention weights are obtained, which are then combined with standardized backbone degradation features. Perform channel-by-channel multiplication to obtain the features to be degraded. The process can be expressed by the following formula: in, express Adaptive pooling layer unit, Indicates bias Pointwise convolutional unit, This represents the Sigmoid operation. This indicates a channel-by-channel multiplication operation; Step B24: Design a degradation feature purification and separation operation, including the GELU activation function and subtraction operation, to separate degradation features and de-degrading features; specifically: [The following is a more detailed explanation of the steps involved in the process:] ... After applying the GELU activation function, de-degenerate features are obtained. , to reduce the characteristics of skeletal degeneration With de-degeneration characteristics Subtraction yields degenerative characteristics. The process can be expressed by the following formula: in, This represents the GELU activation function operation.

5. The multi-degraded image restoration method based on degradation feature decoupling and guidance according to claim 3, characterized in that, The specific implementation steps of step B3 are as follows: Step B31: Extract the first... Scale intermediate features As the first The first degenerate feature guides the input features of the module, then the second... The output feature of the first degradation feature guidance module is used as the first... The second degradation feature guides the input features of the scale module, thus performing the second degradation feature. Scale degradation feature guidance module The stacking process yields the... The output of the scale degradation feature progressive guidance module, i.e., the first... Scale de-degeneration skeletal features For the first Scale No. A degradation feature guidance module will extract the degradation features obtained in step B2. As the second input feature After the degradation feature upsampling module, the result is the same as the first... Scale intermediate features The width and height are the same. Scale degradation guided features At the same time, it will come from the previous one, that is, the first The output features of the degradation feature guidance module are denoted as follows: If this is the first degenerate feature guidance module, then Recorded as , First, input the invertible group normalization unit to perform a forward group normalization operation, then input the invertible feature projection unit to perform a forward convolution operation, and obtain the first... Scale projection features Then the first Scale projection features Decomposed into two features according to the channel dimension, resulting in the first... Scale content features and the Scale degradation features ; will the first Scale content features As direct input, the Scale degradation guided features As an auxiliary input, after passing through the spatial adaptive degradation feature modulation module, the first... Scale content enhancement features ; will the first Scale degradation features As direct input, the Scale degradation guided features As auxiliary input, after the degradation feature-guided separation operation, the first... Scale de-degradation features ; will the first Scale de-degradation features and the Scale content reorganization features The features are merged into a single feature based on the channel dimension. The merged feature is then input into a reversible feature projection unit to perform a backward convolution operation, and then input into a reversible group normalization unit to perform a backward group normalization operation, yielding the output feature. When stacked to the ; When the last degradation feature guides the module at the scale, the output feature obtained is the first... The output feature of the scale degradation feature progressively guided module, i.e., the first... Scale de-degradation intermediate features ; Step B32: Design a degradation feature upsampling module, including group normalization units, Pointwise convolutional units and spatial channel shuffling units; The group normalization unit includes a number of groups. The group normalization layer, Pointwise convolutional units include convolutional kernels of size 1. Step size is 1, padding is 0, number of groups is Unbiased convolutional layers, spatial channel shuffling units include upsampling factors of The pixel shuffling layer; the degradation feature upsampling module is used to spatially reconstruct low-resolution degradation features; specifically, it reconstructs the degradation features... sequentially through group normalization units, After pointwise convolution units and spatial channel shuffling units, we obtain the first... Scale degradation guided features The process can be expressed by the following formula: in, Represents the group normalized unit, express Pointwise convolutional unit, This indicates a spatial channel shuffling unit; Step B33: Design a spatial adaptive degradation feature modulation module, including a degradation feature-guided adaptive spatial scaling module and a degradation feature-guided adaptive channel offset module; the spatial adaptive degradation feature modulation module is used to perform degradation perception from both local and global scales and enhance content features; specifically: ... Scale degradation guided features After inputting the degradation feature-guided adaptive spatial scaling module, the resulting... Scale-based content-based weighting , will the Scale degradation guided features As direct input, the first Scale content features As auxiliary input, after passing through the degradation feature-guided adaptive channel offset module, the first... Scale-based channel-wise weight enhancement , will the Scale content features With the Scale-based content-based weighting Perform element-wise multiplication, then multiply with the first element. Scale-based channel-wise weight enhancement Perform element-wise addition to obtain the first element. Scale content enhancement features The process can be expressed by the following formula: in, This represents an adaptive spatial scaling module guided by degradation features. This represents an adaptive channel offset module guided by degradation features. This indicates an element-wise multiplication operation. This represents an element-wise addition operation; Step B34: Design a degradation feature-guided separation operation, including a subtraction operation, to separate the first... Scale degradation guides features from the first Separation from scale degradation features; specifically, separating the first... Scale degradation features With the Scale degradation guided features Subtraction yields the first... Scale de-degradation features The process can be expressed by the following formula:

6. The method for restoring multiple degraded images based on degradation feature decoupling and guidance according to claim 5, characterized in that, The specific implementation steps of step B33 are as follows: Step B331: Design a degradation feature-guided adaptive spatial scaling module, including... Group convolutional units, group normalized units, GELU activation function, biased activation function Pointwise convolutional units, sigmoid operations, and element-wise addition operations; among which A group of convolutional units includes a convolutional kernel size of Step size is 1, fill size is 1, number of groups is The unbiased convolutional layer, with a group of normalized units including a number of groups. The group normalization layer with bias Pointwise convolutional units include convolutional kernels of size 1. Step size is 1, padding is 0, number of groups is The biased convolutional layer; the degradation feature-guided adaptive spatial scaling module utilizes the first Scale degradation guides the spatial distribution of features, generating pixel-wise spatial weights for the first pixel. Local enhancement or suppression of scale-based content features; Specifically, for the first obtained in step B32... Scale degradation guided features , in turn Group convolutional units, group normalized units, GELU activation function, biased activation function After pointwise convolution units and a sigmoid operation, a unit bias is added to obtain the first... Scale-based content-based weighting The process can be expressed by the following formula: in, express Group convolutional units, Represents the group normalized unit, This represents the GELU activation function. Indicates bias Pointwise convolutional unit, This represents the Sigmoid operation. This indicates an element-wise multiplication operation. This represents an element-wise addition operation; Step B332: Design a degradation feature-guided adaptive channel offset module, including... Adaptive pooling layer with bias Pointwise convolutional units, GELU activation function, biased Pointwise convolutional units with bias Pointwise convolutional units, Tanh activation functions, and channel-wise multiplication operations; the degradation feature-guided adaptive channel offset module is used to perform global conditional control on content features and adjust the overall de-degradation intensity of degradation features; specifically, for the first obtained in step B32... Scale degradation guided features , in turn Adaptive pooling layer with bias Pointwise convolutional units, GELU activation function, biased Pointwise convolutional units with bias After pointwise convolutional units and Tanh activation functions, it is then combined with the first... Scale content features Perform a channel-by-channel multiplication operation to obtain the first... Scale-based channel-wise weight enhancement The process can be expressed by the following formula: in, express Adaptive pooling layer With bias Pointwise convolutional unit, This represents the GELU activation function. This represents the Tanh activation function. This indicates a channel-by-channel multiplication operation.

7. The method for restoring multiple degraded images based on decoupling and guidance of degradation features according to claim 1, characterized in that, The specific implementation steps of step C are as follows: Step C1: Construct a model containing reconstruction loss. Deep semantic guidance loss Decoupling loss of degradation features Total loss ; Its expression is: in, , and For balance parameters; Step C2: Randomly divide the image training dataset obtained in step A into several batches, each batch containing several pairs of images; Step C3: Calculate the gradient of the network parameters using the backpropagation method based on the loss, and update the network parameters using the Adam optimization method; Step C4: Repeat steps C2 to C3 until the model converges, and finally obtain the trained multi-degradation image restoration model.

8. The method for restoring multiple degraded images based on degradation feature decoupling and guidance according to claim 7, characterized in that, The specific implementation steps of step C1 are as follows: Step C11, Design Reconstruction Loss Including L1 loss Structural loss and VGG perceived loss ; Its expression is: in, , and For balance parameters; To restore the image, For the label image, This indicates the operation for calculating the absolute error value. and Indicates adoption The mean obtained by performing local statistics using a Gaussian window. and Indicates adoption The variance obtained by performing local statistics using a Gaussian window. and Two constant values ​​are used to prevent the denominator from being 0; This indicates the operation for calculating the mean square error value. This indicates the extraction of features from layers 3, 8, and 15 using a VGG-16 classification model pre-trained on the ImageNet dataset. Step C12: Design deep semantic guidance loss Including intra-feature similarity loss and feature similarity loss The deep semantic guidance loss This is used to constrain the deep features of the model and enhance its semantic representation capabilities; specifically, it involves using a visual base model, such as DINOv3, to extract label images. semantic features Then, the skeletal features extracted in step B1 are... enter In the pointwise convolutional unit, the backbone projection semantic features are obtained. Next, semantic features and backbone projection semantic features Calculate the intra-feature similarity loss and feature similarity loss Intra-feature similarity loss and feature similarity loss By performing a weighted summation, we obtain the deep semantic guidance loss. ;in, Pointwise convolutional units include convolutional kernels of size 1. Step size is 1, padding is 0, number of groups is Unbiased convolutional layers; the process is expressed by the formula: in, and Represents the balance parameters. Represents the basic visual model. express Pointwise convolutional unit, Indicates features Size aligned to feature Bilinear dimension alignment operation. This indicates a feature normalization operation. This indicates the cosine similarity calculation operation. This indicates the operation for calculating the absolute error value. Step C13: Design the degradation feature decoupling loss. Including non-degradation guided loss Multi-degradation contrast loss and multiple degenerate orthogonal losses The degradation feature decoupling loss The expression is used to guide the degradation features of normal images to tend towards no degradation, and to guide the degradation features of degraded images to tend towards distinguishability and decoupling; its expression is: in, , and For balancing parameters.

9. The method for restoring multiple degraded images based on degradation feature decoupling and guidance according to claim 8, characterized in that, The specific implementation steps of step C13 are as follows: Step C131: Design a degenerate-free boot loss The non-degradation guiding loss Degradation features used to guide normal images toward no degradation; Specifically: for label images The extracted multi-scale non-degradable intermediate and backbone features are then input into the multi-scale image encoder designed in step B1 to obtain the extracted multi-scale non-degradable intermediate features and backbone features. ,Will The non-degradable image backbone features are denoted as... ,Will Input the degradation feature progressive decoupling module designed in step B2 to obtain Degenerative features of the same size ,Will Degradation features of non-degradable images are denoted as [missing information]. Degradation features for non-degradable images The absolute error value is calculated using all-zero features to obtain the non-degenerate guided loss. ;in , and These represent the height, width, and number of channels of the backbone features in a non-degradable image. Indicates the number of degradation features, the first The degradation features of a non-degradable image are: , , The number of channels representing the degradation feature; the process is expressed by the formula: in, This indicates the operation for calculating the absolute error value. Step C132: Design multi-degradation contrast loss and multiple degenerate orthogonal losses The multi-degradation contrast loss The multi-degeneracy orthogonal loss is used to constrain the degenerate features of different degeneracy types to tend towards distinguishability in the feature space. This is used to constrain the degradation features of different degradation types to remain pairwise orthogonal in the feature space, so as to promote the decoupling of each degradation factor; specifically: for the features obtained in step B2 Degenerative features of the same size For each degradation feature After normalization, a stacking operation is performed, and then the matrix product between the stacked degenerate features is calculated to obtain the degenerate feature similarity matrix. Next, the temperature-inclusive information noise contrast estimation loss is calculated on the degenerate feature similarity matrix to obtain the multi-degenerate contrast loss. The absolute error values ​​of the degenerate feature similarity matrix and the identity matrix are calculated to obtain the multi-degenerate orthogonal loss. The process can be expressed by the following formula: in, This indicates a feature normalization operation. This represents a stacking operation of multiple features. This represents the temperature coefficient, with a value of 0.

07. This represents the function for calculating the information-noise contrast estimation loss. Represents the identity matrix. This indicates the operation for calculating the absolute error value.

10. The method for restoring multiple degraded images based on degradation feature decoupling and guidance according to claim 1, characterized in that, The specific implementation steps of step D are as follows: Step D1, change the size to Multiple degraded test images Perform a boundary padding operation, using all zero pixels to fill its height and width to the specified values. and The resulting size is Image ,in , These represent the height and width of the multi-degraded test image, respectively. , These represent the height and width of the multi-degraded test image after filling; Step D2: Fill the multi-degraded test image with the filler. Input the data into the multi-degradation image restoration model trained in step C to obtain... Size restoration image ,Will Let it be denoted as the inference reconstruction image. ; Step D3, change the size to Reasoning and reconstruction image Perform a boundary removal and padding operation to remove the boundary padding area, resulting in a final inference and reconstruction image of the same size as before padding. .

Images