Multi-scale image super-resolution reconstruction method based on profile wave knowledge guided network

Abstract

The present application relates to a kind of multi-scale image super-resolution reconstruction methods based on contour wave knowledge guide network, comprising: obtaining low-resolution image to be reconstructed;Low-resolution image is input into the contour wave knowledge guide network of training completion, and the corresponding super-resolution reconstruction image is output;Contour wave knowledge guide network, including the contourlet filter module, embedding subnetwork, prediction network and super-resolution reconstruction module that are cascaded in turn, wherein, low-resolution image is input into embedding subnetwork, then after passing through prediction network and super-resolution reconstruction module, corresponding super-resolution reconstruction image is obtained;Embedding subnetwork carries out feature extraction to the low-resolution image input, and the feature map obtained provides sufficient information for the prediction contour wave coefficient, prediction network carries out contour wave decomposition coefficient prediction to the feature map input, and obtains contour wave decomposition coefficient prediction value, and super-resolution reconstruction module carries out inverse transform to contour wave decomposition coefficient prediction value and obtains super-resolution reconstruction image.

Description

Technical Field

[0001] This invention belongs to the field of image super-resolution reconstruction technology, specifically relating to a multi-scale image super-resolution reconstruction method based on a contour wave knowledge-guided network. Background Technology

[0002] With the continuous development of technology, the demand for high-resolution images is becoming increasingly urgent. Since hardware-based resolution enhancement is resource-limited and expensive, algorithm-based image super-resolution reconstruction is generally employed. Single Image Super-Resolution (SISR) can recover details lost during image downsampling, resulting in clearer and more detailed images. Therefore, it is widely used in remote sensing satellite imaging, video surveillance, biomedicine, and other fields.

[0003] Traditional interpolation-based SISR methods include nearest neighbor interpolation, bilinear interpolation, cubic interpolation, and bicubic interpolation. These methods are fast and simple, but suffer from accuracy limitations. The emergence of deep learning has opened new avenues for SISR, using end-to-end learning to analyze the statistical relationship between low-resolution images and their corresponding high-resolution images, thereby achieving high-resolution image reconstruction. Representative methods include SRCNN (Super-Resolution Convolutional Neural Network) and FSRCNN (Fast Super-Resolution Convolutional Neural Network). While these methods have to some extent compensated for the shortcomings of traditional SISR methods and improved image clarity visually, they lack the ability to extract high-frequency features, resulting in mostly shallow features and certain accuracy limitations, which affect the reconstruction results to some extent. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention provides a multi-scale image super-resolution reconstruction method based on a contour wave knowledge-guided network. The technical problem to be solved by this invention is achieved through the following technical solution:

[0005] This invention provides a multi-scale image super-resolution reconstruction method based on contour wave knowledge-guided networks, comprising:

[0006] Step 1: Obtain the low-resolution image to be reconstructed;

[0007] Step 2: Input the low-resolution image into the trained contour wave knowledge-guided network, and output the corresponding super-resolution reconstructed image;

[0008] The contourlet knowledge-guided network includes a contourlet filter module, an embedding sub-network, a prediction network, and a super-resolution reconstruction module cascaded in sequence. The low-resolution image is input into the embedding sub-network, and then passes through the prediction network and the super-resolution reconstruction module to obtain the corresponding super-resolution reconstructed image.

[0009] The embedded sub-network extracts features from the input low-resolution image to obtain a feature map. The prediction network predicts the contour wave decomposition coefficients from the input feature map to obtain the predicted contour wave decomposition coefficient values. The super-resolution reconstruction module performs an inverse transformation on the predicted contour wave decomposition coefficient values ​​to obtain the super-resolution reconstructed image.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0011] 1. The multi-scale image super-resolution reconstruction method based on contourlet knowledge-guided network of the present invention combines contourlet feature decomposition in traditional signal processing with a deep learning framework, transforming image reconstruction into a learning problem of contourlet coefficients and their GGD distribution. The parameters of contourlet decomposition coefficients and GGD can effectively represent features sparsely, and multi-scale, multi-directional edge contours and directional textures in the image can be captured with only a small number of coefficients, thereby improving the reconstruction performance of super-resolution images.

[0012] 2. The multi-scale image super-resolution reconstruction method based on contourlet knowledge-guided network of the present invention extracts sparse features in a multi-scale and multi-directional manner through multi-scale contourlet filter modules and integrates them into the network, making full use of the advantages of contourlet feature learning in the frequency domain to ensure detailed reconstruction of super-resolution images.

[0013] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0014] Figure 1 This is a flowchart of a multi-scale image super-resolution reconstruction method based on a contour wave knowledge-guided network provided by an embodiment of the present invention;

[0015] Figure 2 This is a schematic diagram of the structure of a contour wave knowledge-guided network provided in an embodiment of the present invention;

[0016] Figure 3This is a schematic diagram of the 8x high-resolution image reconstruction results using different methods provided in the embodiments of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following describes in detail a multi-scale image super-resolution reconstruction method based on a contour wave knowledge-guided network proposed according to the present invention, in conjunction with the accompanying drawings and specific embodiments.

[0018] The foregoing and other technical contents, features, and effects of the present invention will be clearly presented in the following detailed description of specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and concrete understanding can be gained of the technical means and effects adopted by the present invention to achieve its intended purpose. However, the accompanying drawings are for reference and illustration only and are not intended to limit the technical solutions of the present invention.

[0019] Please see Figure 1 , Figure 1 This is a flowchart of a multi-scale image super-resolution reconstruction method based on a contour wave knowledge-guided network provided by an embodiment of the present invention. As shown in the figure, the multi-scale image super-resolution reconstruction method based on a contour wave knowledge-guided network in this embodiment includes:

[0020] Step 1: Obtain the low-resolution image to be reconstructed;

[0021] Step 2: Input the low-resolution image into the trained contour wave knowledge-guided network, and output the corresponding super-resolution reconstructed image.

[0022] In this embodiment, the contourlet knowledge-guided network includes a contourlet filter module, an embedded sub-network, a prediction network, and a super-resolution reconstruction module, which are cascaded in sequence.

[0023] The low-resolution image is input into the embedding sub-network, and then passes through the prediction network and the super-resolution reconstruction module to obtain the corresponding super-resolution reconstructed image. The embedding sub-network extracts features from the input low-resolution image to obtain a feature map, the prediction network predicts the contourlet decomposition coefficients from the input feature map to obtain the predicted contourlet decomposition coefficient values, and the super-resolution reconstruction module performs an inverse transform on the predicted contourlet decomposition coefficient values ​​to obtain the super-resolution reconstructed image.

[0024] Optionally, the multi-scale image super-resolution reconstruction method based on contour wave knowledge-guided network in this embodiment can be applied to the field of remote sensing satellite imaging. Accordingly, the low-resolution image to be reconstructed is a remote sensing satellite image.

[0025] Furthermore, combined Figure 2The schematic diagram of the contour wave knowledge-guided network shown below provides a detailed explanation of the training process of the contour wave knowledge-guided network.

[0026] In this embodiment, the training process of the contour wave knowledge-guided network includes:

[0027] Step ①: Acquire multiple high-resolution images to form a training set;

[0028] Step 2: Construct a contour wave knowledge guidance network;

[0029] Step 3: Input the training sample images from the training set into the contour wave knowledge-guided network and train it until the preset cutoff condition is reached, thus obtaining the trained contour wave knowledge-guided network.

[0030] During training, the contourlet filter module performs multi-level decomposition on the input training sample image to obtain contourlet decomposition feature maps at different scales and directions. The contourlet decomposition feature map of the low-pass subband is then used as a low-resolution image input to the embedding sub-network. After passing through the prediction network and the super-resolution reconstruction module, the super-resolution reconstructed image corresponding to the training sample image is obtained. Then, the network parameters of the embedding sub-network, prediction network, and super-resolution reconstruction module are learned and updated using the constructed loss function and the backpropagation algorithm.

[0031] Optionally, the contourlet filter module includes multi-stage filter units, each including a scaling filter and a directional filter. The scaling filter is used to decompose the input image to obtain the corresponding low-pass and high-pass components. The high-pass component is input to the directional filter of the current filter unit for directional subband decomposition, and the low-pass component is input to the scaling filter of the next stage filter unit for decomposition.

[0032] The parameters of the contourlet filter module can be set according to the reconstruction factor, including: the number of decomposition levels, the number of sub-bands to be decomposed, and the types of scaling filters and directional filters.

[0033] In this embodiment, the Laplace decomposition is set to "maxflat", and the directional filter is set to "dmaxflat7". Different decomposition levels can be used depending on the required reconstruction factor (2×, 4×, 8×). Generally, the larger the reconstruction factor, the larger the decomposition level.

[0034] In this embodiment, the contourlet filter module has a decomposition level of 3 and a directional filter of [2,2,2], which decomposes the contourlet coefficients. Size adjusted to M 2 ,in, l = 1, ..., L represents the decomposition level of the contourlet filter module, and k = 1, ..., K represents the number of directional subbands of the contourlet filter module.

[0035] In this embodiment, the input training sample image, i.e., the high-resolution image... Perform multi-level decomposition and iteratively use F LP (Scale Filter) and F DFB (Directional filter) for input Decomposition is performed to obtain the contour wave decomposition coefficients. The profile wave decomposition feature map of the low-pass subband obtained by decomposition is then used. The low-resolution image corresponding to the training sample image is used, where p is the reconstruction factor. Taking L=3 and K=12 as an example, the high-pass subband decomposition coefficients are obtained. and low-pass subband decomposition coefficient C low Therefore, the contour wave decomposition coefficients can be expressed as:

[0036]

[0037] Optionally, the embedding subnetwork includes multiple cascaded residual blocks, with the number of channels in each residual block increasing sequentially. The feature map output by each residual block is of the same size and is the same size as the input image of the embedding subnetwork.

[0038] In this embodiment, the embedded subnetwork can maintain the size of the feature map while increasing the number of channels. For the contourlet decomposition feature map of the input low-pass subband, a 3×3 convolution with a stride of 1 and padding of 1 is first performed, and then passed through different residual blocks in sequence to enhance the feature representation. The resulting feature map X is:

[0039]

[0040] In this context, the superscript and subscript of R represent the number of residual blocks and the number of output channels, respectively.

[0041] Optionally, the prediction network includes multiple parallel prediction subnetworks, with the number of prediction subnetworks being L+1, where L is the decomposition level of the contourlet filter module. Each prediction subnetwork performs channel dimensionality reduction and size transformation on the input feature map, and the resulting feature maps form the predicted values ​​of the contourlet decomposition coefficients. The feature maps are matched one-to-one with the channel dimensions and sizes of the contourlet decomposition coefficients of the contourlet filter module.

[0042] In this embodiment, the prediction network mainly learns the contourlet decomposition coefficients of the contourlet filter module. Through appropriate feature learning, each prediction sub-network generates a feature map that matches the channel dimension and size of the corresponding contourlet decomposition coefficient.

[0043] In this embodiment, feature learning for the input feature map X mainly consists of the following three operations:

[0044] (1) Increase the size of the feature map, the size of which is determined by the size factor s:

[0045] op increase =r2(b2(tc2(r1(b1(c1(X))))+X↑s))(3);

[0046] (2) Maintain the number of channels:

[0047] op retain =r2(c2(r1(c1(X))))+X(4);

[0048] (3) Reduce the number of channels:

[0049] op reduce =r2(b2(c2(r1(b1(c1(X)))))+X))(5);

[0050] Where c is convolution, r is rectified linear unit (ReLU), b is batch normalization, and tc is transposed convolution. In the formula, 1 and 2 are used to distinguish one entity or operation from another.

[0051] In this embodiment, the specific settings of the prediction subnetwork are as follows: Figure 2 As shown, in other embodiments, it is only necessary to ensure that each predictive subnetwork generates a feature map that matches the channel dimension and size of the corresponding contour wave decomposition coefficients. This can be achieved by combining the above three operations, without any specific limitations.

[0052] In this embodiment, through an independent prediction network S Net1 S Net2 S Net3 S Net4 The learned contourlet decomposition coefficients, i.e., the predicted values ​​of the contourlet decomposition coefficients, are obtained:

[0053] In this embodiment, during the training process of the contourlet knowledge-guided network, the super-resolution reconstruction module is also used to calculate the GGD (zero-mean generalized Gaussian distribution) parameter vector corresponding to the contourlet decomposition coefficients of the contourlet filter module and the GGD parameter vector corresponding to the predicted contourlet decomposition coefficients obtained by the prediction network.

[0054] Specifically, the zero-mean generalized Gaussian distribution (GGD) can be expressed as:

[0055]

[0056] Wherein, parameter α controls the shape of the GGD distribution, i.e., the rate of decay; parameter σ controls the variance of the GGD distribution; β is the scaling parameter; and Γ(·) is the gamma function.

[0057] Therefore, the calculated profile wave decomposition coefficients The corresponding GGD parameter vector f is:

[0058]

[0059] Similar to the GGD parameter vector f of the true profiled wave decomposition coefficients, the predicted values ​​of the profiled wave decomposition coefficients are calculated. Corresponding GGD parameter vector

[0060] In this embodiment, the network loss consists of four parts, including: spatial global loss l global Frequency domain contour wave loss contourlet Loss of detail retention detail and the loss of GGD parameters l ggd The loss function is expressed as:

[0061] L=νl global +l contourlet +μl detail +l ggd (8);

[0062] In the formula, ν is the weight of the spatial global loss and μ is the weight of the detail preservation loss.

[0063] Wherein, the spatial global loss is the mean squared error (MSE) between the super-resolution reconstructed image and the corresponding training sample image, expressed as:

[0064]

[0065] In the formula, Let I be the super-resolution reconstructed image, and let ||·|| be the training sample image. F Norm operations.

[0066] The frequency domain contourlet loss is the mean square error between the predicted contourlet decomposition coefficients obtained by the prediction network and the contourlet decomposition coefficients of the contourlet filter module. The frequency domain contourlet loss compensates for the high-frequency details ignored in the MSE loss, and it is expressed as:

[0067]

[0068] In the formula, L is the decomposition level of the contourlet filter module, and K is the number of directional subbands of the contourlet filter module. To predict the contour wave decomposition coefficients obtained by the prediction network, These are the contour wave decomposition coefficients of the contourlet filter module.

[0069] To better preserve texture details, the detail preservation loss effectively prevents detail degradation by ensuring that the high-frequency contour wave coefficients are not zero. This is expressed as:

[0070]

[0071] In the formula, γ is the first relaxation value and ε is the second relaxation value;

[0072] To further maintain training stability while ensuring the distribution of contour wave coefficients, a GGD parameter loss based on cosine similarity is proposed, which is expressed as:

[0073]

[0074] In the formula, C similarity ∈[-1,1] represents the cosine similarity, and f is the GGD parameter vector corresponding to the true contourlet coefficients. To predict the GGD parameter vector corresponding to the contourlet coefficients, n is the vector f and The dimension of.

[0075] In this embodiment, the model is trained by minimizing the loss function and using stochastic gradient descent (SGD) until a preset number of training iterations are reached, resulting in a trained contour wave knowledge-guided network.

[0076] The multi-scale image super-resolution reconstruction method based on contourlet knowledge-guided network in this invention combines contourlet feature decomposition in traditional signal processing with a deep learning framework. It transforms image reconstruction into a learning problem of contourlet coefficients and their GGD distribution. The parameters of contourlet decomposition coefficients and GGD can effectively represent features sparsely. It can capture multi-scale and multi-directional edge contours and directional textures in the image with only a small number of coefficients, thereby improving the reconstruction performance of super-resolution images.

[0077] Secondly, the multi-scale image super-resolution reconstruction method based on contourlet knowledge-guided network in this invention extracts sparse features in a multi-scale, multi-directional manner through multi-scale contourlet filter modules and integrates them into the network. This fully leverages the advantages of contourlet feature learning in the frequency domain to ensure detailed reconstruction of the super-resolution image. Furthermore, this contourlet knowledge-guided network addresses the issue of "smoothing" output from reconvolutional neural networks in reconstruction tasks and preserves as much structural information as possible during contourlet decomposition coefficient estimation, thereby effectively capturing edge and texture details of the image and achieving high reconstruction performance on remote sensing scene datasets.

[0078] Furthermore, simulation experiments are conducted to illustrate the effectiveness of the multi-scale image super-resolution reconstruction method based on contour wave knowledge-guided network in this embodiment.

[0079] Four remote sensing experimental datasets were set up. For the UC Merced dataset, 50% of the images were randomly selected as the training set, and the remainder were used for testing. For the AID, NWPU45, and WHU-RS19 datasets, 80% of the images were randomly selected as the training set, and the remainder were used for testing. The final experimental results were averaged across 10 randomized test results.

[0080] In terms of algorithm selection, in order to ensure the fairness of the comparison as much as possible, the comparison algorithms include traditional graph bicubic interpolation, generative networks SRCNN and FSRCNN based on sparse coding, image reconstruction algorithm VDSR based on residual networks, CTN based on multi-layer features and context information, and LGCNet and DCM that combine local and global features.

[0081] The ×8 reconstruction results of the method of this invention and other comparative methods on remote sensing datasets are shown in Table 1.

[0082] Table 1. ×8 reconstruction results of the present invention and other comparative methods on remote sensing datasets.

[0083]

[0084] As shown in Table 1, at an 8x reconstruction scale, the method of this invention outperforms other comparative methods on two objective metrics on the UC MERCED, NWPU45, and AID datasets. Its performance on the WHU-RS19 dataset is slightly worse than the DCM method. Overall, the image reconstruction quality of this invention is good. This is attributed to the fact that the contour wave knowledge-guided network can effectively utilize the multi-scale and multi-resolution characteristics of contours to fully extract detailed features from the image, thereby improving the model's reconstruction performance.

[0085] Please see Figure 3The diagram illustrates the 8x high-resolution image reconstruction results of different methods provided in this embodiment of the invention. From top to bottom, the results are from the UC MERCED, NWPU45, AID, and WHU-RS19 datasets. It can be seen that the images generated by Bicubic, SRCNN, and FSRCNN are generally blurry, with some blurring and jagged edges at the target edges. LGCNet and VDSR's reconstruction results are slightly better than the above three methods, but they exhibit severe artifacts and minor imperfections around their edges. The results generated by DCM and CTN methods are relatively clear and sharp, but still have some imperfections at the edges. This invention generates rich high-frequency details, which has a significant advantage in terms of visual perception, resulting in clearer and sharper edges and reconstruction results closer to the real image. Overall, the method of this invention achieves good reconstruction results in both objective analysis and visual perception.

[0086] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0087] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A multi-scale image super-resolution reconstruction method based on contour wave knowledge-guided networks, characterized in that, include: Step 1: Obtain the low-resolution image to be reconstructed; Step 2: Input the low-resolution image into the trained contour wave knowledge-guided network, and output the corresponding super-resolution reconstructed image; The contourlet knowledge-guided network includes a contourlet filter module, an embedding sub-network, a prediction network, and a super-resolution reconstruction module cascaded in sequence. The low-resolution image is input into the embedding sub-network, and then passes through the prediction network and the super-resolution reconstruction module to obtain the corresponding super-resolution reconstructed image. The embedded sub-network extracts features from the input low-resolution image to obtain a feature map. The prediction network predicts the contour wave decomposition coefficients from the input feature map to obtain the predicted contour wave decomposition coefficient values. The super-resolution reconstruction module performs an inverse transformation on the predicted contour wave decomposition coefficient values ​​to obtain the super-resolution reconstructed image.

2. The multi-scale image super-resolution reconstruction method based on contour wave knowledge-guided network according to claim 1, characterized in that, The training process of the contour wave knowledge-guided network includes: Step ①: Acquire multiple high-resolution images to form a training set; Step 2: Construct a contour wave knowledge guidance network; Step 3: Input the training sample images from the training set into the contour wave knowledge guidance network to train it until the preset cutoff condition is reached, and obtain the trained contour wave knowledge guidance network. During training, the contourlet filter module performs multi-level decomposition on the input training sample image to obtain contourlet decomposition feature maps at different scales and directions. The contourlet decomposition feature map of the low-pass subband is then input as a low-resolution image into the embedding sub-network. After passing through the prediction network and the super-resolution reconstruction module, the super-resolution reconstructed image corresponding to the training sample image is obtained. The network parameters of the embedding sub-network, the prediction network, and the super-resolution reconstruction module are learned and updated using the constructed loss function and the backpropagation algorithm.

3. The multi-scale image super-resolution reconstruction method based on contour wave knowledge-guided network according to claim 2, characterized in that, The contourlet filter module includes a multi-stage filter unit, which includes a scale filter and a directional filter. The scaling filter is used to decompose the input image to obtain the corresponding low-pass and high-pass components. The high-pass component is input to the directional filter of the current filter unit for directional subband decomposition, and the low-pass component is input to the scaling filter of the next stage filter unit for decomposition.

4. The multi-scale image super-resolution reconstruction method based on contour wave knowledge-guided network according to claim 2, characterized in that, The embedded sub-network includes multiple cascaded residual blocks, the number of channels of the multiple residual blocks increases sequentially, and the feature map output by each residual block has the same size and is the same size as the input image of the embedded sub-network.

5. The multi-scale image super-resolution reconstruction method based on contour wave knowledge-guided network according to claim 2, characterized in that, The prediction network includes multiple parallel prediction subnetworks, and the number of prediction subnetworks is L+1, where L is the decomposition level of the contourlet filter module. Each prediction subnetwork performs channel dimensionality reduction and size transformation on the input feature map, and the resulting feature map forms the predicted value of the contourlet decomposition coefficient. The feature map corresponds one-to-one with the channel dimension and size of the contourlet decomposition coefficient of the contourlet filter module.

6. The multi-scale image super-resolution reconstruction method based on contour wave knowledge-guided network according to claim 5, characterized in that, During the training process of the contour wave knowledge-guided network, the super-resolution reconstruction module is also used to calculate the GGD parameter vector corresponding to the contour wave decomposition coefficients of the contourlet filter module and the GGD parameter vector corresponding to the predicted contour wave decomposition coefficients obtained by the prediction network.

7. The multi-scale image super-resolution reconstruction method based on contour wave knowledge-guided network according to claim 6, characterized in that, The loss function is expressed as: L=νl global +l contourlet +μl detail +l ggd ; In the formula, l global For airspace global loss, l contourlet For frequency domain contour wave loss, l detail To preserve details, l ggd ν represents the GGD parameter loss, μ represents the weight of the spatial global loss, and μ represents the weight of the detail preservation loss.

8. The multi-scale image super-resolution reconstruction method based on contour wave knowledge-guided network according to claim 7, characterized in that, The spatial global loss is the mean square error between the super-resolution reconstructed image and the corresponding training sample image, expressed as: In the formula, Let I be the super-resolution reconstructed image, and let ||·|| be the training sample image. F Norm operations; The frequency domain contour wave loss is the mean square error between the predicted contour wave decomposition coefficients obtained by the prediction network and the contour wave decomposition coefficients of the contourlet filter module, expressed as: In the formula, L is the decomposition level of the contourlet filter module, and K is the number of directional subbands of the contourlet filter module. To predict the contour wave decomposition coefficients obtained by the prediction network, These are the contour wave decomposition coefficients of the contourlet filter module.

9. The multi-scale image super-resolution reconstruction method based on contour wave knowledge-guided network according to claim 8, characterized in that, The detail retention loss is expressed as: In the formula, γ is the first relaxation value and ε is the second relaxation value; The GGD parameter loss is expressed as: l ggd ＝1-C similarity In the formula, C similarity ∈[-1,1] represents the cosine similarity, and f is the GGD parameter vector corresponding to the true contourlet coefficients. To predict the GGD parameter vector corresponding to the contourlet coefficients, n is the vector f and The dimension of.

Images