A super-resolution reconstruction method for extremely deep images based on non-residual structure

Abstract

The present invention discloses a non-residual structured ultra-deep image super-resolution reconstruction network model and method, the overall structure of the network includes three parts: a network head, a network backbone, and a network tail; the network head is used to extract shallow features of an input low-resolution image; the network backbone is composed of 16 feature channel parallel groups connected in series by adaptive weights and then connected in parallel with each other, the image feature tensor inferred by the 16 feature channel parallel groups will be integrated through global adaptive feature fusion, and the integrated information will pass through two ordinary convolutional layers with different kernels in sequence; the network tail is used to aggregate the shallow features and deep features extracted by the network, and integrate them back into the three RGB channels to obtain the final amplified super-resolution image. The network model of the present invention uses channel parallel connection without the identity mapping of residual learning to achieve jump connections between layers, solving the problem of network model design being too large and too deep and difficult to train in deep learning algorithms.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and deep learning technology, and in particular to a technology for establishing a corresponding high-resolution image based on a low-resolution image. Background Art

[0002] Image super-resolution (SR) reconstruction refers to the reconstruction of low-resolution images to corresponding high-resolution without changing the imaging hardware through machine learning and signal processing. Single image super-resolution is the mainstream direction of super-resolution research. Since super-resolution reconstruction is an underdetermined infinite solution problem in mathematics, traditional methods such as interpolation, sparse expression, iterative back-projection, etc. are difficult to reconstruct the high-frequency detail information of the image.

[0003] With the development of deep learning in the field of computer vision, single image super-resolution methods based on deep learning are currently the main methods in this field. From the initial convolution-based neural network algorithm to new network structures, such as deep residual networks, recursive neural networks, dense link networks and generative adversarial networks, they all have corresponding applications in this field and show very good performance, making the reconstructed high-resolution images more realistic and the details clearer.

[0004] In order to make the deep learning model have better performance, one of the current mainstream methods is to continuously increase the depth of the neural network and the number of feature channels. However, increasing the depth of the neural network and the number of feature channels will expand the size of the neural network, which will bring great difficulties to the training of the model. The model will experience gradient vanishing and gradient explosion during the training process, and the inability of the model to converge will lead to the failure of the final training. In addition, overly large network models also have the phenomenon of degradation of representation learning ability. Residual Learning is currently the most widely used method to reduce the difficulty of training. The image super-resolution method based on enhanced deep residual learning network (EDSR) is a classic image super-resolution method based on residual learning. For example Figure 1 As shown in the figure: This method introduces the method of identity jump mapping to linearly superimpose the image feature tensor that has been inferred by the nonlinear layer and the image feature tensor that has not been inferred by the nonlinear layer, which greatly promotes the transmission of feature information in the deep neural network model. However, the introduction of identity mapping in residual learning brings a lot of representation information redundancy, weakens the representation learning ability of the deep learning model, and makes the network obtain lower generalization under the same number of parameters. Therefore, directly introducing residual learning in deep deep learning models cannot fully develop the potential of the model.

[0005] In deep learning, the feature channels in the neural network model have a significant impact on the performance of the model. In order to enhance the representation learning ability of the deep learning model under residual learning and to improve the limitations of residual learning, many deep learning models use a combination of residual learning and feature channel parallelism. For example, the image super-resolution method based on residual dense network (RDN), the image super-resolution method based on memory network (MemNet), the image super-resolution method based on serial residual learning network (CARN), etc. This type of method stabilizes the performance of the model through residual learning, while strengthening the ability to share information between feature channels of different network layers. Compared with the model based on residual learning alone, the performance of this type of method has been greatly improved.

[0006] Although the combination of residual learning and feature channel concatenation has improved the model's representation learning ability to a certain extent, the identity mapping still exists, so the model's representation learning ability is still limited. In addition, this method directly connects different neural network layers in parallel, ignoring the fact that different neural network layers have different contribution capabilities to the model. As a result, the model based on this method cannot adaptively adjust the weights of the parallel connection between different layers, thereby limiting the generalization ability of the model. Summary of the invention

[0007] In response to the above-mentioned problems existing in the existing deep learning model that combines residual learning with parallel connection of feature channels, the present invention provides a non-residual structured ultra-deep image super-resolution reconstruction network model, which is a deep learning network model that does not contain residual learning at all.

[0008] The non-residual structured ultra-deep image super-resolution reconstruction network model provided by the present invention has an overall network structure including three parts: a network head, a network trunk, and a network tail. The overall expression of the network is:

[0009] Y=F t (F b (F h (X)))

[0010] In the formula, F h (·) indicates the network header, F b (·) represents the backbone of the network, F t (·) represents the tail of the network, X represents the input low-resolution RGB image, and Y represents the output high-resolution RGB image.

[0011] The network head is used to extract shallow features of the input low-resolution image. The network head is a common convolutional layer with a kernel of 3×3, 3 input channels, and 32 output channels. The expression of the network head is:

[0012] X 0 =F h (X) = C 3×3 (X)

[0013] Where, X 0 Represents the image feature tensor output by the network head; F h (·) indicates the network header; C 3×3 (·) denotes a 3×3 convolutional layer; X denotes the input low-resolution RGB image.

[0014] The network backbone is composed of 16 feature channel parallel groups connected in series and then connected in parallel by adaptive weights. The image feature tensor inferred by the 16 feature channel parallel groups will be integrated by global adaptive feature fusion, and the integrated information will pass through a common convolution layer with a kernel of 1×1, 32 input channels, and 32 output channels, and a common convolution layer with a kernel of 3×3, 32 input channels, and 32 output channels.

[0015] The expression of the network backbone is:

[0016] X n+1 =F b (X 0 )=C 3×3 (C 1×1 (X c ))

[0017] Where, X n+1 Represents the features output by the backbone of the network; F b (·) indicates the backbone of the network; X 0 represents the image feature tensor output by the network head. Since the backbone contains 16 feature extraction modules (feature channel parallel groups), n represents the total number of feature extraction modules, so n = 16; X c represents the global image feature tensor aggregated by the global adaptive feature fusion method, C 3×3 (·) represents a 3×3 convolutional layer, C 1×1 (·) denotes a 1×1 convolutional layer.

[0018] The expression of the characteristic channel parallel group is:

[0019] X′ i =C gi (X i-1 )=C gi (C gi-1 (...C g1 (X 1 )...))

[0020] X i =C1×1 (X′ i )

[0021] In the formula, X′ i Cg represents the image feature tensor before the 1×1 convolution layer and after feature extraction by the feature channel parallel group. i (·) represents the nth feature channel parallel group. Since the network model has a total of 16 channel parallel groups, i = 1, 2..., 16; X i-1 Cg represents the image feature tensor after inference by the i-1th feature channel parallel group; i-1 (·) represents the i-1th parallel group of feature channels; Cg 1 (·) represents the first characteristic channel parallel group; C 1×1 (·) is a 1×1 convolutional layer, X i Represents the image feature tensor after inference by the parallel group of the i-th feature channel.

[0022] The expression of the global adaptive feature fusion is:

[0023] X c =G(X 0 , ...X i , ..., X n )=[Λ 0 X 0 , ..., Λ i X i , ..., Λ n X n ]

[0024] In the formula, Λ i represents the adaptive weight parameter corresponding to the image feature tensor after the i-th feature channel parallel group inference, X i represents the image feature tensor after inference of the i-th feature channel parallel group, where i represents the index of the position. Since the network model has a total of 16 channel parallel groups, i = 1, 2, ..., 16; G(·) represents the global adaptive feature fusion method, X c represents the global image feature tensor aggregated by the global adaptive feature fusion method, [·] represents the parallel operation of channels; X 0 Represents the image feature tensor output by the network head, X n represents the image feature tensor after the nth (last) feature channel parallel group inference. Since the network model has a total of 16 channel parallel groups, n = 16; Λ 0 Represents the adaptive weight parameter corresponding to the image feature tensor output by the network head, Λ nRepresents the adaptive weight parameter corresponding to the image feature tensor after the nth (last) feature channel parallel group inference.

[0025] The image feature tensor integrated by global adaptive feature fusion will also undergo global correction and adjustment of the aggregated features through two layers of convolutional layers. 1×1 (·) represents the first 1×1 convolutional layer, C 3×3 (·) represents the second 3×3 convolutional layer, X n+1 Represents the final output part of the network backbone, then the expression of the final output part of the network backbone is:

[0026] X n+1 =C 3×3 (C 1×1 (X c ))

[0027] In the formula, C 1×1 (·) represents the first 1×1 convolutional layer, C 3×3 (·) represents the second 3×3 convolutional layer, X n+1 Represents the final output part of the network backbone, X c It represents the global image feature tensor aggregated by the global adaptive feature fusion method. Since the network model has a total of 16 channel parallel groups, n=16.

[0028] The feature parallel module is the most basic unit in the network backbone, so the expression of the feature channel parallel module is:

[0029] W(X m，i-1 )=C 3×3 (R(C 3×3 (X m，i-1 ))

[0030] x m，i =C 1×1 ([β 1 X m，i-1 , β 2 W(x m，i-1 )])

[0031] Where W(·) represents nonlinear mapping; X m，i-1 represents the image feature tensor input through the i-1th feature channel parallel module in the mth feature channel parallel group. Since the network model has a total of 16 channel parallel groups, each channel parallel group contains 8 channel parallel modules, so m = 1, 2, ..., 16; i = 1, 2, ..., 8; C 3×3 (·) represents a 3×3 convolutional layer, R(·) represents a linear rectification function; X m，i represents the image feature tensor input through the i-th feature channel parallel module in the m-th feature channel parallel group; C1×1 (·) indicates a 1×1 convolutional layer; [·] indicates a channel parallel operation; β 1 represents the adaptive weight of the jump mapping part; β 2 Represents the adaptive weight of the wide channel activation part.

[0032] The end of the network is the image reconstruction part: it consists of a sub-pixel convolution layer with a kernel of 3×3, 32 input channels, and 32 output channels, and a normal convolution layer with a kernel of 3×3, 32 input channels, and 3 output channels. Its function is to aggregate the shallow and deep features extracted by the network and integrate them back into the three RGB channels to obtain the final enlarged super-resolution image. Let C sub (·) represents the sub-pixel convolution layer, Y represents the final super-resolution RGB three-channel image, and the expression at the end of the network is:

[0033] Y=C 3×3 (C sub (X n+1 ))

[0034] Where Y represents the final super-resolution RGB three-channel image; C sub (·) represents the sub-pixel convolution layer; C 3×3 (·) indicates a 3×3 convolutional layer; X n+1 Represents the final output part of the network backbone; since the network model has a total of 16 channel parallel groups, n=16.

[0035] Compared with the prior art, the present invention is beneficial in that:

[0036] The present invention is a deep learning network model that does not contain residual learning at all. It is based on a neural network with full feature channels in parallel. Through a relatively simple idea, it uses channel parallelization without identity mapping to achieve jump connections between layers. The performance is better than residual learning, and it provides a new solution to solve the problem that the model is too large and too deep in design and difficult to train in the deep learning algorithm. Moreover, the channel parallelization between layers of this method contains adaptive weight parameters, which can make the network adaptively weigh the contribution of different layers to the representation ability, so that the representation learning ability of the model constructed based on this method is no longer limited even in very deep cases (up to 400 layers), so that the representation learning ability of the model itself can be fully explored. The image super-resolution reconstruction method with full feature channels in parallel is based on a deep neural network with full feature channels in parallel. Its essence is a convolutional neural network that does not contain identity mapping in residual learning at all.

[0037] The deep network model image super-resolution method built based on this method not only has a significant performance improvement compared to the image super-resolution method with residual learning of the same scale, with a peak signal-to-noise ratio (PSNR) of about 0.2db; but also has a significant reduction in the number of parameters compared to the model with the same performance, containing only 60% of the parameters of similar methods. This type of method can also be generalized to other computer vision tasks, such as target detection, image recognition, semantic segmentation, etc., and can provide an important reference for improving the representation learning ability and full development model performance of deep learning models for such tasks.

[0038] Other advantages, objectives and features of the present invention will be embodied in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 , a detailed overview of the modules of traditional residual learning identity mapping.

[0040] Figure 2 , the overall structure diagram of the algorithm model.

[0041] Figure 3 , detailed overview of the structure in the model layer.

[0042] Figure 4 , channel characteristic parallel group details.

[0043] Figure 5 , channel feature parallel module details. DETAILED DESCRIPTION

[0044] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0045] like Figure 2 As shown, the non-residual structured ultra-deep image super-resolution reconstruction network model provided by the present invention is based on a deep neural network with all feature channels in parallel, and its essence is a convolutional neural network that completely does not contain the identity mapping in residual learning.

[0046] The overall structure of the network is shown in the figure Figure 2 The specific model layer is shown in Figure 3 As shown in Figure 1, the input image of the network will pass through the head, trunk, and tail of the network in sequence. In order to more accurately represent the results of the network, let the head of the network be F h (·), the main cadres are divided into F b (·), the network tail is F t(·), the input low-resolution RGB image is X, the output high-resolution RGB image is Y, and the overall expression of the network is:

[0047] Y=F t (F b (F h (X)))

[0048] In the formula, F h (·) indicates the network header, F b (·) represents the backbone of the network, F t (·) represents the tail of the network, X represents the input low-resolution RGB image, and Y represents the output high-resolution RGB image.

[0049] 1. Network Head

[0050] The head of the network is the shallow feature extraction part: the network head is a common convolution layer with a kernel of 3×3, 3 input channels, and 32 output channels. The expression of the network head is:

[0051] X 0 =F h (X) = C 3×3 (X)

[0052] In the formula, X 0 Represents the image feature tensor output by the network head; F h (·) indicates the network header; C 3×3 (·) denotes a 3×3 convolutional layer; X denotes the input low-resolution RGB image.

[0053] 2. Network backbone

[0054] The backbone of the network is the most important part: it is composed of 16 feature channel parallel groups connected in series and then connected in parallel through adaptive weights. The image feature tensor inferred by the 16 feature channel parallel groups will be integrated by global adaptive feature fusion, and the integrated information will pass through a normal convolution layer with a kernel of 1×1, 32 input channels, and 32 output channels, and a normal convolution layer with a kernel of 3×3, 32 input channels, and 32 output channels.

[0055] The expression of the network backbone is:

[0056] X n+1 =F b (X 0 )=C 3×3 (C 1×1 (X c ))

[0057] In the formula, X n+1Represents the features output by the backbone of the network; F b (·) indicates the backbone of the network; X 0 represents the image feature tensor output by the network head. Since the backbone contains 16 feature extraction modules (feature channel parallel groups), n represents the total number of feature extraction modules, so n = 16; X c represents the global image feature tensor aggregated by the global adaptive feature fusion method, C 3×3 (·) represents a 3×3 convolutional layer, C 1×1 (·) denotes a 1×1 convolutional layer.

[0058] The expression of the characteristic channel parallel group is:

[0059] X′ i =C gi (X i-1 )=Cg i (Cg i-1 (...Cg 1 (X 1 )...))

[0060] X i =C 1×1 (X′ i )

[0061] In the formula, X′ i Cg represents the image feature tensor before the 1×1 convolution layer and after feature extraction by the feature channel parallel group. i (·) represents the nth feature channel parallel group. Since the network model has a total of 16 channel parallel groups, i = 1, 2..., 16; X i-1 Cg represents the image feature tensor after inference by the i-1th feature channel parallel group; i-1 (·) represents the i-1th parallel group of feature channels; Cg 1 (·) represents the first characteristic channel parallel group; C 1×1 (·) is a 1×1 convolutional layer, X i Represents the image feature tensor after inference by the parallel group of the i-th feature channel.

[0062] The expression of the global adaptive feature fusion is:

[0063] X c =G(X 0 , ...X i , ..., X n )=[Λ 0 X 0 , ..., Λ i X i , ..., Λ n Xn ]

[0064] In the formula, Λ i represents the adaptive weight parameter corresponding to the image feature tensor after the i-th feature channel parallel group inference, X i represents the image feature tensor after inference of the i-th feature channel parallel group, where i represents the index of the position. Since the network model has a total of 16 channel parallel groups, i = 1, 2, ..., 16; G(·) represents the global adaptive feature fusion method, X c represents the global image feature tensor aggregated by the global adaptive feature fusion method, [·] represents the parallel operation of channels; X 0 Represents the image feature tensor output by the network head, X n represents the image feature tensor after the nth (last) feature channel parallel group inference. Since the network model has a total of 16 channel parallel groups, n = 16; Λ 0 Represents the adaptive weight parameter corresponding to the image feature tensor output by the network head, Λ n Represents the adaptive weight parameter corresponding to the image feature tensor after the nth (last) feature channel parallel group inference.

[0065] The image feature tensor integrated by global adaptive feature fusion will also undergo global correction and adjustment of the aggregated features through two layers of convolutional layers. 1×1 (·) represents the first 1×1 convolutional layer, C 3 ×3 (·) represents the second 3×3 convolutional layer, X n+1 Represents the final output part of the network backbone, then the expression of the final output part of the network backbone is:

[0066] X n+1 =C 3×3 (C 1×1 (X c ))

[0067] In the formula, C 1×1 (·) represents the first 1×1 convolutional layer, C 3×3 (·) represents the second 3×3 convolutional layer, X n+1 Represents the final output part of the network backbone, X c It represents the global image feature tensor aggregated by the global adaptive feature fusion method. Since the network model has a total of 16 channel parallel groups, n=16.

[0068] The feature parallel module is the most basic unit in the network backbone, so the expression of the feature channel parallel module is:

[0069] W(X m，i-1)=C 3×3 (R(C 3×3 (X m，i-1 ))

[0070] X m，i =C 1×1 ([β 1 x m，i-1 , β 2 W(X m，i-1 )])

[0071] Where W(·) represents nonlinear mapping; X m，i-1 represents the image feature tensor input through the i-1th feature channel parallel module in the mth feature channel parallel group. Since the network model has a total of 16 channel parallel groups, each channel parallel group contains 8 channel parallel modules, so m = 1, 2, ..., 16; i = 1, 2, ..., 8; C 3×3 (·) represents a 3×3 convolutional layer, R(·) represents a linear rectification function; X m，i represents the image feature tensor input through the i-th feature channel parallel module in the m-th feature channel parallel group; C 1×1 (·) indicates a 1×1 convolutional layer; [·] indicates a channel parallel operation; β 1 represents the adaptive weight of the jump mapping part; β 2 Represents the adaptive weight of the wide channel activation part.

[0072] 3. Network tail

[0073] The end of the network is the image reconstruction part: it consists of a sub-pixel convolution layer with a kernel of 3×3, 32 input channels, and 32 output channels, and a normal convolution layer with a kernel of 3×3, 32 input channels, and 3 output channels. Its function is to aggregate the shallow and deep features extracted by the network and integrate them back into the three RGB channels to obtain the final enlarged super-resolution image. Let C sub (·) represents the sub-pixel convolution layer, Y represents the final super-resolution RGB three-channel image, and the expression at the end of the network is:

[0074] Y=C 3×3 (C sub (X n+1 ))

[0075] Where Y represents the final super-resolution RGB three-channel image; C sub (·) represents the sub-pixel convolution layer; C 3×3 (·) indicates a 3×3 convolutional layer; X n+1 Represents the final output part of the network backbone; since the network model has a total of 16 channel parallel groups, n=16.

[0076] (1) Comparative analysis of the image super-resolution reconstruction method of the present invention and other existing methods:

[0077] The image super-resolution reconstruction method of the present invention is compared with other deep learning super-resolution algorithms of the same scale in terms of parameters, and the comparison results are shown in Table 1. Among them, EDSR is an image super-resolution method based on an enhanced deep residual learning network (reference Lim B, Son S, Kim H, et al. Enhanced deep residual networks for single image super-resolution [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition workshops. 2017: 136-144.). RDN is an image super-resolution method based on a residual dense connection network (reference Zhang Y, Tian Y, Kong Y, et al. Residual dense network for image super-resolution [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 2472-2481.). RCAN is an image super-resolution method based on deep residual learning channel attention (reference Zhang Y, Li K, Li K, et al. Image super-resolution using very deep residual channel attention networks [C] / / Proceedings of the European conference on computer vision (ECCV). 2018: 286-301.). SAN is an image super-resolution method based on a second-order attention network (reference Dai T, Cai J, Zhang Y, et al. Second-order attention network for single image super-resolution [C] / / Proceedings of the IEEE / CVF conference on computer vision and pattern recognition. 2019: 11065-11074).

[0078] Table 1 Comparison of parameters between the method of the present invention and deep learning super-resolution algorithms of the same scale

[0079]

[0080] As shown in Table 1, compared with image super-resolution networks with residual learning of the same network size and effect, the image super-resolution network with all feature channels in parallel is significantly ahead in terms of parameter quantity. Compared with the classic method using residual learning: image super-resolution method based on reinforced deep residual learning network, the image super-resolution reconstruction method with all feature channels in parallel reduces the number of parameters by about 30 million.

[0081] (2) Table 2 is a comparison of the effects of the method of the present invention and the deep learning super-resolution algorithm of the same scale. The indicators in the table are peak signal-to-noise ratio (db) / image result similarity. Among them, EDSR, RDN, SAN, and RCAN are four existing methods. For detailed explanations, see the previous paragraph. The bold parts in the table represent the best results under the same multiples and similar benchmark data sets.

[0082] Table 2. Comparison of the effects of the method of the present invention and deep learning super-resolution algorithms of the same scale

[0083]

[0084] As shown in Table 2, compared with the image super-resolution network with residual learning of the same network scale and effect, the performance of the image super-resolution reconstruction method based on the full feature channel parallel network is about 0.2db ahead of the same network in terms of image peak signal-to-noise ratio.

[0085] This invention was supported by the National Natural Science Foundation of China Youth Science Fund Project 62102330, the Sichuan Natural Science Foundation Youth Science Fund Project 2022NSFSC0947, and the Central Universities Basic Research Business Expenses Special Fund 2682021CX040.

[0086] The above description is only a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. A very deep image super-resolution reconstruction method based on non-residual structure, characterized in that: The overall structure of the adopted network model includes three parts: network head, network backbone, and network tail; The network head is used to extract shallow features of the input low-resolution image; The network backbone is composed of 16 feature channel parallel groups connected in series and then connected in parallel by adaptive weights; the image feature tensor inferred by the 16 feature channel parallel groups will be integrated by global adaptive feature fusion, and the integrated information will pass through a common convolution layer with a kernel of 1×1, 32 input channels, and 32 output channels, and a common convolution layer with a kernel of 3×3, 32 input channels, and 32 output channels in sequence; The basic unit of the network backbone is the characteristic parallel module, and the expression of the characteristic channel parallel module is: W(X m，i-1 )=C 3×3 (R(C 3×3 (X m，i-1 )) X m，i =C 1×1 ([β1X m，i-1 ,β2W(X m，i-1 )]) Where W(·) represents the nonlinear mapping part; X m，i-1 represents the image feature tensor input through the i-1th feature channel parallel module in the mth feature channel parallel group. Since each channel parallel group contains 8 channel parallel modules, there are 16 channel parallel groups in total, so m = 1, 2, ..., 16; i = 1, 2, ..., 8; X m，i represents the image feature tensor input through the i-th feature channel parallel module in the m-th feature channel parallel group; C 3×3 (·) represents a 3×3 convolutional layer, R(·) represents a linear rectification function; C 1×1 (·) represents a 1×1 convolutional layer; [·] represents channel parallel operation; β1 represents the adaptive weight of the skip mapping part; β2 represents the adaptive weight of the wide channel activation part; The tail of the network is used to aggregate the shallow features and deep features extracted by the network, and integrate them back into the three RGB channels to obtain the final enlarged super-resolution image.

2. The method for super-resolution reconstruction of extremely deep images based on non-residual structure according to claim 1, characterized in that: The overall expression of the model network is: Y=F t (F b (F h (X))) In the formula, F h (·) indicates the network header, F b (·) represents the network backbone, F t (·) represents the tail of the network, X represents the input low-resolution RGB image, and Y represents the output high-resolution RGB image.

3. The method for ultra-deep image super-resolution reconstruction based on non-residual structure according to claim 1, characterized in that: The network head is a common convolutional layer with a kernel of 3×3, 3 input channels, and 32 output channels. The expression of the network head is: X0=F h (X)=C 3×3 (X) In the formula, X represents the input low-resolution RGB image; X0 represents the image feature tensor output by the network head; F h (·) indicates the network header; C 3×3 (·) indicates a 3×3 convolutional layer.

4. The method for super-resolution reconstruction of extremely deep images based on non-residual structure according to claim 1, characterized in that: The expression of the network backbone is: X n+1 =F b (X0)=C 3×3 (C 1×1 (X c )) Where, X n+1 Represents the features output by the backbone of the network; F b (·) represents the backbone of the network; X0 represents the image feature tensor output by the network head. Since the backbone contains 16 feature extraction modules, i.e., feature channel parallel groups, n represents the total number of feature extraction modules, so n = 16; X c represents the global image feature tensor aggregated by the global adaptive feature fusion method, C 3×3 (·) represents a 3×3 convolutional layer, C 1×1 (·) indicates a 1×1 convolutional layer.

5. The method for super-resolution reconstruction of extremely deep images based on non-residual structure according to claim 1, characterized in that: The expression of the characteristic channel parallel group is: X′ i =Cg i (X i-1 )=Cg i (Cg i-1 (...Cg1(X1)...)) X i =C 1×1 (X′ i ) In the formula, X′ i Cg represents the image feature tensor before the 1×1 convolution layer and after feature extraction by the feature channel parallel group. i (·) represents the i-th feature channel parallel group, where i represents the index of the position. Since the network model has a total of 16 channel parallel groups, i = 1, 2, ..., 16; X i-1 Cg represents the image feature tensor after inference by the i-1th feature channel parallel group; i-1 (·) represents the i-1th parallel group of characteristic channels; Cg1(·) represents the first parallel group of characteristic channels; C 1×1 (·) represents a 1×1 convolutional layer, X i Represents the image feature tensor after the final inference by the i-th feature channel parallel group.

6. The method for ultra-deep image super-resolution reconstruction based on non-residual structure according to claim 1, characterized in that: The expression of the global adaptive feature fusion is: X c =G(X0,...,X i ,...,X n )=[Λ0X0,...,Λ i X i ,...,L n X n ] In the formula, Λ i represents the adaptive weight parameter corresponding to the image feature tensor after the i-th feature channel parallel group inference, X i represents the image feature tensor after inference of the i-th feature channel parallel group, where i represents the index of the position. Since the network model has a total of 16 channel parallel groups, i = 1, 2, ..., 16; G(·) represents the global adaptive feature fusion method, X c represents the global image feature tensor aggregated by the global adaptive feature fusion method, [·] represents the parallel operation of channels; X0 represents the image feature tensor output by the network head, X n represents the image feature tensor after inference through the nth feature channel parallel group. Since the network model has a total of 16 channel parallel groups, n = 16; Λ0 represents the adaptive weight parameter corresponding to the image feature tensor output by the network head, Λ n Represents the adaptive weight parameter corresponding to the image feature tensor after the nth feature channel parallel group inference.

7. The method for ultra-deep image super-resolution reconstruction based on non-residual structure according to claim 1, characterized in that: The expression of the final output part of the network backbone is: X n+1 =C 3×3 (C 1×1 (X c )) In the formula, C 1×1 (·) represents the first 1×1 convolutional layer, C 3×3 (·) represents the second 3×3 convolutional layer, X n+1 Represents the final output part of the network backbone, X c Represents the global image feature tensor aggregated by the global adaptive feature fusion method; since the network model has a total of 16 channel parallel groups, n = 16.

8. The method for ultra-deep image super-resolution reconstruction based on non-residual structure according to claim 1, characterized in that: The expression of the network tail is: Y=C 3×3 (C sub (X n+1 )) Where Y represents the final super-resolution RGB three-channel image; C sub (·) represents the sub-pixel convolution layer; C 3×3 (·) indicates a 3×3 convolutional layer; X n+1 Represents the final output part of the network backbone; since the network model has a total of 16 channel parallel groups, n=16.

Images