A method for removing neck wrinkles based on hidden space feature editing and multi-scale feature fusion
By employing latent space feature editing and multi-scale feature fusion, and utilizing neck mask guidance, this method addresses the dependence of existing neck wrinkle removal algorithms on neck wrinkle segmentation models, achieving more accurate and flexible neck wrinkle removal results suitable for various portrait beautification applications.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIAMEN ZHENJING TECH CO LTD
- Filing Date
- 2023-06-02
- Publication Date
- 2026-06-30
AI Technical Summary
Existing neck wrinkle removal algorithms rely on neck wrinkle segmentation models, resulting in high data annotation costs and unsatisfactory results. They are difficult to accurately remove neck wrinkles and avoid accidentally damaging other wrinkles or accessories.
A multi-scale feature fusion method based on latent space feature editing is adopted. Guided by a neck mask, the method uses multi-scale feature fusion and brush network to edit the features of the neck wrinkle region, avoiding accidental damage to non-neck wrinkles and jewelry, thus achieving accurate removal of neck wrinkles.
It improves the accuracy and flexibility of the neck wrinkle removal algorithm, making it applicable to a wider range of portrait beautification algorithms. It reduces the reliance on the neck wrinkle segmentation model, ensuring skin tone uniformity and natural results.
Smart Images
Figure CN116630759B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of beauty technology, and in particular to a method for removing neck wrinkles based on multi-scale feature fusion using latent space feature editing. Background Technology
[0002] Current neck wrinkle removal algorithms primarily use a neck wrinkle segmentation model to obtain a neck wrinkle mask in the image to be removed, then process the neck wrinkle region, and finally fuse the original image with the processed result to obtain the final neck wrinkle removal image. This type of neck wrinkle removal algorithm heavily relies on the neck wrinkle segmentation model, and the cost of annotating neck wrinkle segmentation data during development is relatively high. If the neck wrinkle segmentation model cannot obtain an accurate neck wrinkle mask, the effect will be unsatisfactory, for example, incomplete removal, ineffectiveness, or accidental damage to other types of wrinkles. With the increasing demand for selfies and photography, users have placed higher demands on the effectiveness and reliability of portrait beautification algorithms. For neck wrinkle removal algorithms, removing various types of neck wrinkles is a basic goal, avoiding accidental damage to the jawline, hand lines, necklaces, etc., is an important aspect, and maintaining the uniform skin tone of the original neck wrinkle area after removing "tire tread" type neck wrinkles is another key focus for the algorithm. Summary of the Invention
[0003] In view of this, the purpose of this invention is to propose a multi-scale feature fusion method for removing neck wrinkles based on latent space feature editing, which can get rid of the dependence of neck wrinkle removal algorithms on neck wrinkle segmentation models, remove neck wrinkles in portraits, and beautify portraits.
[0004] According to one aspect of the present invention, a multi-scale feature fusion method for neck wrinkle removal based on latent space feature editing is provided, comprising: segmenting an image of the neck wrinkle to be removed to obtain a neck mask; obtaining a neck image patch and a mask image patch based on the neck mask; obtaining a latent space feature with n channels based on the neck image patch and the mask image patch, wherein the latent space feature is a feature map with dimensions n×h×w obtained by feature extraction of a 4-channel image obtained by concatenating the neck image patch and the mask image patch by channel using a CNN, where n represents the number of layers of the feature map, h represents the height of the feature map, and w represents the width of the feature map; preprocessing the latent space feature to obtain preprocessed latent space feature; fusing and decoding based on the preprocessed latent space feature to obtain a preliminary result image of neck wrinkle removal; and postprocessing based on the preliminary result image of neck wrinkle removal to obtain a final result image after neck wrinkle removal.
[0005] In the above technical solution, a neck mask is used as a guide to avoid accidentally damaging non-neck wrinkles (such as lip lines or handprints on the neck) and neck jewelry, making the neck wrinkle removal effect more accurate and reliable. It does not rely on high-precision neck wrinkle segmentation models, making it more flexible and easier to use. Because it does not depend on segmentation models specific to particular wrinkle types, it is applicable to a wider range of portrait beautification algorithms.
[0006] In some embodiments, obtaining a neck image block and a mask image block based on a neck mask specifically includes: calculating an alignment matrix based on the neck mask, aligning the image to be removed with the neck mask, and obtaining neck image blocks and neck mask image blocks of three different resolutions: high, medium, and low.
[0007] In the above technical solution, the advantage of this setup is that it enables the algorithm to better capture features related to neck wrinkles. Because neck wrinkles are diverse, with light, dark, fine, and coarse lines intermingling, and the receptive field of a fixed-size CNN convolutional kernel is limited, features of light and fine lines can be captured more accurately at higher resolutions, while the relationship between pixels in areas of deep and coarse lines and their surrounding pixels can be better learned at lower resolutions.
[0008] In some embodiments, latent space features with n channels are obtained based on a neck image patch and a mask image patch, specifically including:
[0009] Following a progression from low to high resolution, at each resolution, the neck image patch and the neck mask image patch are concatenated by channel for feature extraction, resulting in a latent space feature map with n channels. This latent space feature map is a feature map with dimensions n×h×w obtained by CNN feature extraction from the 4-channel image obtained by concatenating the neck image patch and the mask image patch by channel, where n represents the number of layers in the feature map, h represents the height of the feature map, and w represents the width of the feature map. Features extracted at lower resolutions are reused in the feature extraction process at higher resolutions. The neck image patch and the neck mask image patch are combined into a 4-channel image, with the neck mask image patch providing information about the region that the algorithm needs to process.
[0010] In the above technical solution, the advantage of this setup is that feature reuse fuses information acquired by the algorithm under different receptive fields. This allows fine-grained features that cannot be captured at lower resolutions to be supplemented by features extracted at higher resolutions, and neighborhood features that cannot be captured at higher resolutions to be supplemented by features extracted at lower resolutions. By using a neck mask image patch, the algorithm's processing area is limited to the skin region of the neck, avoiding accidental damage to other irrelevant areas.
[0011] In some embodiments, preprocessing the latent space features to obtain preprocessed latent space features specifically includes:
[0012] The first latent space features are input into the brush network to obtain the second latent space features with m channels.
[0013] The preprocessed latent space features are obtained by concatenating the second latent space features with m channels with the last nm feature channels of the first latent space features. The second latent space features with m channels represent the features of the neck wrinkle region in the image to be removed after the neck wrinkles have been removed.
[0014] In the above technical solution, the purpose of this setting is to replace the feature channels representing neck wrinkles in the first latent space features with the latent space features without neck wrinkles output by the brush network, thereby editing the first latent space features and removing neck wrinkle information from the image to be removed in the latent space.
[0015] In some embodiments, the values of m and n are determined through cross-reconstruction. The cross-reconstruction process is as follows: given a data pair (src, tar, mask), src represents the image to be removed from the neck wrinkles, tar represents the src after neck wrinkle removal, and mask represents the neck mask, aligning (src, tar, mask) yields data pairs (src, tar, mask) at high, medium, and low resolutions. hr tar hr mask hr ),(src mr tar mr mask mr ),(src lr tar lr mask lr At each resolution, set the src... *r With mask *r After stitching by channel, a 4-channel image is obtained. *r , input1 *r Inputting the feature extraction network at the corresponding resolution yields latent space features (feature1) with n channels. *r ; tar *r With mask *r After stitching by channel, a 4-channel image is obtained. (input2) *r , input2 *r Inputting the feature extraction network at the corresponding resolution yields latent space features with n channels, feature2. *r ; will feature1 *r The first m feature channels and feature2 *r The feature is obtained by splicing the last nm feature channels. 1’ *r feature2 *r The first m feature channels and feature1 *r The feature is obtained by splicing the last nm feature channels. 2’ *r ; to combine features at different resolutions 1’ *rFeature mixing and decoding are performed to obtain output1, which combines features at different resolutions. 2’ *r Feature mixing and decoding are performed to obtain output2; consistency constraints are applied to make output1 close to tar and output2 close to src. The optimal values of n and m are then selected using this method.
[0016] In the above technical solution, the purpose of this setting is to determine the optimal m and n before formally training the CNN, so as to ensure that the features of the neck wrinkle region and the features of the non-neck wrinkle region can be well decoupled during the formal training of the CNN.
[0017] In some embodiments, a preliminary result image for neck wrinkle removal is obtained by fusing and decoding based on preprocessed latent space features. Specifically, the preliminary result image for neck wrinkle removal is obtained by fusing and decoding preprocessed latent space features of three different resolutions: high, medium, and low.
[0018] In the above technical solutions, multi-scale fusion can improve feature representation. Compared with single-scale features, multi-scale feature fusion can produce more discriminative and robust features, thereby improving the performance of classification, detection, and recognition. Multi-scale feature fusion utilizes features from different scales to capture the spatial information of objects. By fusing features from different scales, more accurate and reliable spatial information can be obtained, thereby improving the accuracy and robustness of the model.
[0019] In some embodiments, post-processing is performed on the preliminary result image of neck wrinkle removal to obtain the final result image after neck wrinkle removal. Specifically, the preliminary result image of neck wrinkle removal is subjected to skin color unification processing; an anti-alignment matrix is calculated based on the alignment matrix, and the neck boundary information of the image to be removed is calculated; the final result image after neck wrinkle removal is anti-aligned to the neck position of the image to be removed using the neck boundary information, and combined with the degree coefficient representing the intensity of neck wrinkle removal, alpha fusion is performed to obtain the final result image after neck wrinkle removal.
[0020] In the above technical solution, skin tone unification processing is used to solve the problem of uneven skin tone after the initial wrinkle removal. Alpha fusion is further used to achieve image mixing and fusion, making the final synthesized image more natural and realistic.
[0021] According to another aspect of the present invention, a multi-scale feature fusion device for neck wrinkle removal based on latent space feature editing is provided, comprising an acquisition module, an analysis module, a feature extraction module, a feature and processing module, a fusion decoding module, and a post-processing module electrically connected in sequence; the acquisition module is used to segment the image of the neck wrinkle to be removed to obtain a neck mask; the analysis module is used to acquire neck image blocks and mask image blocks based on the neck mask; the feature extraction module is used to obtain latent space features with n channels based on the neck image blocks and mask image blocks; the feature and processing module is used to preprocess the latent space features to obtain preprocessed latent space features; the fusion decoding module is used to perform fusion decoding based on the preprocessed latent space features to obtain a preliminary result image of neck wrinkle removal; and the post-processing module is used to perform post-processing based on the preliminary result image of neck wrinkle removal to obtain a final result image after neck wrinkle removal.
[0022] In the above technical solution, a multi-scale feature fusion method for neck wrinkle removal based on latent space feature editing is constructed. This method uses a neck mask as a guide to avoid accidentally damaging non-neck wrinkles (such as lip lines or handprints on the neck) and neck accessories, making the neck wrinkle removal effect more accurate and reliable. It does not rely on high-precision neck wrinkle segmentation models, making it more flexible and easier to use. Because it does not depend on segmentation models specific to particular wrinkle types, it is applicable to a wider range of portrait beautification algorithms.
[0023] According to another aspect of the present invention, a multi-scale feature fusion neck wrinkle removal device based on latent space feature editing is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the multi-scale feature fusion neck wrinkle removal method based on latent space feature editing as described in any of the preceding claims.
[0024] According to another aspect of the present invention, a computer-readable storage medium is provided, storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the multi-scale feature fusion method for removing neck wrinkles based on latent space feature editing as described in any of the preceding claims. Attached Figure Description
[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0026] Figure 1 This is one of the flowcharts of an embodiment of the multi-scale feature fusion method for removing neck wrinkles based on latent space feature editing of the present invention;
[0027] Figure 2 This is the second flowchart of an embodiment of the multi-scale feature fusion method for removing neck wrinkles based on latent space feature editing of the present invention;
[0028] Figure 3 This is a schematic diagram of the S400 process of an embodiment of the multi-scale feature fusion method for removing neck wrinkles based on latent space feature editing of the present invention. Detailed Implementation
[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0030] This invention provides a multi-scale feature fusion method for removing neck wrinkles based on latent space feature editing, which can get rid of the dependence of neck wrinkle removal algorithms on neck wrinkle segmentation models, remove neck wrinkles in portraits, and beautify portraits.
[0031] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the multi-scale feature fusion method for removing neck wrinkles based on latent space feature editing according to the present invention. It should be noted that if substantially the same result is obtained, the method of the present invention is not necessarily identical. Figure 1 The illustrated process sequence is limited. For example... Figure 1 As shown, the method includes the following steps:
[0032] S100. Segment the image of the neck wrinkles to be removed to obtain a neck mask;
[0033] In this embodiment, specifically, the image of the neck wrinkles to be removed is input into a neck segmentation network to obtain a neck mask in the image. The structure of the neck segmentation network in this embodiment is based on a lightweight version of Unet.
[0034] In this embodiment, compared with the existing technology that uses a neck wrinkle segmentation model to directly remove neck wrinkles, this embodiment uses a more mature neck segmentation network that can quickly segment the neck region in the image, reducing annotation and modeling costs.
[0035] S200: Obtain the neck image block and the mask image block based on the neck mask;
[0036] In this embodiment, specifically, an alignment matrix is calculated based on the neck mask, and the image to be removed and the neck mask are aligned to obtain neck image blocks and neck mask image blocks of three different resolutions: high, medium, and low.
[0037] In this embodiment, the height and width of the high-resolution image block are consistent with the aligned image block, while the height and width of the medium-resolution and low-resolution image blocks are half of the previous resolution.
[0038] S300. Obtain latent space features with n channels based on neck image patches and mask image patches;
[0039] In this embodiment, specifically, following the order from low resolution to high resolution, at each resolution, the neck image block and the neck mask image block are stitched together by channel and then feature extraction is performed to obtain latent space features with n channels.
[0040] S400. Preprocess the latent space features to obtain the preprocessed latent space features.
[0041] In this embodiment, please refer to Figure 2 The first latent space features are input into the brush network to obtain the second latent space features with m channels. The brush network learns the skin features of the neck wrinkle region in the image to be removed from the first latent space features. There is a brush network for each of the three resolutions: high, medium and low. The brush network is a CNN network.
[0042] The preprocessed latent space features are obtained by concatenating the second latent space feature with m channels with the last nm feature channels of the first latent space feature.
[0043] In this embodiment, the values of m and n are determined by cross-reconstruction.
[0044] S500: Based on the preprocessed latent space features, a preliminary result image of neck wrinkle removal is obtained by fusion decoding.
[0045] In this embodiment, specifically, the preprocessed latent space features of three different resolutions (high, medium, and low) are fused and decoded to obtain a preliminary result image of neck wrinkle removal.
[0046] S600. Based on the preliminary results of neck wrinkle removal, post-processing is performed to obtain the final result image after neck wrinkle removal.
[0047] In this embodiment, specifically, the skin tone of the preliminary results image of neck wrinkle removal is uniformly processed;
[0048] The anti-alignment matrix is calculated based on the alignment matrix, and the neck boundary information of the image to be treated for neck wrinkles is calculated.
[0049] The neck boundary information is used to align the result image after neck wrinkle removal with the neck position of the image to be removed, and then alpha fusion is performed using a coefficient representing the intensity of neck wrinkle removal to obtain the result image after neck wrinkle removal.
[0050] Please see Figure 3 , Figure 3 This is a method framework diagram for a specific example in this embodiment, which specifically includes the following steps:
[0051] 1. Input the image of the neck wrinkles to be removed into the neck segmentation network to obtain the neck mask in the image.
[0052] 2. Calculate the alignment matrix based on the neck mask, align the image to be removed with the neck mask, and obtain neck image blocks and neck mask image blocks of three different resolutions: high, medium, and low.
[0053] 3. Following a progression from low to high resolution, at each resolution, the neck image patch and the neck mask image patch are concatenated by channel for feature extraction, resulting in a latent space feature with n channels. The first m channels in the latent space feature represent features related to the neck wrinkle region, while the last nm channels represent features unrelated to the neck wrinkle region. The optimal values of n and m were determined during the development phase using data pairs through cross-reconstruction. Features extracted at low resolution are reused in the feature extraction process at medium resolution, and vice versa. Throughout the feature extraction process using CNN at each resolution, the resolution of the feature map remains consistent. Since neck wrinkles are diverse, primarily manifested in differences in depth and thickness, feature extraction from images at different resolutions allows the algorithm to have different receptive fields, enabling more accurate capture of neck wrinkle information.
[0054] 4. Edit the first m channels of the latent space features to obtain new latent space features. The latent space feature editing process is as follows:
[0055] 4.1 Input the latent space features with n channels into the brush network to obtain latent space features with m channels.
[0056] 4.2. After concatenating the latent space features of the m channels obtained in 4.1 with the last nm feature channels of the original latent space features, a new latent space feature is obtained.
[0057] 5. After fusing and decoding the new latent space features obtained from the three resolutions, a preliminary result image of neck wrinkle removal is obtained.
[0058] 6. Input the preliminary results of neck wrinkle removal obtained in step 5 into a skin tone unification network to remove uneven skin tone that may occur after the initial wrinkle removal.
[0059] 7. Calculate the anti-alignment matrix based on the alignment matrix, calculate the boundary information, and combine the anti-alignment matrix of the result image obtained in step 6 with the degree coefficient to perform alpha fusion to obtain the result image after neck wrinkle removal.
[0060] This invention uses a neck mask as a guide to avoid accidentally damaging non-neck wrinkles (such as lip lines or handprints on the neck) and neck jewelry, making neck wrinkle removal more accurate and reliable. It does not rely on high-precision neck wrinkle segmentation models, making it more flexible and easier to use. Because it does not depend on segmentation models specific to particular wrinkle types, it is applicable to a wider range of portrait beautification algorithms.
[0061] Example 2
[0062] A multi-scale feature fusion device for neck wrinkle removal based on latent space feature editing is proposed, comprising an acquisition module, an analysis module, a feature extraction module, a feature processing module, a fusion decoding module, and a post-processing module, which are electrically connected in sequence. The acquisition module is used to segment the image to be removed from the neck wrinkle to obtain a neck mask. The analysis module is used to obtain neck image blocks and mask image blocks based on the neck mask. The feature extraction module is used to obtain latent space features with n channels based on the neck image blocks and mask image blocks. The feature processing module is used to preprocess the latent space features to obtain preprocessed latent space features. The fusion decoding module is used to perform fusion decoding based on the preprocessed latent space features to obtain a preliminary result image of neck wrinkle removal. The post-processing module is used to perform post-processing based on the preliminary result image of neck wrinkle removal to obtain the final result image after neck wrinkle removal.
[0063] It should be noted that the above modules correspond one-to-one with each step of the multi-scale feature fusion method for removing neck wrinkles based on latent space feature editing. Since each step has been explained in detail above, the specific principles of the function of each module will not be explained in detail here.
[0064] Example 3
[0065] A multi-scale feature fusion neck wrinkle removal device based on latent space feature editing is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the multi-scale feature fusion neck wrinkle removal method based on latent space feature editing as described above.
[0066] It should be noted that, given that a multi-scale feature fusion method for removing neck wrinkles based on latent space feature editing has already been described in detail above, the specific principles of this method will not be elaborated upon here.
[0067] Example 4
[0068] A computer-readable storage medium is provided, storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the multi-scale feature fusion method for removing neck wrinkles based on latent space feature editing as described in any of the preceding claims.
[0069] It should be noted that, given that a multi-scale feature fusion method for removing neck wrinkles based on latent space feature editing has already been described in detail above, the specific principles of this method will not be elaborated upon here.
[0070] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A method for editing neck wrinkle based on multi-scale feature fusion in hidden space features, characterized in that, include: The image of the neck wrinkles to be removed is input into a neck segmentation network to obtain a neck mask; The alignment matrix is calculated based on the neck mask. The image to be removed from the neck wrinkles is aligned with the neck mask to obtain neck image blocks of high, medium and low resolutions as well as mask image blocks. Following the order from low resolution to high resolution, at each resolution, the neck image patch and the neck mask image patch are stitched together by channel and then feature extraction is performed to obtain latent space features with n channels. The latent space features are input into a brush network to obtain a second latent space feature with m channels. The second latent space feature with m channels is concatenated with the last nm feature channels of the latent space feature to obtain the preprocessed latent space feature. The brush network is used to learn the skin features of the neck wrinkle region in the image to be treated from the latent space features, and there is a brush network for each of the high, medium and low resolutions. The first m channels of the latent space feature represent features related to the neck wrinkle region, and the last nm channels represent features unrelated to the neck wrinkle region. The preprocessed latent space features of high, medium and low resolutions are fused and decoded to obtain a preliminary result image of neck wrinkle removal; Skin tone unification is performed on the preliminary result image of neck wrinkle removal; the anti-alignment matrix is calculated based on the alignment matrix, and the neck boundary information of the image to be removed is calculated; the preliminary result image of neck wrinkle removal is anti-aligned to the neck position of the image to be removed using the neck boundary information, and alpha fusion is performed in combination with the degree coefficient representing the strength of neck wrinkle removal to obtain the result image after neck wrinkle removal.
2. The method for removing neck wrinkles based on multi-scale feature fusion using latent space feature editing as described in claim 1, characterized in that, The values of m and n are determined by cross-reconstruction.
3. A multi-scale feature fusion anti-wrinkle device based on latent space feature editing, characterized in that, Based on the method according to any one of claims 1-2, the apparatus comprises: It includes an acquisition module, an analysis module, a feature extraction module, a feature and processing module, a fusion decoding module, and a post-processing module that are connected in sequence. The acquisition module is used to input the image of the neck wrinkle removal target into the neck segmentation network for segmentation to obtain a neck mask; The analysis module is used to calculate the alignment matrix based on the neck mask, align the image of the neck wrinkles to be removed with the neck mask, and obtain neck image blocks and mask image blocks of three different resolutions: high, medium and low. The feature extraction module is used to extract features by concatenating the neck image patch and the neck mask image patch by channel at each resolution, from low resolution to high resolution, to obtain latent space features with n channels. The feature and processing module is used to input latent space features into a brush network to obtain a second latent space feature with m channels; the second latent space feature with m channels is concatenated with the last nm feature channels of the latent space feature to obtain a preprocessed latent space feature; wherein, the brush network is used to learn the skin features of the neck wrinkle region in the image to be treated from the latent space feature, and there is a brush network for each of the high, medium and low resolutions; the first m channels of the latent space feature represent features related to the neck wrinkle region, and the last nm channels represent features unrelated to the neck wrinkle region; The fusion decoding module is used to fuse and decode the preprocessed latent space features of high, medium and low resolutions to obtain a preliminary result image of neck wrinkle removal. The post-processing module is used to unify the skin tone of the preliminary neck wrinkle removal result image; calculate the anti-alignment matrix based on the alignment matrix, and calculate the neck boundary information of the image to be removed; use the neck boundary information to anti-align the preliminary neck wrinkle removal result image to the neck position of the image to be removed, and combine it with the degree coefficient representing the strength of neck wrinkle removal for alpha fusion to obtain the final result image after neck wrinkle removal.
4. A multi-scale feature fusion device for removing neck wrinkles based on latent space feature editing, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform a multi-scale feature fusion method for removing neck wrinkles based on latent space feature editing as described in any one of claims 1 to 2.
5. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the multi-scale feature fusion method for removing neck wrinkles based on latent space feature editing as described in any one of claims 1 to 2.