A semantically aligned image virtual try-on method
By using a four-stage virtual try-on framework with semantic alignment, the problems of misalignment between clothing and target dressing areas and loss of semantic information are solved, generating virtual try-on images that conform to the human body layout and improving the realism of the images.
Patent Information
- Application Number
- CN202310438278.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-23
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2043-04-23
AI Technical Summary
Existing image-based virtual try-on methods suffer from misalignment between clothing and the target wearing area and loss of semantic information during the clothing alignment and try-on synthesis stages, resulting in generated images that do not conform to the human body layout.
A four-stage virtual try-on framework based on semantic alignment is adopted, including data preprocessing, human semantic synthesis module, clothing alignment module and try-on synthesis module. Virtual try-on images that conform to human body layout are generated through regional restriction strategy and spatial adaptive normalization layer.
The generated virtual try-on images align the clothing with the target dressing area, preserve the person's posture and features, and conform to the semantic layout of the human body, thus improving the photorealistic quality of the images.
Smart Images

Figure CN116452705B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image virtual try-on, specifically relating to a semantically aligned image virtual try-on method. Background Technology
[0002] Image-based virtual try-on technology aims to transfer specific clothing images onto target person images and generate photorealistic images or videos. Its biggest difference from graphics-based virtual try-on methods is that it does not require 3D reconstruction of the clothing and person; instead, it relies on the model's semantic understanding of the image to synthesize the image. Given a target clothing image and a person image, the model's output image should meet the following requirements: (1) the target clothing should be worn at the target location on the person; (2) the person's posture and features should be preserved; (3) the texture and details of the target clothing should be preserved; and (4) clothing outside the target clothing area should be well preserved.
[0003] The most popular image-based virtual try-on method is mainly based on clothing deformation, which generally includes two main stages: (1) clothing alignment: deforming the clothing to align with the human body; (2) try-on synthesis: merging the deformed clothing with the human body to generate the try-on outfit. Common geometric methods used for clothing deformation include thin plate spline interpolation, optical flow, and 3D-assisted methods. At the same time, some methods choose to predict human semantic maps to better guide try-on synthesis. The advantage of clothing deformation is that it can align the clothing and the human body in advance to provide more clothing texture priors in the try-on synthesis stage. However, existing clothing alignment methods only provide a rough alignment effect, such as the inability to handle self-occlusion, which will cause misalignment between the clothing and the target wearing area. In addition, in the try-on synthesis stage, semantically guided methods tend to directly input the human semantic map into the module, which often loses semantic information, thereby exacerbating the misalignment between the clothing and the target wearing area. Summary of the Invention
[0004] To address the aforementioned problems, this invention discloses a semantically aligned image-based virtual try-on method that can generate virtual try-on images that better conform to the human body layout. First, the human body image and human semantic map are preprocessed to remove arm and clothing information. Second, a human semantic synthesis module predicts the target human semantic map. Then, a clothing alignment module aligns the clothing and the human body. Finally, a try-on synthesis module obtains the final try-on image. In the clothing alignment module, a region restriction strategy is proposed to limit the range of clothing deformation and reduce excessive clothing distortion. In the try-on synthesis module, a spatially adaptive normalization layer is used instead of a regular normalization layer to fix the human semantic map information.
[0005] To achieve the above objectives, the technical solution of the present invention is as follows:
[0006] A semantically aligned image virtual try-on method, given a person image I∈R 3 × H × W And a target clothing image c∈R 3 × H × W The model will output the target clothing c∈R. 3 × H × W Images of people Preserving both posture and human features simultaneously involves the following steps:
[0007] (1) Data Preprocessing: This part aims to remove clothing information from any target clothing areas in Person Image I, thereby obtaining a representation where clothing is unrelated to the person—Person Image I. a and human semantic map S a Both removed information about clothing and arms.
[0008] (2) Human semantic synthesis module: The input is a human semantic graph S with arm and clothing information removed. a And pose image P and clothing mask c m The model will generate a semantic map of the target human body wearing the target clothing.
[0009] (3) Clothing Alignment Module: This module distorts the clothing image c using a non-rigid TPS transformation to align it with the clothing area of the target human body, and avoids distortion of the clothing T after deformation through a region restriction strategy. θ (c) with target human semantic map The misalignment issue also prevents the clothing from becoming excessively distorted.
[0010] (4) Fitting Synthesis Module: This module receives the person image I, the pose P, and the aligned clothing T. θ (c) As input, and using the target human semantic map Conditional normalization information is used as the generator to generate the target image.
[0011] In step (1), the data preprocessing process is as follows: given a human image I, human pose maps P∈R are extracted using a pre-trained human pose extractor and a human parsing extractor, respectively. 18HW and human semantic graph S∈L HW (L represents the total number of semantic map labels). First, the clothing portion S of the human semantic map S. cThe coordinates provided by the pose graph P are set to 0, and then a columnar mask is created centered on the coordinates provided by the pose graph P to mask the arm portion of image S. For the human body image I, the clothing portion is masked using the semantic graph S, and then a columnar mask is created centered on the coordinates provided by the pose graph P to expand the mask area so that the mask completely covers the upper body of the human body. After the above data preprocessing, the clothing-independent representation S is obtained. a and I a They completely removed clothing information while preserving information about non-target clothing areas to the greatest extent possible.
[0012] In step (2), human semantic synthesis uses U-Net as the generator. Generator G S Receive human semantic map S a Pose image P and clothing mask c m As input, predict the target human semantic map. A multi-scale discriminator structure is adopted. Two discriminators D are used. S1 and D S2 Human semantic maps at original size and twice the downsampled size To increase the receptive field of the discriminator, discrimination is performed. Specifically, the obtained... And input S a ⊕P⊕c m Channel combination is used as the input to the discriminator. The two discriminators receive the original input and the input after being downsampled by 2 times, respectively. They are then passed through a Sigmoid layer to output confidence scores, and the average of the scores is taken as the final output of the discriminator.
[0013] In step (3), the garment alignment module twists the garment c to align it with the target garment I. c This allows the virtual fitting module to easily integrate clothing and human body images. The clothing alignment module learns from a set of TPS parameters and uses these parameters to perform a TPS transformation on clothing c to obtain aligned clothing c. w Two encoders, E1 and E2, each composed of a convolutional neural network, are used to process the human semantic map generated in the previous stage. Encode the clothing c to generate two feature vectors. The obtained features will be L2 normalized, spatially flattened, and organized into a matrix. Then, adjust the shapes of the two and perform matrix multiplication to obtain... Corr, the correlation matrix between c and c:
[0014]
[0015] The Corr matrix contains f E1 and f E2The similarity between each pair of pixels is then determined. A regressor R is then used to downsample Corr to obtain the corresponding TPS parameters θ∈R. 50 By performing a TPS transformation on garment c, the deformed garment T can be obtained. θ (c)
[0016] In step (3), a region restriction strategy is used to control the deformation range of the clothing. Specifically, this is achieved by using a human semantic map... Extract the target clothing mask M c =S c Then, with the transformed clothing T-shirt θ (c) Perform a dot product to obtain the final output of this module, i.e., the aligned clothing c. w :
[0017] c w =T θ (c)⊙M c
[0018] In step (3), the target loss is the alignment of the clothing c. w And target clothing I c The L1+ perceptual loss is used. Simultaneously, to ensure the module focuses on the overall shape of the clothing, rather than just deforming a portion of the target clothing area, a shape loss is introduced. Specifically, the same set of TPS parameters as the clothing deformation is used to deform the in-store clothing mask c. m , obtained with T θ (c) Clothing mask T with the same shape θ (c m Then calculate it and the target clothing mask M. c The L1 loss between them can be expressed by the following formula:
[0019]
[0020] Therefore, the overall loss function of the module is defined as:
[0021]
[0022] In step (4), the fitting room synthesis module merges the aligned clothing and person images obtained in the previous stage to obtain the final fitting room image. The input is a clothing-independent representation of the person. a Posture heatmap P and aligned clothing c w Channel combination I a ⊕P⊕c w Output fitting images Human semantic graph The input is a conditionally normalized SPADE ResBlock. The fitting synthesis module consists of a series of stacked SPADE ResBlocks and upsampling layers, employing a pure decoder architecture. The input feature map is gradually enlarged to the size of the final output.
[0023] In step (4), the module first adjusts I a ⊕P⊕c w The size is determined, and then a 3×3 convolutional layer is used to output a feature map of size 1024×4×3. The output feature map will be resized to match the size of the input. a ⊕P⊕c w Channel combination is used as the input to the SPADE ResBlock. The structure of the SPADE ResBlock is similar to that of the Residual Block, consisting of three convolutional layers paired with residual connections. Similarly, the conditionally normalized input to the SPADE layer... It also needs to be adjusted to the same size as the current SPADEResBlock input. The output of SPADEResBlock will be upsampled by a factor of two, and then its channel count will be halved by a 3×3 convolutional layer. The resulting output will be used as the input to the next SPADEResBlock. As can be seen, the model employs a multi-scale fusion strategy, which ensures the normalization effect of feature maps at different scales under the guidance of the semantic map, thereby achieving alignment between the output image and the semantic layout. The module's objective loss function consists of adversarial loss, reconstruction loss (L1 loss), perceptual loss, and the discriminator's feature matching loss:
[0024]
[0025] In the formula λ rec , λ perc and λ FM These are the coefficients for the corresponding losses.
[0026] The beneficial effects of this invention are as follows:
[0027] (1) A four-stage virtual fitting framework with semantic alignment can be used to generate virtual fitting images that conform to the human body layout.
[0028] (2) The clothing alignment module with regional restriction is adopted. By introducing a regional restriction strategy, the range of clothing deformation is limited, the excessive distortion of clothing is reduced, and the clothing is aligned with the target clothing area.
[0029] (3) A fitting synthesis module with semantic layout alignment is adopted, and a spatial adaptive normalization layer is used to fix human semantic map information, thereby generating fitting images that are more in line with human semantic layout. Attached Figure Description
[0030] Figure 1 This is a flowchart of the image virtual try-on method with semantic alignment according to the present invention.
[0031] Figure 2 This is a flowchart of the data preprocessing process of the present invention.
[0032] Figure 3 This is a schematic diagram of the human semantic synthesis module of the present invention.
[0033] Figure 4 This is a structural diagram of the human semantic synthesis module of the present invention.
[0034] Figure 5 This is a schematic diagram of the clothing alignment module with area restriction according to the present invention.
[0035] Figure 6 This is a structural diagram of the clothing alignment module with area limitation according to the present invention.
[0036] Figure 7 This is a schematic diagram of the fitting room synthesis module with semantic layout alignment according to the present invention. Figure 1 .
[0037] Figure 8 This is a schematic diagram of the fitting room synthesis module with semantic layout alignment according to the present invention. Figure 2 .
[0038] Figure 9 This is a structural diagram of the fitting room synthesis module with semantic layout alignment according to the present invention.
[0039] Figure 10 This is a schematic diagram illustrating the virtual try-on effect of the present invention. Detailed Implementation
[0040] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.
[0041] like Figure 1 As shown, a semantically aligned image virtual try-on method includes the following steps:
[0042] Step S1: Data Preprocessing: This part aims to remove clothing information from any target clothing areas in Person Image I, thereby obtaining a representation where clothing is irrelevant to the person—Person Image I. a and human semantic map S a Both removed information about clothing and arms.
[0043] Step S2: Human semantic synthesis module: The input is a human semantic graph S with arm and clothing information removed. a And pose image P and clothing mask c m The model will generate a semantic map of the target human body wearing the target clothing.
[0044] Step S3: Clothing Alignment Module: This module distorts the clothing image c using a non-rigid TPS transformation to align it with the clothing area of the target human body, and uses a region constraint strategy to avoid the deformed clothing Tθ(c) from aligning with the semantic map of the target human body. The misalignment issue also avoids excessive distortion of the clothing;
[0045] Step S4: Fitting and Composition Module: This module receives the person image I, the pose P, and the aligned clothing T. θ (c) As input, and using the target human semantic map Conditional normalization information is used as the generator to generate the target image.
[0046] like Figure 2 As shown, the data preprocessing flow of this invention is as follows: Given a human image I, human pose maps P∈R are extracted using a pre-trained human pose extractor and a human parsing extractor, respectively. 18×H×W and human semantic graph S∈L H×W (L represents the total number of semantic map labels). First, the clothing portion S of the human semantic map S. c The coordinates provided by the pose graph P are set to 0, and then a columnar mask is created centered on the coordinates provided by the pose graph P to mask the arm portion of image S. For the human body image I, the clothing portion is masked using the semantic graph S, and then a columnar mask is created centered on the coordinates provided by the pose graph P to expand the mask area so that the mask completely covers the upper body of the human body. After the above data preprocessing, the clothing-independent representation S is obtained. a and I a They completely removed clothing information while preserving information about non-target clothing areas to the greatest extent possible.
[0047] like Figure 3 As shown, the human semantic synthesis module of this invention follows a process. Human semantic synthesis uses U-Net as the generator. Generator G S Receive human semantic map S a Pose image P and clothing mask c m As input, predict the target human semantic map. A multi-scale discriminator structure is adopted. Two discriminators D are used. S1 and D S2 Human semantic maps at original size and twice the downsampled size To increase the receptive field of the discriminator, discrimination is performed. Specifically, the obtained... And input S a ⊕P⊕c mChannel combination is used as input to the discriminator. Two discriminators receive the original input and the input after being downsampled by a factor of two, respectively. Each input is then processed through a Sigmoid layer to output confidence scores, and the average of these scores is taken as the final output of the discriminator. The overall loss function of the module is:
[0048]
[0049] In the formula λ CE is the coefficient of cross-entropy loss.
[0050] like Figure 4 As shown, the structure of the human semantic synthesis module of the present invention is as follows. The first layer of the encoder is a convolutional layer with a 4×4 convolutional kernel downsampled by a factor of 2, followed by two identical convolutional blocks. Each encoder convolutional block contains a LeakyReLU layer, a convolutional layer with a 4×4 convolutional kernel downsampled by a factor of 2, and an instance normalization layer. The middle layer of the module is a convolutional block containing a size-invariant convolutional layer and a transposed convolutional layer. The encoder consists of three convolutional blocks. Except for the last convolutional block, each convolutional block contains a ReLU layer, a transposed convolutional layer upsampled by a factor of 2, and an instance normalization layer. The last convolutional block replaces the instance normalization layer with a Tanh layer.
[0051] like Figure 5 As shown, the process of the region-restricted clothing alignment module proposed in this invention is as follows: The clothing alignment module twists clothing c to align with target clothing I. c This allows the virtual fitting module to easily integrate clothing and human body images. The clothing alignment module learns from a set of TPS parameters and uses these parameters to perform a TPS transformation on clothing c to obtain aligned clothing c. w Two encoders, E1 and E2, each composed of a convolutional neural network, are used to process the human semantic map generated in the previous stage. Encode the clothing c to generate two feature vectors. The obtained features will be L2 normalized, spatially flattened, and organized into a matrix. Then, adjust the shapes of the two and perform matrix multiplication to obtain... Corr, the correlation matrix between c and c:
[0052]
[0053] The Corr matrix contains f E1 and f E2 The similarity between each pair of pixels is then determined. A regressor R is then used to downsample Corr to obtain the corresponding TPS parameters θ∈R. 50 By performing a TPS transformation on garment c, the deformed garment T can be obtained. θ (c)
[0054] In the clothing alignment module, a region restriction strategy is used to control the deformation range of the clothing. Specifically, this is achieved by analyzing the human semantic map... Extract the target clothing mask M c =S c Then, with the transformed clothing T-shirt θ (c) Perform a dot product to obtain the final output of this module, i.e., the aligned clothing c. w :
[0055] c w =T θ (c)⊙M c
[0056] The target loss is the alignment of the clothing c. w And target clothing I c The L1+ perceptual loss is used. Simultaneously, to ensure the module focuses on the overall shape of the clothing, rather than just deforming a portion of the target clothing area, a shape loss is introduced. Specifically, the same set of TPS parameters as the clothing deformation is used to deform the in-store clothing mask c. m , obtained with T θ (c) Clothing mask T with the same shape θ (c m Then calculate it and the target clothing mask M. c The L1 loss between them can be expressed by the following formula:
[0057]
[0058] like Figure 6 As shown, the structure of the region-restricted clothing alignment module proposed in this invention is as follows. E1 and E2 have the same structure, consisting of four 4×4 convolutional layers with double downsampling and two 3×3 convolutional layers. The outputs of E1 and E2 will be matrix-multiplied in the Correlation Layer. The regressor R consists of two 4×4 convolutional layers with double downsampling and three 3×3 convolutional layers. Finally, a set of TPS parameters θ∈R will be output. 50 .
[0059] like Figure 7 , Figure 8 , Figure 9 As shown, the process and structure of the semantic layout alignment fitting-out synthesis module proposed in this invention are as follows. The fitting-out synthesis module merges the aligned clothing and person images obtained in the previous stage to obtain the final fitting-out person image. The input is a clothing-independent representation I. a Posture heatmap P and aligned clothing c w Channel combination I a ⊕P⊕c w Output fitting images Human semantic graph The input is the conditionally normalized SPADEResBlock. The fitting synthesis module consists of a series of stacked SPADEResBlocks and upsampling layers, employing a pure decoder architecture. The input feature map is gradually enlarged to the size of the final output.
[0060] The module first adjusts I a ⊕P⊕c w The size is determined, and then a 3×3 convolutional layer is used to output a feature map of size 1024×4×3. The output feature map will be resized to match the size of the input. a ⊕P⊕c w Channel combination is used as the input to the SPADE ResBlock. The structure of the SPADE ResBlock is similar to that of the Residual Block, consisting of three convolutional layers paired with residual connections. Similarly, the conditionally normalized input to the SPADE layer... It also needs to be adjusted to the same size as the current SPADEResBlock input. The output of SPADEResBlock will be upsampled by a factor of two, and then its channel count will be halved by a 3×3 convolutional layer. The resulting output will be used as the input to the next SPADEResBlock. As can be seen, the model employs a multi-scale fusion strategy, which ensures the normalization effect of feature maps at different scales under the guidance of the semantic map, thereby achieving alignment between the output image and the semantic layout. The module's objective loss function consists of adversarial loss, reconstruction loss (L1 loss), perceptual loss, and the discriminator's feature matching loss:
[0061]
[0062] In the formula λ rec , λ perc and λ FM These are the coefficients for the corresponding losses.
[0063] The following section will use the VITON dataset (a dataset for image-based virtual try-on, containing 16,253 pairs of person-clothing images, of which 14,221 are training sets and 2,032 are test sets) to describe in detail the operating environment, parameter settings, and virtual try-on generation effects of this invention.
[0064] The experimental environment in this invention is as follows. The specific environment configuration is as follows: (1) CPU, AMD Epyc 7542 32 cores 64 threads (2) GPU, NVIDIA GeForce RTX 3090×4 (3) Memory, Samsung DDR4 32GB×8 (4) Operating system, Ubuntu 22.04.1LTS (5) Deep learning environment, Pytorch 1.11. All modules of this invention are optimized using the Adam optimizer, with hyperparameters β1=0.5, β2=0.999, and the learning rate is gradually reduced to 0 using a linear decay strategy. The training batch size of the human semantic synthesis module is 32, the generator learning rate is 2e-4, and the discriminator learning rate is 1e-4. The training batch size of the clothing alignment module is 64, the learning rate is 2e-4, and the coefficients in the loss function are: λ app =1,λ vgg =1,λ shp =0.1. The training batch size of the try-on synthesis module is 32, the generator learning rate is 2e-4, the discriminator learning rate is 2e-5, and the coefficients in the loss function are: λ rec =0.1,λ perc =1,λ FM =10.
[0065] like Figure 10 The image shows the virtual try-on effect of this invention on various garments, including short-sleeved, half-sleeved, long-sleeved, and sleeveless clothing. It can be seen that this invention can allow a model to change into multiple outfits simultaneously and generate sufficiently convincing try-on images.
[0066] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.
Claims
1. A semantic-aligned image virtual fitting method, characterized in that, Includes the following steps: (1) Data Preprocessing: This part aims to erase the clothing information of any target clothing region in the person image I, so as to obtain the clothing-irrelevant expression of the person image I a and the human body semantic map S a , both of which are free of clothing and arm information; (2) Human semantic synthesis module: input is a human semantic graph S without arm and clothing information a and pose graph P and clothing mask c m , the model will generate a target human semantic graph wearing the target clothing ; (3) Garment alignment module: This module warps the garment image c by a non-rigid TPS transformation to align the target human's clothed region, and avoids the misalignment problem of the warped garment with the target human semantic map and also avoids the over-warping of the garment. The specific approach involves using human semantic maps... Extract the target clothing mask Then, with the transformed clothing Perform a dot product to obtain the final output of this module, which is the aligned clothing c. w : ; The target loss is the alignment of the clothing c. w And target clothing I c The L1+ perceptual loss was replaced; a shape loss was introduced to make the module focus on the overall shape of the clothing, rather than just deforming a portion of the target clothing area; specifically, the same set of TPS parameters as the clothing deformation was used to deform the in-store clothing mask c. m , get and Clothing masks with the same shape Then calculate it and the target clothing mask. L1 loss between; Expressed using the following formula: ; Therefore, the overall loss function of the module is defined as: ; (4) Clothing Composition Module: This module receives the person image I, the pose image P, and the aligned clothing. As input, and using the target human semantic map Conditional normalization information is used as a generator to generate the target image.
2. The image virtual try-on method with semantic alignment as described in claim 1, characterized in that, In step (1), given a human image I, human pose maps P∈R are extracted using a pre-trained human pose extractor and a human parsing extractor, respectively. 18 H W and human semantic graph S∈L H W L represents the total number of semantic graph labels; S represents the clothing portion of the human semantic graph S. c Set the coordinates to 0, and then use the coordinates provided by the pose graph P as the center to create a columnar mask to cover the arm part of S; for the human body image I, use the semantic graph S to cover the clothing part, and then use the coordinates provided by the pose graph P as the center to create a columnar mask to expand the mask range so that the mask completely covers the upper body of the human body.
3. The image virtual try-on method with semantic alignment as described in claim 1, characterized in that, In step (2), human semantic synthesis uses U-Net as the generator, and generator G... S Receive human semantic map S a Pose image P and clothing mask c m As input, predict the target human semantic map. A multi-scale discriminator structure is adopted; two discriminators D are used. S1 and D S2 Human semantic maps at original size and twice the downsampled size To make distinctions, the receptive field of the discriminator is increased; The specific method is to obtain and input Channel combination is used as the input to the discriminator. The two discriminators receive the original input and the input after being downsampled by 2 times, respectively. They are then passed through a Sigmoid layer to output confidence scores, and the average of the scores is taken as the final output of the discriminator.
4. The image virtual try-on method with semantic alignment as described in claim 1, characterized in that, In step (4), the input to the clothing synthesis module is an expression indicating that the clothing is unrelated to the person. a Posture diagram P and aligned clothing c w channel combination Output fitting images Human semantic graph The input is a conditionally normalized SPADE ResBlock; the fitting synthesis module consists of a series of stacked SPADE ResBlocks and upsampling layers, using a pure decoder architecture; the input feature map is gradually enlarged to the size of the final output. The module was adjusted first. The size is determined, and then a 3×3 convolutional layer is used to output a feature map of size 1024×4×3; the output feature map will be resized to match its size. Channel combination is used as the input to SPADEResBlock; the SPADEResBlock consists of three convolutional layers with residual connection structures; the conditionally normalized input of the SPADE layer... It needs to be adjusted to the same size as the current SPADEResBlock input; the output of SPADEResBlock will be upsampled by a factor of two, and then its channel count will be halved by a 3×3 convolutional layer. The resulting output will be used as the input of the next SPADEResBlock. The model employs a multi-scale fusion strategy, which ensures the normalization effect of feature maps at different scales under the guidance of semantic maps, thereby achieving alignment between the output image and the semantic layout. The module's objective loss function consists of adversarial loss, reconstruction loss, perceptual loss, and the discriminator's feature matching loss. ; In the formula , and These are the coefficients for the corresponding losses.
Citation Information
Patent Citations
Semantic-based multi-posture virtual fitting method
CN113361560A
Computer angiography imaging synthesis method based on multi-scale discrimination
CN115239674A
