Processing method for realizing rapid restoration and enhancement of image
By constructing an image restoration enhancement model containing recursive residual groups and fusion modules, the problems of loss of image restoration details and insufficient accuracy in the prior art are solved, and a higher accuracy image restoration effect is achieved.
Patent Information
- Application Number
- CN202411915261.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-24
- Publication Date
- 2025-05-06
AI Technical Summary
The existing image restoration technology has problems such as loss of detail features and poor accuracy, resulting in poor image restoration effect.
The image restoration enhancement model is constructed, including the first convolutional layer, recursive residual group and fusion module. The fusion module integrates high-resolution features and context information through multi-scale residual blocks. The fusion module fuses the residual image and the original image to output the restored image.
By taking into account context information and spatial details, the accuracy and effect of image restoration are improved, which significantly improves the image quality after restoration.
Smart Images
Figure CN119941577A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and in particular to a processing method for realizing rapid restoration and enhancement of an image. Background Art
[0002] Image degradation of varying severity often occurs during image acquisition due to physical limitations of the camera or complex lighting conditions. For example, mobile phone cameras have narrow apertures and are equipped with small sensors with limited dynamic range, which often produce noisy and low-contrast images. Similarly, images taken under inappropriate lighting conditions can appear too dark or too bright. Image restoration technology aims to recover the original clear image from its damaged measurement results.
[0003] Recent advances in image restoration and enhancement are related to the development of deep learning models because they are able to learn powerful and generalizable prior knowledge from large-scale datasets. Existing convolutional neural networks generally follow two architectural designs: one is an encoder-decoder structure, and the other is high-resolution feature processing. The first architecture first maps the input to a low-resolution representation in steps, and then back-maps it to the original resolution in steps. Although this method learns enough contextual information by reducing the spatial resolution, the precise spatial details are lost, making it difficult to recover them in subsequent stages. The high-resolution networks of the second architecture do not use any processing to downsample. Although they can restore better spatial details, these networks have limited receptive fields and are not good at encoding contextual information.
[0004] Therefore, existing solutions generally have problems such as loss of detail features and poor accuracy when restoring images, resulting in poor image restoration effects. Summary of the invention
[0005] In view of this, in order to address the above shortcomings, it is necessary to propose a processing method for realizing rapid image restoration and enhancement, so as to improve the accuracy of image restoration and thus improve the effect of image restoration.
[0006] The present invention provides a processing method for realizing rapid restoration and enhancement of an image, comprising:
[0007] Constructing an image restoration and enhancement model; wherein the image restoration and enhancement model includes a first convolution layer, at least one recursive residual group RRG, a second convolution layer and a fusion module; the first convolution layer is used to extract underlying features of the original image; each of the recursive residual groups RRG includes at least one multi-scale residual block MRB, which is used to integrate high-resolution features by transferring context information from a low-resolution stream for the extracted underlying features, so as to restore both context information and spatial detail information; the second convolution layer is used to extract features of the deep features output by the recursive residual group RRG to obtain a residual image; the fusion module is used to fuse the residual image with the original image and output a restored image;
[0008] Obtaining the image to be restored;
[0009] The image to be restored is input into the image restoration enhancement model, and a restored image is output.
[0010] Preferably, each multi-scale residual block MRB includes a selective kernel feature fusion module and a residual context block;
[0011] The selective kernel feature fusion module is used to dynamically adjust the receptive field through a fusion unit and a selection unit; wherein the fusion unit is used to generate a global feature descriptor by combining information from multi-resolution streams, and the selection unit is used to recalibrate feature maps of different streams using the global feature descriptor and aggregate them to obtain a fused feature;
[0012] The residual context block is used to extract features from the fused features output by the selective kernel feature fusion module, so as to transfer informative features and suppress useless features.
[0013] Preferably, the fusion unit is configured to perform the following operations:
[0014] Receive feature information from two parallel convolutional streams carrying different scale information output by the first convolutional layer, and combine the multi-scale features by element-wise sum to obtain a combined feature L;
[0015] Applying global average pooling to the combined feature L in the spatial dimension to obtain channel statistical information s;
[0016] Processing the channel statistics s through a channel downscaling convolution layer to generate a compact feature vector z;
[0017] The feature vector z is passed through two parallel channel-upscale convolutional layers to obtain feature descriptors v1 and v2 respectively.
[0018] Preferably, the selection unit is configured to perform the following operations:
[0019] Apply the softmax function to the feature descriptors v1 and v2 respectively to obtain the attention values s1 and s2;
[0020] The obtained attention values s1 and s2 are used to calibrate the multi-scale features to obtain the fusion features of multi-resolution information.
[0021] Preferably, the calibrating the multi-scale features using the obtained attention values s1 and s2 includes:
[0022] The fusion features are obtained using the following calculation formula:
[0023] U=s1·L1+s2·L2
[0024] Among them, U is used to characterize the fusion feature, L1 is used to characterize the feature information of the channel where s1 is located, and L2 is used to characterize the feature information of the channel where s2 is located.
[0025] Preferably, the residual context block includes a grouped convolution module, a context module, a third convolution layer and an aggregation module;
[0026] The grouped convolution module is used to process the fused features through two 3×3 grouped convolution layers to obtain grouped convolution features;
[0027] The context module is used to generate attention aggregation features through the grouped convolutional features;
[0028] The third convolutional layer is used to process the attention aggregation feature;
[0029] The aggregation module is used to aggregate the output of the third convolutional layer and the fusion features to obtain contextual aggregation features that retain high-resolution spatial details.
[0030] Preferably, the context module is configured to perform the following operations:
[0031] The grouped convolution feature F b ∈R H×W×C Apply a 1×1 convolution layer, reshape and softmax to generate a new feature F c ∈R 1×1×HW ; Where H and W are the height and width of the feature map respectively;
[0032] The grouped convolutional features F b ∈R H×W×C Reshape to F b ∈R 1×HW×C , and with F c ∈R 1×1×HW Perform matrix multiplication to obtain the global feature descriptor F d ∈R1×1×C ; Where C is the number of channels;
[0033] The global feature descriptor F d ∈R 1×1×C Through two 1×1 convolutional layers, we get the new attention feature F e ∈R 1 ×1×C ;
[0034] The new attention feature F e ∈R 1×1×C Aggregate to group convolutional features F b ∈R H×W×C at each position.
[0035] Preferably, the aggregation module is specifically configured to perform aggregation processing using the following calculation formula:
[0036] F RCB =F a +ω(CM(F b ))
[0037] Among them, F RCB It is used to characterize the contextual aggregation features that retain high-resolution spatial details, F a is used to characterize the fusion feature, ω is used to characterize the third convolutional layer, and CM is used to characterize the context module.
[0038] Preferably, when constructing the image restoration and enhancement model, the following loss function is used to optimize the model:
[0039]
[0040] Among them, L is used to represent the loss function, A is used to characterize the image after the model is restored. * It is used to characterize the basic real image, and ε is the regularization term.
[0041] Preferably, when training the image restoration and enhancement model, smaller image blocks are used for training in the early stage, and gradually larger image blocks are used for training in the later stage.
[0042] It can be seen from the above technical solution that in the processing method for realizing rapid image restoration and enhancement provided by the present invention, an image restoration and enhancement model is first constructed, and then the image to be restored is obtained and input into the constructed image restoration and enhancement model, and the restored image can be output. When constructing the image restoration and enhancement model, its architecture includes a recursive residual group, and the recursive residual group is composed of a number of multi-scale residual blocks, which can integrate high-resolution features by transferring context information from the low-resolution stream for the extracted underlying features, and can not only obtain multi-scale context information, but also retain the high-resolution spatial details of the image, that is, the restored image can take into account both context information and spatial details, improve the accuracy of image restoration, and make the restored image effect better. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 A flowchart of a processing method for realizing rapid image restoration and enhancement provided by an embodiment of the present invention.
[0044] Figure 2 A general architecture diagram of an image restoration provided by an embodiment of the present invention.
[0045] Figure 3 An architectural diagram of a selective kernel feature fusion provided in an embodiment of the present invention.
[0046] Figure 4 An architectural diagram of a residual context block provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0048] like Figure 1 As shown, the present invention provides a processing method for realizing rapid restoration and enhancement of an image, and the method may include the following steps:
[0049] Step 101: construct an image restoration and enhancement model; wherein the image restoration and enhancement model includes a first convolution layer, at least one recursive residual group RRG, a second convolution layer and a fusion module; the first convolution layer is used to extract underlying features of the original image; each recursive residual group RRG includes at least one multi-scale residual block MRB, which is used to integrate high-resolution features by transferring context information from a low-resolution stream for the extracted underlying features, so as to restore both context information and spatial detail information; the second convolution layer is used to extract features of the deep features output by the recursive residual group RRG to obtain a residual image; the fusion module is used to fuse the residual image with the original image and output a restored image;
[0050] Step 102: Obtain the image to be restored;
[0051] Step 103: input the image to be restored into the image restoration enhancement model, and output the restored image.
[0052] In this embodiment, when constructing an image restoration and enhancement model, its architecture includes a recursive residual group, and the recursive residual group is composed of a number of multi-scale residual blocks, which can integrate high-resolution features by transmitting contextual information from a low-resolution stream for the extracted underlying features, and can not only obtain multi-scale contextual information, but also retain the high-resolution spatial details of the image, so that the restored image can take into account both contextual information and spatial details, thereby improving the accuracy of image restoration and making the restored image better.
[0053] Specifically, when constructing the image restoration enhancement model in step 101, it is considered to be obtained by training using sample image data. Specifically, the architecture used can be as follows Figure 2 As shown. It mainly includes the first convolution layer, several recursive residual groups RRG, the second convolution layer and the fusion module. Each recursive residual group can include several multi-scale residual blocks MRB, which can extract parallel multi-resolution convolution streams with richer semantics and spatially precise feature representation, realize information exchange across multi-resolution streams, realize attention-based feature aggregation from different streams, extract attention-based features, and then realize the transmission of context information from low-resolution streams and the integration of high-resolution features, so as to realize the recovery of both context information and spatial detail information.
[0054] like Figure 2 As shown, for the overall process, a graph A∈R is given H×W×3 , apply a convolutional layer to extract the underlying features F0∈R of image A H×W×C , then the extracted feature map F0 is passed through N recursive residual groups RRG to generate deep features F n ∈R H×W×CEach recursive residual group RRG contains several multi-scale residual blocks, and then the deep feature F n ∈R H×W×C Apply a convolutional layer to get the residual image Δ∈R H×W×3 The final restored image is
[0055] Among them, consider using the following loss function to optimize the model:
[0056]
[0057] Among them, L is used to represent the loss function, A~ is used to represent the image after the model is restored, and A * It is used to characterize the basic real image. ε is a regular term, which can be taken as 10 according to experimental experience. -3 .
[0058] In order to encode contrastive contextual information, existing convolutional neural networks usually adopt the following architectural design: (1) the receptive field of neurons in each stage is fixed; (2) the spatial size of feature maps is gradually reduced to generate strong low-resolution representations, and (3) high-resolution representations are gradually restored from low-resolution representations. However, in visual science, the local receptive field sizes of neurons in the same area of the visual cortex of primates are different. Therefore, it is more effective to combine with CNN to collect multi-scale spatial information in the same layer. Specifically, we propose a multi-scale residual block (MRB), which can generate spatially accurate outputs by maintaining high-resolution representations while receiving rich contextual information from low resolution. MRB consists of multiple parallel connected full convolution streams that process feature maps of different resolutions and integrate high-resolution features by transferring contextual information from low-resolution streams.
[0059] Specifically, for each multi-scale residual block MRB, it may include a selective kernel feature fusion module and a residual context block;
[0060] A selective kernel feature fusion module, used for dynamically adjusting the receiving field through a fusion unit and a selection unit; wherein the fusion unit is used to generate a global feature descriptor by combining information from multi-resolution streams, and the selection unit is used to recalibrate feature maps of different streams using the global feature descriptor and aggregate them to obtain a fused feature;
[0061] The residual context block is used to extract the fused features output by the selective kernel feature fusion module so as to transfer informative features and suppress useless features.
[0062] 1. For the selective kernel feature fusion module SKFF
[0063] A fundamental property of neurons in the visual cortex is the ability to change their receptive fields in response to stimuli. This mechanism of adaptively adjusting the receptive fields can be incorporated into CNNs by using multi-scale feature generation followed by feature aggregation and selection. The most commonly used feature aggregation methods include simple concatenation or summation. However, these approaches provide limited expressive power to the network. Therefore, a nonlinear process of fusing features of different resolution streams based on a self-attention mechanism is considered through a selective feature fusion module.
[0064] Specifically, Figure 3 As shown, the selective kernel feature fusion module dynamically adjusts the receptive field through “fusion” and “selection”. The “fusion” operator generates a global feature descriptor by combining information from multi-resolution streams, and the “selection” operator uses the global feature descriptor to recalibrate the feature maps of different streams and then aggregates them.
[0065] 1.1. “Fusion”
[0066] For fusion, consider configuring the fusion unit to do the following:
[0067] Receive feature information from two parallel convolutional streams carrying different scale information output by the first convolutional layer, and combine the multi-scale features by element-wise sum to obtain a combined feature L;
[0068] Apply global average pooling to the combined feature L in the spatial dimension to obtain channel statistics s;
[0069] The channel statistics s are processed through the channel downscaling convolution layer to generate a compact feature vector z;
[0070] The feature vector z is passed through two parallel channel-wise upscaling convolutional layers to obtain feature descriptors v1 and v2 respectively.
[0071] In this embodiment, the selective kernel feature fusion module receives inputs from two parallel convolutional streams carrying information of different scales, combines these multi-scale features with element-wise sum L = L1 + L2, and then performs L∈R H×W×C Apply global average GAP in the spatial dimension to calculate channel statistics s∈R H×W×C . Then a channel downscaling convolutional layer is used to generate a compact feature representation z∈R 1×1×r , the empirical value r = C / 8. Then the feature vector z passes through two parallel channel upscaling convolutional layers, one for each resolution stream, and obtains two feature descriptors v1 and v2, respectively. The dimensions of feature descriptors v1 and v2 are both 1×1×C.
[0072] 1.2. "Choice"
[0073] For selection, consider configuring the selection unit to do the following:
[0074] Apply the softmax function to the feature descriptors v1 and v2 respectively to obtain the attention values s1 and s2;
[0075] The obtained attention values s1 and s2 are used to calibrate the multi-scale features to obtain the fusion features of multi-resolution information.
[0076] In this embodiment, consider applying the softmax function to the feature descriptors v1 and v2 respectively to activate the attention values s1 and s2, and then use them to adaptively recalibrate the multi-scale feature maps L1 and L2 respectively. The overall process of feature recalibration and aggregation can be defined as U=s1·L1+s2·L2, where U is used to characterize the fused features, L1 is used to characterize the feature information of the channel where s1 is located, and L2 is used to characterize the feature information of the channel where s2 is located. In this way, the selective kernel feature fusion method uses about 5 times fewer parameters than the splicing aggregation method, but the results show that the selective kernel feature fusion method is better.
[0077] 2. For residual context blocks
[0078] When the selective kernel feature fusion module SKFF fuses information across multiple resolution branches, it is also necessary to extract useful information from the feature tensor through a refinement mechanism. Therefore, it is considered to extract features in the convolutional flow through a residual context block RCB to allow more informative features to be further passed while suppressing less useful features.
[0079] Specifically, Figure 4 As shown, the residual context block may include a grouped convolution module, a context module, a third convolution layer, and an aggregation module;
[0080] The grouped convolution module is used to process the fused features through two 3×3 grouped convolution layers to obtain grouped convolution features;
[0081] The context module is used to generate attention aggregation features by grouping convolutional features;
[0082] The third convolutional layer is used to process the attention aggregation features;
[0083] The aggregation module is used to aggregate the output and fusion features of the third convolutional layer to obtain contextual aggregation features that retain high-resolution spatial details.
[0084] Specifically, the context module is configured to perform the following operations:
[0085] Group convolution feature F b ∈R H×W×C Apply a 1×1 convolution layer, reshape and softmax to generate a new feature Fc ∈R 1×1×HW ; Where H and W are the height and width of the feature map respectively;
[0086] The grouped convolution features F b ∈R H×W×C Reshape to F b ∈R 1×HW×C , and with F c ∈R 1×1×HW Perform matrix multiplication to obtain the global feature descriptor F d ∈R 1×1×C ; Where C is the number of channels;
[0087] The global feature descriptor F d ∈R 1×1×C Through two 1×1 convolutional layers, we get the new attention feature F e ∈R 1 ×1×C ;
[0088] The new attention feature F e ∈R 1×1×C Aggregate to group convolutional features F b ∈R H×W×C at each position.
[0089] The aggregation module is specifically configured to perform aggregation processing using the following calculation formula:
[0090] F RCB =F a +ω(CM(F b ))
[0091] Among them, F RCB It is used to represent the contextual aggregation features that retain high-resolution spatial details, F a It is used to represent the fusion features, ω is used to represent the third convolutional layer, and CM is used to represent the context module.
[0092] In this embodiment, F b ∈R H×W×C It means that by inputting feature F at the beginning of RCB a The feature map obtained by applying two 3×3 grouped convolutional layers. These grouped convolutions are more resource-efficient than standard convolutions and are able to learn unique representations in each filter group. ω represents the last convolutional layer with a filter size of 1×1. CM stands for context module, which consists of three parts: 1) Context modeling: From the original feature map F b ∈R H×W×C To begin, we first generate new features F by applying a 1×1 convolution followed by a reshape and softmax operation. c ∈R 1×1×HW, where the reshaping method can be achieved by rotating, deforming, flipping and other data enhancement methods. Further, F b ∈R H×W×C Reshape to R 1×HW×C , and with F c ∈R 1×1×HW Perform matrix multiplication to obtain the global feature descriptor F d ∈R 1×1×C ; 2) Feature transformation: In order to capture the dependencies between channels, the feature descriptor F d ∈R 1×1×C Through two 1×1 convolutional layers, we get the new attention feature F e ∈R 1×1×C ; 3) Feature fusion: add the elements one by one and add the attention feature value F e ∈R 1×1×C Aggregate to the original grouped convolutional feature F b ∈R H ×W×C At each position of . Finally, after passing through the output of the third convolutional layer, it is combined with the fusion feature F a Perform aggregation.
[0093] In addition, the size of the image blocks that the network is trained on needs to be balanced with training speed and test accuracy. In terms of large image blocks, convolutional neural networks can capture fine image details, providing better results, but the training speed is relatively slow. In terms of small image blocks, training is faster but at the expense of reduced accuracy. In order to strike a proper balance between training speed and accuracy, this scheme considers training the network with smaller image blocks in the early stages of training and gradually using larger image blocks in the later stages of training. In this way, the network sequentially transitions from learning simple tasks to learning more complex tasks. The use of a progressive learning strategy on mixed-size image blocks not only improves the training speed, but also enhances the performance of the model at test time.
[0094] In summary, the processing method for realizing rapid image restoration and enhancement provided by the present invention has at least the following advantages:
[0095] Beneficial effects:
[0096] (1) Maintaining the original high-resolution features along the network layers, reducing the loss of precise spatial details;
[0097] (2) The model processes features with lower spatial resolution through parallel convolutional streams and can encode multi-scale contextual information;
[0098] (3) The multi-resolution parallel branch complements the main high-resolution branch and provides a more accurate and context-rich feature representation for image restoration;
[0099] (4) In terms of aggregating contextual information, a selective kernel fusion approach is adopted, and a set of useful kernels is dynamically selected through a self-attention mechanism, while retaining their unique complementary properties when fusing features.
[0100] The present specification also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed in a computer, the computer is caused to execute a method in any one of the embodiments in the specification.
[0101] The present specification also provides a computing device, including a memory and a processor, wherein executable codes are stored in the memory, and when the processor executes the executable codes, a method in any embodiment of the present specification is implemented.
[0102] Since the device embodiments provided by the present invention are based on the same inventive concept as the method embodiments of this specification, the specific contents can be found in the description of the method embodiments of this specification and will not be repeated here.
[0103] The modules or units in the device of the embodiment of the present invention can be combined, divided and deleted according to actual needs. The above disclosure is only the preferred embodiment of the present invention, and of course it cannot be used to limit the scope of the rights of the present invention. Those skilled in the art can understand that all or part of the processes of the above embodiment are implemented, and the equivalent changes made according to the claims of the present invention still fall within the scope of the invention.
Claims
1. A processing method for realizing rapid image restoration and enhancement, characterized in that: include: Constructing an image restoration and enhancement model; wherein the image restoration and enhancement model includes a first convolution layer, at least one recursive residual group RRG, a second convolution layer and a fusion module; the first convolution layer is used to extract underlying features of the original image; each of the recursive residual groups RRG includes at least one multi-scale residual block MRB, which is used to integrate high-resolution features by transferring context information from a low-resolution stream for the extracted underlying features, so as to restore both context information and spatial detail information; the second convolution layer is used to extract features of the deep features output by the recursive residual group RRG to obtain a residual image; the fusion module is used to fuse the residual image with the original image and output a restored image; Obtaining the image to be restored; The image to be restored is input into the image restoration enhancement model, and a restored image is output.
2. The method for realizing rapid image restoration and enhancement according to claim 1, characterized in that: Each multi-scale residual block MRB includes a selective kernel feature fusion module and a residual context block; The selective kernel feature fusion module is used to dynamically adjust the receptive field through a fusion unit and a selection unit; wherein the fusion unit is used to generate a global feature descriptor by combining information from multi-resolution streams, and the selection unit is used to recalibrate feature maps of different streams using the global feature descriptor and aggregate them to obtain a fused feature; The residual context block is used to extract features from the fused features output by the selective kernel feature fusion module, so as to transfer informative features and suppress useless features.
3. The method for realizing rapid image restoration and enhancement according to claim 2, characterized in that: The fusion unit is configured to perform the following operations: Receive feature information from two parallel convolutional streams carrying different scale information output by the first convolutional layer, and combine the multi-scale features by element-wise sum to obtain a combined feature L; Applying global average pooling to the combined feature L in the spatial dimension to obtain channel statistical information s; Processing the channel statistics s through a channel downscaling convolution layer to generate a compact feature vector z; The feature vector z is passed through two parallel channel-upscale convolutional layers to obtain feature descriptors v1 and v2 respectively.
4. The method for realizing rapid image restoration and enhancement according to claim 3, characterized in that: The selection unit is configured to perform the following operations: Apply the softmax function to the feature descriptors v1 and v2 respectively to obtain the attention values s1 and s2; The obtained attention values s1 and s2 are used to calibrate the multi-scale features to obtain the fusion features of multi-resolution information.
5. The method for realizing rapid image restoration and enhancement according to claim 4, characterized in that: The obtained attention values s1 and s2 are used to calibrate the multi-scale features. include: The fusion features are obtained using the following calculation formula: U=s1·L1+s2·L2 Among them, U is used to characterize the fusion feature, L1 is used to characterize the feature information of the channel where s1 is located, and L2 is used to characterize the feature information of the channel where s2 is located.
6. The method for realizing rapid image restoration and enhancement according to claim 2, characterized in that: The residual context block includes a grouped convolution module, a context module, a third convolution layer and an aggregation module; The grouped convolution module is used to process the fused features through two 3×3 grouped convolution layers to obtain grouped convolution features; The context module is used to generate attention aggregation features through the grouped convolutional features; The third convolutional layer is used to process the attention aggregation feature; The aggregation module is used to aggregate the output of the third convolutional layer and the fusion features to obtain contextual aggregation features that retain high-resolution spatial details.
7. The method for realizing rapid image restoration and enhancement according to claim 6, characterized in that: The context module is configured to perform the following operations: The grouped convolution feature F b ∈R H×W×C Apply a 1×1 convolution layer, reshape and softmax to generate a new feature F c ∈R 1×1×HW ; Where H and W are the height and width of the feature map respectively; The grouped convolutional features F b ∈R H×W×C Reshape to F b ∈R 1×HW×C , and with F c ∈R 1×1×HW Perform matrix multiplication to obtain the global feature descriptor F d ∈R 1×1×C ; Where C is the number of channels; The global feature descriptor F d ∈R 1×1×C Through two 1×1 convolutional layers, we get the new attention feature F e ∈R 1×1×C ; The new attention feature F e ∈R 1×1×C Aggregate to group convolutional features F b ∈R H×W×C at each position.
8. The method for realizing rapid image restoration and enhancement according to claim 7, characterized in that: The aggregation module is specifically configured to perform aggregation processing using the following calculation formula: F RCB =F a +ω(CM(F b )) Among them, F RCB It is used to characterize the contextual aggregation features that retain high-resolution spatial details, F a is used to characterize the fusion feature, ω is used to characterize the third convolutional layer, and CM is used to characterize the context module.
9. The method for realizing rapid image restoration and enhancement according to claim 1, characterized in that: When constructing the image restoration and enhancement model, the following loss function is used to optimize the model: Among them, L is used to represent the loss function, A~ is used to represent the image after the model is restored, and A * It is used to characterize the basic real image, and ε is the regularization term.
10. The method for realizing rapid image restoration and enhancement according to claim 1, characterized in that: When training the image restoration enhancement model, smaller image blocks are used for training in the early stage, and gradually larger image blocks are used for training in the later stage.