A face super-resolution method and system based on a convenient alternating projection network

Abstract

The present invention provides a face super-resolution method and system based on a convenient alternating projection network. Combined with a convenient alternating projection unit, the network enables each high-resolution spatial feature to efficiently interact and fuse across levels, thereby improving the network's feature expression capabilities. The convenient alternating projection network proposed in the present invention achieves performance and visual quality similar to that of a non-compact model, and significantly reduces parameters and testing time, thereby achieving the function of improving the reconstruction efficiency and performance of super-resolution face images. The present invention solves the technical problems of the current non-compact face super-resolution reconstruction algorithm, such as the existence of additional redundant features, low algorithm efficiency, and certain limitations in the reconstruction performance of lightweight networks. The network efficiency and performance proposed in the present invention are superior to other latest face image super-resolution algorithms, and the generated high-resolution face images are of higher quality.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision facial image super-resolution, and in particular relates to a facial super-resolution method and system based on a convenient alternating projection network. Background Art

[0002] Facial images are widely used in everyday life, such as in public surveillance videos, social applications, and personal photos. However, when face detection, verification, and parsing are used as input for facial image understanding, the performance of each task degrades significantly. Super-resolution (SR) of facial images is a simple approach to address this problem. It works by restoring low-resolution (LR) facial images of the same scene and converting them into potential high-resolution (HR) facial images.

[0003] Existing face super-resolution methods based on convolutional neural networks (CNNs) have achieved significant improvements over traditional methods due to their superior computational power. This is because deeper and wider networks based on deep learning can provide powerful feature extraction capabilities. However, due to the limited performance of computing devices in practical applications, large and complex neural networks are not suitable for face super-resolution scenarios. Replacing high-performance computing equipment with these devices can significantly increase the overall system cost, raising operating costs. Therefore, balancing the model size and operational efficiency of super-resolution networks is particularly important.

[0004] Recently, Dong et al. pioneered the application of convolutional neural networks to image super-resolution (Single Image Super-Resolution Using Deep Learning, SRCNN), which consists of only three convolutional layers to learn nonlinear relationships between images. Using concepts based on residual learning networks, Kim et al. proposed a very deep SR network (VDSR) with 20 convolutional layers. Lim et al. improved the deep residual SR network (EDSR), expanding the model size to more than 60 convolutional layers and removing unnecessary modules from traditional residual networks. Zhang et al. proposed a residual dense network (RDN) with over 100 layers, fully utilizing the hierarchical features of LR images and effectively improving the visual quality of imaging. Subsequently, Zhang et al. utilized the residual in the residual structure and introduced a channel attention mechanism, forming a very deep residual channel attention network (RCAN) with more than 400 layers. Niu et al. introduced the Holistic Attention Network (HAN), which greatly enhanced the feature correlation between layers. Wang et al. used multi-view texture compensation reconstruction and introduced a texture attention mechanism to enhance visual performance while reducing the overall depth of the network. Chen et al. introduced the Spatial Attention Residual Network (SPARNet). The spatial attention mechanism avoids focusing on areas with poor features and improves the overall training efficiency of the network. Wang et al. proposed a Sparse Mask Super-Resolution (SMSR) network to learn sparse masks, studying image sparsity to improve the inference efficiency of image sparse networks and reduce redundant computations in the network.

[0005] While the aforementioned non-compact networks achieve satisfactory results, these deep network models incur significant computational overhead and generate redundant features, often limiting their performance on mobile and edge devices. To improve efficiency, lightweight reconstruction networks often ignore the overall contours and fine textures of the face, resulting in limited feature representation. Therefore, minimizing redundant features and leveraging convenient alternating projection methods to enhance the efficiency and performance of face SR reconstruction are crucial. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a face super-resolution method and system based on a convenient alternating projection network, which is used to improve the reconstruction efficiency and performance of super-resolution of face images.

[0007] The technical solution adopted by the present invention to solve the above technical problems is: a face super-resolution method based on a convenient alternating projection network, comprising the following steps:

[0008] S0: Construct a face super-resolution system based on a convenient alternating projection network, including a downsampling module, a coarse feature extractor module, a fine feature extractor module, a feature fusion module and a face image reconstruction module;

[0009] S1: downsample the high-resolution face image to the target low-resolution face image;

[0010] S2: Perform a separable convolution operation on the target low-resolution face image and use a coarse feature extractor to extract a coarse facial feature map;

[0011] S3: Construct a convenient alternating projection unit as a fine feature extractor, input the rough facial feature map into the convenient alternating projection unit, and obtain the fine facial feature map by alternating multiple upward-downward projection operations;

[0012] S4: Let concat be the feature fusion operation, Y i t The high-resolution feature output of each upward projection unit is fused through dense connections across levels. The fused face fine feature map is:

[0013] concat(Y i 1 ,Y i 2 ,...,Y i t );

[0014] S5: Perform convolution operation on the fused facial fine feature map to reconstruct the facial feature map into the target high-resolution facial image; let f represent the reconstruction operation, which is implemented by ordinary convolution, then the target high-resolution facial image Y i sr Expressed as:

[0015] Y i sr =f[Y i 1 ,Y i 2 ,...,Y i t ].

[0016] According to the above scheme, in step S2, the specific steps are:

[0017] Use three convolutional layers to build a rough feature extractor, and use the rough feature extractor to extract a rough facial feature map; Represents a rough facial feature map, Conv(·) represents a normal convolution operation, DeConv(·) represents a Depthwise convolution operation, and PConv(·) represents a point convolution operation. Represents the input low-resolution face image, and the rough facial feature map is expressed as:

[0018]

[0019] According to the above solution, in step S3, the convenient alternating projection unit includes an upward-downward projection combination consisting of a plurality of upward projection units and downward projection units;

[0020] The upward projection unit uses operations including deconvolution, point convolution, convenience connection and ordinary convolution to obtain low-resolution residual features, and then deconvolution is performed on the low-resolution residual features and projected back to the high-resolution space to obtain high-resolution spatial fusion features;

[0021] The downward projection unit projects the high-resolution spatial fusion features to the low-resolution spatial learning residual features to obtain low-resolution spatial features by performing operations including point convolution, convenience connection and ordinary convolution.

[0022] According to the above solution, in step S3, the specific steps of the upward projection operation are:

[0023] The rough facial features Enter the convenient alternate projection unit;

[0024] The rough facial features Deconvolution obtains high-resolution spatial features:

[0025]

[0026] Y i t1 Perform point convolution operation, and then use convenient connection operation Φ and ordinary convolution respectively to fuse the features extracted by the two and then perform ordinary convolution to obtain low-resolution features:

[0027]

[0028] Will and Subtract to obtain low-resolution residual features

[0029] The residual features are then amplified through a deconvolution operation The output is:

[0030]

[0031] The high-resolution spatial feature Y i t1 With Y i t2 Add to get high-resolution features

[0032] Furthermore, in step S3, the convenient link operation Φ is used to perform a convenient linear transformation operation on each inherent feature, and according to L ij =Φ i,j (PConv(Y i t )) function performs channel-by-channel mapping and conveniently connects features through multiple categories Get the entire n=m·t ordinary convolution Conv(Y i t )=[L 11 ,L 12 ,...,L mt ] to quickly obtain the required n feature maps.

[0033] According to the above solution, in step S3, the specific steps of the downward projection operation are:

[0034] For high-resolution feature Y i t Perform point convolution operation, and then use convenient connection operation Φ and ordinary convolution respectively to fuse the features extracted by the two and then perform ordinary convolution to obtain low-resolution features:

[0035]

[0036] Then it will Deconvolution obtains high-resolution features:

[0037]

[0038] The feature Y i t3 With Y i t Subtract to obtain high-resolution residual features

[0039] For high-resolution residual features Perform point convolution again, fuse the features extracted by the convenient connection operation Φ with those extracted by ordinary convolution, and perform ordinary convolution to obtain low-resolution features:

[0040]

[0041] The final low-resolution output features are:

[0042]

[0043] According to the above scheme, the following steps are also included:

[0044] Let N be the number of face images in the dataset, H(·) be the entire network, and Y i lr and Y i hr Denotes the i-th low-resolution face image and the i-th high-resolution face image in the dataset, then the loss function L(θ) of the entire network is expressed as:

[0045]

[0046] A face super-resolution system based on a convenient alternating projection network, comprising a downsampling module, a coarse feature extractor module, a fine feature extractor module, a feature fusion module and a face image reconstruction module;

[0047] The downsampling module is used to downsample the high-resolution face image to the target low-resolution face image;

[0048] The coarse feature extractor module is used to perform ordinary convolution and depth-separable convolution operations on the target low-resolution image to extract a coarse facial feature map;

[0049] The fine feature extractor module is used to construct a convenient alternating projection network as a fine feature extractor. The rough facial feature map is input into the convenient alternating projection network to obtain a fine facial feature map. The convenient alternating projection network includes several upward-downward projection combinations. Each unit includes several point convolutions, deconvolutions, convenient connection operations and ordinary convolutions. The high-resolution features generated by each unit are fused through dense connections, allowing features at each stage to interact across levels, realizing the network's powerful feature expression capabilities.

[0050] The feature fusion module is used to perform feature fusion on the fine facial feature maps obtained by each projection unit;

[0051] The face image reconstruction module is used to perform ordinary convolution on the fused face feature map to reconstruct the target high-resolution face image.

[0052] A computer storage medium stores a computer program executable by a computer processor, wherein the computer program executes a face super-resolution method based on a convenient alternating projection network.

[0053] The beneficial effects of the present invention are:

[0054] 1. The present invention provides a face super-resolution method and system based on a convenient alternating projection network. Combined with a convenient alternating projection unit, the network enables efficient cross-layer interactive fusion of high-resolution spatial features, thereby improving the network's feature expression capabilities. The convenient alternating projection network proposed in the present invention achieves performance and visual quality similar to non-compact models, significantly reduces parameters and testing time, and realizes the function of improving the reconstruction efficiency and performance of super-resolution facial images.

[0055] 2. The present invention solves the technical problems that the current non-compact face super-resolution reconstruction algorithm has additional redundant features, low algorithm efficiency, and certain limitations in lightweight network reconstruction performance.

[0056] 3. The network efficiency and performance proposed in this invention are superior to other latest face image super-resolution algorithms, and the generated high-resolution face images are of higher quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a flow chart of an embodiment of the present invention.

[0058] Figure 2 It is a network structure diagram of an embodiment of the present invention.

[0059] Figure 3 2 is a system diagram of an embodiment of the present invention.

[0060] Figure 4 2 is a comparison chart of test results of an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0062] See also Figure 2 and Figure 3 , the embodiment of the present invention includes a downsampling module 101, a coarse feature extractor module 102, a fine feature extractor module 103, a feature fusion module 104 and a face image reconstruction module 105;

[0063] The downsampling module 101 is used to downsample the high-resolution face image to a target low-resolution face image;

[0064] The coarse feature extractor module 102 is used to perform ordinary convolution and depth-separable convolution operations on the target low-resolution image to extract a coarse facial feature map;

[0065] The fine feature extractor module 103 is used to construct a convenient alternating projection network as a fine feature extractor. The rough facial feature map is input into the convenient alternating projection network to obtain a fine facial feature map. The convenient alternating projection network includes several upward-downward projection combinations. Each unit includes several point convolutions, deconvolutions, convenient connection operations and ordinary convolutions. The high-resolution features generated by each unit are fused through dense connections, so that the features of each stage interact across levels, realizing the powerful feature expression capability of the network.

[0066] The feature fusion module 104 is used to perform feature fusion on the fine facial feature maps obtained by each projection unit;

[0067] The facial image reconstruction module 105 is used to perform ordinary convolution on the fused facial feature map to reconstruct the target high-resolution facial image.

[0068] The specific implementation of each module can refer to the description of the above method embodiment, and the embodiment of the present invention will not be repeated.

[0069] See also Figure 1 , an embodiment of the present invention includes the following steps:

[0070] S0: Construct a face super-resolution system based on a convenient alternating projection network, including a downsampling module, a coarse feature extractor module, a fine feature extractor module, a feature fusion module and a face image reconstruction module;

[0071] S1: downsample the high-resolution face image to the target low-resolution face image;

[0072] In the embodiment of the present invention, bicubic interpolation is used to downsample the high-resolution face image to a target low-resolution face image.

[0073] In this embodiment, the FFHQ face dataset is preferably used as the training, validation, and test sets. It contains 1000 images as the training dataset, 10 images as the validation dataset, and 40 images as the test dataset. The images in the dataset are 256×256 pixels in size. This invention uses a bicubic degradation model to downsample the dataset with a downsampling factor of 4, resulting in low-resolution images of 64×64 pixels.

[0074] S2: Perform a separable convolution operation on the target low-resolution face image and use a coarse feature extractor to extract a coarse facial feature map. The specific steps are as follows:

[0075] A coarse feature extractor is constructed using three convolutional layers. In the embodiment of the present invention, a 3*3 normal convolution, a 3*3 depthwise convolution and a 1*1 point convolution operation are used to construct the coarse feature extractor, and the coarse feature extractor is used to extract a coarse facial feature map. Represents a rough facial feature map, Conv(·) represents a 3*3 normal convolution operation, DeConv(·) represents a 3*3 Depthwise convolution operation, and PConv(·) represents a 1*1 point convolution operation. Represents the input low-resolution face image, and the rough facial feature map is expressed as:

[0076]

[0077] S3: Construct a Convenient Alternate Projection Unit (CAPU) as a fine feature extractor. The rough facial feature map is input into the fine facial feature map obtained by the CAPU. This is used to reduce feature redundancy in the network and enhance cross-layer interaction of facial features at each stage.

[0078] The convenient alternating projection unit includes an upward-downward projection combination consisting of a plurality of upward projection units and downward projection units;

[0079] The upward projection unit uses operations including deconvolution, point convolution, convenience connection and ordinary convolution to obtain low-resolution residual features, deconvolution the low-resolution residual features and re-project them into the high-resolution space to obtain high-resolution spatial fusion features; the specific steps are:

[0080] The rough facial features Input to the convenient alternating projection unit, in the upward projection unit, first the rough facial feature map Use 8*8 size deconvolution operation to obtain high-resolution spatial features: Y i t1 Perform point convolution operation, and then use convenient connection operation Φ and ordinary convolution respectively to fuse the features extracted by the two and then perform ordinary convolution to obtain low-resolution features: Will and Subtract to obtain low-resolution residual features The residual features are then amplified through a deconvolution operation The output is: The high-resolution spatial feature Y i t1 With Y i t2 Add to get high-resolution features

[0081] The convenient link operation Φ is to perform a convenient linear transformation operation on each inherent feature and ij =Φ i,j (PConv(Y i t )) function performs channel-by-channel mapping, and multiple types of convenient connection features The entire n=m·t ordinary convolution Conv(Y i t )=[L 11 ,L 12 ,...,L mt ] to quickly obtain the required n feature maps.

[0082] The down-projection unit projects the high-resolution spatial fusion features onto the low-resolution spatial learning residual features by performing operations including point convolution, convenience connection and ordinary convolution to obtain low-resolution spatial features. The specific steps are as follows:

[0083] High-resolution feature Y i t It is then input into the downward projection unit for point convolution, and then the features extracted by the convenient connection operation Φ and ordinary convolution are fused and then ordinary convolution is performed to obtain low-resolution features: Then it will Deconvolution obtains high-resolution features: The feature Y i t3 With Y i t Subtract to obtain high-resolution residual features Perform the point convolution operation again, and through the convenient connection operation Φ and ordinary convolution, fuse the features extracted by the two and perform ordinary convolution to obtain low-resolution features: The final low-resolution output features are

[0084] Alternately iterate multiple upward-downward projection steps to extract fine facial feature maps;

[0085] S4: Let concat be the feature fusion operation, Y i t The high-resolution feature output of each upward projection unit is fused through dense connections across levels. The fused face fine feature map is:

[0086] concat(Y i 1 ,Y i 2 ,...,Y i t );

[0087] S5: Perform convolution operation on the fused facial fine feature map to reconstruct the facial feature map into the target high-resolution facial image; let f represent the reconstruction operation, which is implemented by 3*3 ordinary convolution, then the target high-resolution facial image Y i sr Expressed as:

[0088] Y i sr =f[Y i 1 ,Y i 2 ,...,Y i t ];

[0089] Let N be the number of face images in the dataset, H(·) be the entire network, and Y i lr and Y i hr Denotes the i-th low-resolution face image and the i-th high-resolution face image in the dataset, then the loss function L(θ) of the entire network is expressed as:

[0090]

[0091] Finally, the present invention provides a test embodiment. The present invention uses the FFHQ face dataset as the training set, validation set and test set, which contains 1000 images as the training dataset, 10 images as the validation dataset, and 40 images as the test dataset. The image size in the dataset is 256×256 pixels. The present invention uses a bicubic degradation model to downsample the dataset, where the downsampling factor is 4, so the size of the downsampled low-resolution image is 64×64 pixels. It should be noted that all training, validation and testing are based on the luminance channel in the YCbCr color space, and a 4-fold magnification factor is used for training and testing. The SR reconstruction results are evaluated using three evaluation indicators: Peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) and feature similarity index (FSIM) to test the performance of SR reconstruction under the luminance channel. In each iteration of training, an LR block of size 16×16 is extracted as input. The model of the present invention is trained by Adaptive Moment Estimation (Adam) optimizer, β 1 =0.9,β 2 = 0.999. The initial learning rate is set to 10 -4 , and then reduced to one tenth of the original value every 100 cycles. The model of the present invention was trained using the Pytorch platform and a GTX2080 graphics card. Table 1 shows the comparison results under the condition of a reconstruction multiple of 4 using the above three evaluation indicators. Figure 4 Comparison of the reconstruction results of 4x face images, where (a) represents the Bicubic result, (b) represents the EDSR result, (c) represents the RDN result, (d) represents the HAN result, (e) represents the MTC result, (f) represents the SPAR result, (g) represents the SMSR result, (h) represents the result of the present invention, and (i) represents the original high-resolution face reference image.

[0092] The face SR methods selected for comparison include: Bicubic, EDSR, RDN, DBPN, HAN, MTC, SPAR, and SMSR. Bicubic is a classic image interpolation algorithm; EDSR is a classic SR method based on deep residual networks; RDN is a deep residual densely connected network; HAN is an SR network based on overall attention; MTC is a face SR method based on multi-view texture compensation; SPAR is a face SR method based on spatial attention; SMSR is a lightweight SR method that constructs a sparse mask network by studying image sparsity. Figure 4(i) is the original high-resolution face reference image, and (h) is the high-resolution face image generated by the present invention. It can be seen that the present invention achieves a very high visual effect in the visual results.

[0093] Table 1 Comparison results between the present invention and seven excellent algorithms

[0094] method Bicubic EDSR RDN HAN MTC SPAR SMSR The present invention PSNR / dB 29.80 32.00 31.99 32.17 31.98 32.21 32.22 32.33 SSIM 0.8459 0.8854 0.8873 0.8896 0.8889 0.8896 0.8916 0.8929 FSIM 0.8886 0.9280 0.9249 0.9293 0.9278 0.9293 0.9283 0.9295

[0095] It can be seen from the experimental results in the above table that the present invention has achieved obvious advantages compared with the other seven methods.

[0096] Any portion not described in the specification is prior art or common knowledge. This embodiment is intended only to illustrate the invention and is not intended to limit the scope of the invention. Any modifications such as equivalent substitutions made by those skilled in the art to the invention are deemed to fall within the scope of protection of the claims.

[0097] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, or two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0098] The above embodiments are intended only to illustrate the design concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The scope of protection of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design concepts disclosed in the present invention are within the scope of protection of the present invention.

Claims

1. A face super-resolution method based on a convenient alternating projection network, characterized by: The following steps are involved: S0: Construct a face super-resolution system based on a convenient alternating projection network, including a downsampling module, a coarse feature extractor module, a fine feature extractor module, a feature fusion module and a face image reconstruction module; S1: downsample the high-resolution face image to the target low-resolution face image; S2: Perform a separable convolution operation on the target low-resolution face image and use a coarse feature extractor to extract a coarse facial feature map; S3: Construct a convenient alternating projection unit as a fine feature extractor, input the rough facial feature map into the convenient alternating projection unit, and obtain the fine facial feature map by alternating multiple upward-downward projection operations; The convenient alternating projection unit includes an upward-downward projection combination consisting of a plurality of upward projection units and downward projection units; The upward projection unit uses operations including deconvolution, point convolution, convenience connection and ordinary convolution to obtain low-resolution residual features, and then deconvolution is performed on the low-resolution residual features and projected back to the high-resolution space to obtain high-resolution spatial fusion features; The down-projection unit projects the high-resolution spatial fusion features onto the low-resolution spatial learning residual features to obtain low-resolution spatial features by performing operations including point convolution, convenience connection and ordinary convolution; The specific steps of the upward projection operation are: The rough facial features Enter the convenient alternate projection unit; The rough facial features Deconvolution obtains high-resolution spatial features: ； right Perform point convolution operation, and then perform convenient connection operation After fusing the features extracted by the two and performing ordinary convolution, the low-resolution features are obtained: ； Will and Subtract to obtain low-resolution residual features ; The residual features are then amplified through a deconvolution operation The output is: ； High-resolution spatial features and Add to get high-resolution features ; Convenient link operation Used to perform convenient linear transformation operations on each intrinsic feature and according to Function performs channel-by-channel mapping and conveniently connects features through multiple categories Get the whole Ordinary convolution , to quickly get what you need feature maps; The specific steps of the downward projection operation are: For high-resolution features Perform point convolution operation, and then perform convenient connection operation After fusing the features extracted by the two and performing ordinary convolution, the low-resolution features are obtained: ； Then it will Deconvolution obtains high-resolution features: ； The features and Subtract to obtain high-resolution residual features ; For high-resolution residual features Perform the point convolution operation again and conveniently connect the operation The low-resolution features are obtained by fusing them with the features extracted by ordinary convolution and performing ordinary convolution: ； The final low-resolution output features are: ； S4: Set is the feature fusion operation, The high-resolution feature output of each upward projection unit is fused through dense connections across levels. The fused face fine feature map is: ； S5: Perform convolution operation on the fused facial fine feature map to reconstruct the facial feature map into the target high-resolution facial image; Represents the reconstruction operation, which is implemented using ordinary convolution, so the target high-resolution face image Expressed as: 。 2. The face super-resolution method based on a convenient alternating projection network according to claim 1, characterized in that: In the step S2, the specific steps are: Use three convolutional layers to build a rough feature extractor, and use the rough feature extractor to extract a rough facial feature map; represents a rough facial feature map, represents the normal convolution operation, express Depthwise Convolution operation, represents the point convolution operation, Represents the input low-resolution face image, and the rough facial feature map is expressed as: 。 3. The face super-resolution method based on a convenient alternating projection network according to claim 1, characterized in that: The following steps are also included: set up Indicates the number of face images in the dataset, Represents the entire network, and Indicates the first low-resolution face image and the A high-resolution face image, then the loss function of the entire network Expressed as: 。 4. A system for the face super-resolution method based on a convenient alternating projection network according to any one of claims 1 to 3, characterized in that: It includes a downsampling module, a coarse feature extractor module, a fine feature extractor module, a feature fusion module and a face image reconstruction module; The downsampling module is used to downsample the high-resolution face image to the target low-resolution face image; The coarse feature extractor module is used to perform ordinary convolution and depth-separable convolution operations on the target low-resolution image to extract a coarse facial feature map; The fine feature extractor module is used to construct a convenient alternating projection network as a fine feature extractor. The rough facial feature map is input into the convenient alternating projection network to obtain a fine facial feature map. The convenient alternating projection network includes several upward-downward projection combinations. Each unit includes several point convolutions, deconvolutions, convenient connection operations and ordinary convolutions. The high-resolution features generated by each unit are fused through dense connections, allowing features at each stage to interact across levels, realizing the network's powerful feature expression capabilities. The feature fusion module is used to perform feature fusion on the fine facial feature maps obtained by each projection unit; The face image reconstruction module is used to perform ordinary convolution on the fused face feature map to reconstruct the target high-resolution face image.

5. A computer storage medium, characterized in that: A computer program executable by a computer processor is stored therein, and the computer program executes a face super-resolution method based on a convenient alternating projection network as described in any one of claims 1 to 3.

Images