CN113793397B - Garment attribute editing method based on improved WGAN - Google Patents
Garment attribute editing method based on improved WGAN Download PDFInfo
- Publication number
- CN113793397B CN113793397B CN202110871983.6A CN202110871983A CN113793397B CN 113793397 B CN113793397 B CN 113793397B CN 202110871983 A CN202110871983 A CN 202110871983A CN 113793397 B CN113793397 B CN 113793397B
- Authority
- CN
- China
- Prior art keywords
- network
- clothing
- wgan
- garment
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 33
- 230000000694 effects Effects 0.000 claims abstract description 15
- 238000012549 training Methods 0.000 claims description 27
- 230000006870 function Effects 0.000 claims description 19
- 238000012360 testing method Methods 0.000 claims description 8
- 238000010276 construction Methods 0.000 claims description 5
- 238000009826 distribution Methods 0.000 claims description 5
- 230000014509 gene expression Effects 0.000 claims description 5
- 238000002372 labelling Methods 0.000 claims description 5
- 238000004458 analytical method Methods 0.000 claims description 3
- 238000013461 design Methods 0.000 abstract description 14
- 230000008034 disappearance Effects 0.000 abstract description 8
- 238000004880 explosion Methods 0.000 abstract description 4
- 238000013528 artificial neural network Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 238000013473 artificial intelligence Methods 0.000 description 3
- 230000015556 catabolic process Effects 0.000 description 3
- 238000006731 degradation reaction Methods 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 230000011218 segmentation Effects 0.000 description 3
- 208000009119 Giant Axonal Neuropathy Diseases 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 201000003382 giant axonal neuropathy 1 Diseases 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 239000003086 colorant Substances 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000008921 facial expression Effects 0.000 description 1
- 230000014759 maintenance of location Effects 0.000 description 1
- 230000000116 mitigating effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000009958 sewing Methods 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
- G06T11/001—Texturing; Colouring; Generation of texture or colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Abstract
The application belongs to the field of computer vision, and discloses a clothing attribute editing method based on an improved WGAN. The method can generate new clothing attributes based on the existing clothing attributes, and achieves the goal of intelligent clothing aided design. The present application adds a cross-connection residual structure to the producer's decoding network relative to the original WGAN network. The residual structure can promote gradient circulation, merge shallow features and alleviate gradient disappearance; secondly, on the basis of the WGAN original loss function, the L2 norm term is added to the reconstruction loss, and the method has the main advantages that the gradient is smoother, and the problems of gradient disappearance and gradient explosion can be relieved. By optimizing the network structure and the loss function, the WGAN has more excellent effect than before. The method has the characteristics of rapid end-to-end generation, multiple generation attributes and good effect.
Description
Technical Field
The application belongs to the field of computer vision, and particularly relates to a clothing attribute editing method based on an improved WGAN.
Background
Prior to the application of the standard garment size, all garments were made by tailor's fitting to the individual stature. With the development of standard clothing sizes, sewing machines and modern industry and commerce, the time and cost of mass production of clothing is greatly reduced, and ready-made clothing mass-produced by a production line is very popular with consumers in the past decades. This is because of its low price, which can be affordable to the consumer; on the other hand, consumers have more choices because of the variety of styles. However, such garments are typically designed to be standard-sized and in fact it is difficult for most non-standard-sized consumers to find a fully fitted, satisfactory garment.
The modern society economy is rapidly increased, the clothing related industry is rapidly developed, the wearing requirements of people are higher and higher, besides the basic functional requirements of protection, comfort, warmth retention and the like in the past, people have requirements on fashion sense, uniqueness and the like of clothing styles gradually, and the clothing is suitable for the heart instrument clothing of the people to bring confidence and satisfaction to the people, so that on the basis of the selection standard of the clothing, the clothing is gradually changed from comfort and attractive popular standard to personalized design comprehensively considering personal stature characteristics, gas quality characteristics and aesthetic preferences based on fit, and personalized customization and consumption of the clothing of the user become future trends and trends.
Artificial intelligence has evolved rapidly in recent years and has also begun to gradually serve the apparel industry. At present, artificial intelligence technology has been explored in the aspects of detection and segmentation processing of fashion images, virtual fitting, clothing recommendation, fashion trend prediction and the like. However, there has been little research into the design of properties of garments themselves. In recent years, the generation of the development of the countermeasure network technology brings a lot of inspires to the design of clothing attribute in the field of facial expression change and the like. Therefore, the application will explore intelligent design technology of clothing based on deep learning to improve clothing properties.
The traditional clothing component design method is based on the inspiration of a designer or auxiliary material library for collocation design, and completely depends on the design level of the designer. The academic community is also researching and applying artificial intelligence technology to conduct intelligent aided design, and the primary task is to establish a high-quality clothing data set. The deep fashion garment dataset was established by the university of hong Kong multimedia laboratory team Shang Xiaoou. The data set contains labels of the attribute of the clothing, and data support is provided for training a deep learning algorithm to conduct intelligent clothing design. Fashionpedia Dataset is a garment attribute segmentation dataset established by the university of kannel and Google, comprising 48, 825 total garment images, and finely subdividing the garment item and providing fine-grained attributes for the subdivided categories. The data set unifies the example segmentation and the visual attribute identification, and lays a solid foundation for structural understanding of clothing attributes. Based on the data set, wei-Lin Hsiao of UT Austin proposes a Fashion++ clothing design network which uses texture and shape encoders to extract clothing corresponding attribute features and combines the same, and the algorithm uses a generated neural network to learn clothing attribute codes to synthesize clothing. And then the implicit code is explicitly decomposed according to the shape and the texture, so that the attribute editing of the clothing is realized. The algorithm mainly focuses on the editing of the overall shape of the garment, and does not have the editing of local properties such as real sleeves, collars and the like. Ping Q et al propose a Fashion-AttGAN algorithm that focuses on editing properties such as collar, sleeve length, etc. of the coat, which encodes some properties and then uses AttGAN for training generation. The algorithm opens up new possibilities for user-driven apparel design and may be beneficial for virtual try-on, apparel recommendation, visual search, and the like. But the current algorithm generation effect still needs to be improved.
Disclosure of Invention
The present application addresses the above-described problems by providing a garment property editing method based on an improved WGAN. The method can generate new clothing attributes based on the existing clothing attributes, and achieves the goal of intelligent clothing aided design.
In order to solve the technical problems, the specific technical scheme of the clothing attribute editing method based on the improved WGAN is as follows:
the clothing attribute editing method based on the improved WGAN is divided into a training stage and a testing stage, wherein the training stage is optimized by adopting a supervised learning mode; the testing stage adopts the converged network to generate clothing attribute; the method specifically comprises the following steps:
step 1: in the training stage, firstly, acquiring clothing construction clothing data sets of various styles, and then, labeling attributes of each clothing sample; the attributes comprise three primary granularity attributes of sleeve length, color and collar;
step 2: training data with attribute labels is input to generate an countermeasure network for training.
Further, the generating countermeasure network in the step 2 consists of a generating network and a judging network; the generating network comprises a generator and a decoder for learning attribute features of the garment from training data, reconstructing input garment data based on the learned feature expressions on a normal distribution; the distinguishing network consists of a decoder and a distinguishing device; the decoder of the judging network shares weight with the decoder of the generating network, and the decoder of the judging network reconstructs false clothing data based on the generating network and then inputs the false clothing data and the original data label to the judging device for analysis and judgment; the generating network and the discriminating network maintain Nash equilibrium.
Further, the three primary granularity attributes of the sleeve length, the color and the collar comprise 14 fine granularity attributes of a high collar, a V-shaped collar, a round collar, red, orange, yellow, green, cyan, blue, purple, long sleeves, half sleeves, short sleeves, sleeveless sleeves and the like.
Further, the step 1 comprises the following specific steps:
firstly, collecting clothing construction clothing data sets X of various styles, and then labeling attributes of each clothing sample; the collar attribute labels are divided into a high collar, a V-shaped collar and a round collar; color attributes are marked as red, orange, yellow, green, cyan, blue and purple; the sleeve length attribute is marked as long sleeve, half sleeve, short sleeve and no sleeve.
Further, the step 2 comprises the following specific steps:
training data X with labels 0 Inputting a generated countermeasure network Z for training; the generating countermeasure network Z consists of a generating network and a judging network; the generating network comprises a generator Genc and a decoder Gdec for learning attribute features of the garment from the training data and then reconstructing the input garment data X based on the learned feature expressions over a normal distribution 0 Obtaining a reconstructed picture X 1 The method comprises the steps of carrying out a first treatment on the surface of the The generation network calculates a reconstruction loss L based on the reconstructed picture and the original input sample, and the reconstruction loss function is shown as follows:
wherein ,
the generating network optimizes the reconstruction effect of the generating network by minimizing the reconstruction loss L;
the discrimination network consists of a decoder Gdec, a discriminator C and a discriminator D; the decoder Gdec of the discriminating network shares weights with the decoder Gdec of the generating network, which reconstructs false clothing data X based on the generating network Z 2 Then X is taken up 2 Raw data X 0 The corresponding label is input to the discriminator D for discrimination.
Further, the discriminator D learns the features of the input data to discriminate that the input picture is a true sample X 0 Or reconstructed dummy sample X 2 Then optimizing through a cross entropy loss function; the discriminator C discriminates the input dummy sample data X with finer granularity 2 Which classes of attributes are generated and then compared with the real sample X 0 Attribute comparison of (2); and the attribute generated by the generator is approximated to the real sample by optimizing the cross entropy loss function.
Further, the discriminator D uses a coarse grain discriminator to discriminate true and false generated data, and the discriminator C uses a fine grain discriminator to discriminate false sample attributes.
Further, a cross-connection residual structure is added to the decoding network of the generator on a WAGN basis.
Further, an integral loss function is established, wherein the integral loss function is formed by reconstructing loss L, true and false discrimination loss0 and attribute loss att loss Three parts are composed, and the following formula is shown:
L all ＝L+att loss +loss0
further, the weight of the generated network Z is frozen, then a clothing picture without an attribute tag is input, the network is improved to generate a new clothing style with new attribute, and the intelligent clothing editing target is realized.
The clothing attribute editing method based on the improved WGAN has the following advantages: the present application adds a cross-connection residual structure to the producer's decoding network relative to the original WGAN network. The residual structure can promote gradient circulation, merge shallow features and alleviate gradient disappearance; secondly, on the basis of the WGAN original loss function, the L2 norm term is added to the reconstruction loss, and the method has the main advantages that the gradient is smoother, and the problems of gradient disappearance and gradient explosion can be relieved. By optimizing the network structure and the loss function, the WGAN has more excellent effect than before. Therefore, the method has the characteristics of rapid end-to-end generation, multiple generation attributes and good effect.
Drawings
FIG. 1 is a flow chart of a method for garment property editing based on an improved WGAN of the present application;
fig. 2 is a WGAN generation result diagram;
fig. 3 is a graph of the results of the improved WGAN generation of the present application;
FIG. 4 is a graph of DCGAN generation results;
fig. 5 is a diagram of LSGAN generation results.
Detailed Description
For a better understanding of the objects, structures and functions of the present application, a method for tailoring properties based on an improved WGAN will be described in further detail with reference to the accompanying drawings.
1. Experimental environment and data set
The computer hardware and software configuration adopts Ubuntu 64-bit system, and the processor is i7 8700k, memory 24G and display card 1070ti. The software environment is PyTorch 1.4.
The dataset contained 14,221 garment images, three primary granularity attributes of sleeve length, collar, color, 14 fine granularity attributes using CV-PTON.
2. The specific implementation steps are as follows:
as shown in fig. 1, the garment attribute editing method based on the improved WGAN of the present application is divided into a training phase and a testing phase, wherein the training phase is optimized by adopting a supervised learning mode. And in the test stage, the converged network is adopted for clothing attribute generation.
Particularly, the algorithm designs three primary granularity attributes of sleeve length, color and collar, and 14 fine granularity attributes of high collar, V-shaped collar, round collar, red, orange, yellow, green, blue, purple, long sleeve, half sleeve, short sleeve, sleeveless and the like;
in the training stage, firstly, a clothing construction clothing data set X of various styles is collected, and then, attribute labeling is carried out on each clothing sample. The collar attribute labels are divided into { high collar, V-shaped collar and round collar }; color attributes are labeled { red, orange, yellow, green, cyan, blue, purple }; the sleeve length attribute is marked as { long sleeve, half sleeve, short sleeve, no sleeve }.
Training data X with labels 0 The input generation antagonizes the network Z for training. The generating countermeasure network Z consists of a generating network and a distinguishing network. The generating network comprises a generator Genc and a decoder Gdec for learning attribute features of the garment from the training data and then reconstructing the input garment data X based on the learned feature expressions over a normal distribution 0 Obtaining a reconstructed picture X 1 . The generation network calculates a reconstruction loss L based on the reconstructed picture and the original input samples. The reconstruction loss function is shown as follows:
wherein ,
the generating network optimizes the reconstruction effect of the generating network by minimizing the reconstruction loss L.
The discrimination network is composed of a decoder Gdec and a discriminator C, D. In particular, the decoder Gdec of the discriminating network shares weights with the decoder Gdec of the generating network, which reconstructs false garment data X based on the generating network Z 2 . Then X is taken up 2 Raw data X 0 The corresponding label is input to the discriminator D for discrimination.
In particular, the discriminator D learns the features of the input data to discriminate that the input picture is a true sample X 0 Or reconstructed dummy sample X 2 Then optimized by a cross entropy loss function. Further, the arbiter C discriminates the input dummy sample data X with finer granularity 2 Which classes of attributes are generated and then compared with the real sample X 0 Is a comparison of attributes of (a). And the attribute generated by the generator is approximated to the real sample by optimizing the cross entropy loss function. Network training is performedThe process is that the generating network generates false pictures close to real data as much as possible, and the judging network distinguishes the false pictures generated by the generating network as much as possible, and the generating network and the false pictures maintain Nash balance.
In particular, the application adopts a coarse grain discriminator D to discriminate the true and false of the generated data, and then adopts a fine grain discriminator C to discriminate the false sample attribute. The method effectively improves the discrimination effect of the discriminator and further improves the generation effect of the whole algorithm.
In particular, in terms of network structure, the present application adds a cross-connection residual structure to the decoding network of the generator on a WAGN basis. The residual structure can promote gradient fluxion, fuse shallow layer characteristics, and alleviate gradient disappearance problem.
The integral loss function of the application is formed by reconstructing loss L, true and false discrimination loss0 and attribute loss att loss Three parts. The following formula is shown:
L all ＝L+att loss +loSS0
in the testing stage, the weight of the generated network Z is frozen, then the clothing picture without the attribute tag is input, the new clothing style with new attribute can be generated by improving the network, the intelligent editing target of the clothing is realized, and the improved network generates new clothing with long-sleeve attribute based on the input short-sleeve picture as shown in fig. 1.
In particular, the application has the characteristics of rapid end-to-end generation, multiple generation attributes and good effect.
3. Test results and analysis
(1) Contrast with WGAN
On the CV-PTON dataset, the loss of WGAN was 10%, whereas the improved WGAN of the present application can reduce the loss to 5%. Fig. 2 and 3 show the results of the creation of WAGN and the improved WGAN for garment property editing. Each row in the image is true value, reconstruction, sleeveless, short sleeve, long sleeve, high collar, POLO collar/T shirt collar, V collar, red, gray, black, white, beige, navy, blue and green.
Overall, the improved WGAN produced a clearer edge and better color effect. For example, for the "cotta" column, there are many cases in WAGN that create incomplete sleeves, which rarely occurs in the results of improving WAGN; in addition to shape attributes, color attributes are also more effective in improving the production of WGANs, such as for the last column "green" the green produced by the improved WGAN is more deeply colored, while the WGAN is more affected by the original color of the garment. Table 1 shows the WGAN generation result color control, and table 2 shows the improvement of the WGAN generation result color control.
Table 1: WGAN generation result color comparison table
Table 2: improved WGAN generation result color comparison table
| True value | Red color | Gray color | Black color | White color | Beige color | Navy blue | Blue color | Green colour |
| #462726 | #372e35 | #9d999d | #1f1f21 | #f7f7f9 | #3d2c30 | #282f42 | #42608d | #444639 |
| #8496ab | #c63c39 | #a7a6ab | #141519 | #f7f7f7 | #937373 | #222c3f | #b16768 | #748170 |
| #151517 | #7c1f20 | #e0d7df | #4e485a | #f9f9fb | #2b232a | #16202b | #436cac | #36463f |
| #615f54 | #b73a33 | #828682 | #101015 | #f6f6f6 | #765f52 | #2a3846 | #478090 | #7e5951 |
| #1f2223 | #802524 | #b7b3b5 | #595150 | #f4f2f5 | #4b4143 | #2c3446 | #2b4f82 | #24332a |
| #292a3e | #6a1f24 | #9a96a8 | #1b1e23 | #f2f0f6 | #342d39 | #382b34 | #2e5280 | #324245 |
| #36322f | #6c201f | #8a9290 | #44423e | #f0f0f0 | #423b36 | #2e3847 | #3a6479 | #405747 |
| #60607d | #ad4e6b | #9f95ab | #3a3442 | #fbfbff | #877586 | #444459 | #65869d | #515151 |
| #71696f | #c23b3d | #80484f | #17181c | #f4f4f6 | #805c5f | #263142 | #4e85a1 | #687262 |
| #323746 | #77282f | #5f6268 | #202126 | #f2f2f4 | #3f3d49 | #483f46 | #305c86 | #374646 |
(2) In contrast to other GANs
As shown in fig. 4, the DCGAN is generated. It can be seen that DCGAN does not substantially modify the garment shape properties, and in terms of color modification, can only color a small amount on the basis of the original garment, and cannot make the garment adhere to vivid colors as in the improvement of WAGN. As shown in table 3, the DCGAN generated result color control.
Table 3: DCGAN generation result color comparison table
As shown in fig. 5, while for LSGAN, it can be seen that the shape properties of the garment are slightly affected but not well modified, it can be said that the modification of the garment shape by LSGAN is as ineffective as DCGAN. LSGAN is also poorly colored with respect to color. Thus, comparison with other GANs further verifies the effectiveness of the improved WAGN. As shown in table 4, LSGAN generated result color control.
Table 4: LSGAN generation result color comparison table
| True value | Red color | Gray color | Black color | White color | Beige color | Navy blue | Blue color | Green colour |
| #462726 | #4a292d | #482b2d | #3e2829 | #f3f0eb | #4d2a2d | #37222a | #35222b | #3a2624 |
| #8496ab | #766b7e | #8595a9 | #5d6778 | #cfd4d5 | #929ab0 | #7988a0 | #8c99aa | #748994 |
| #151517 | #140c10 | #101317 | #1d2325 | #89867d | #1b1a20 | #15151a | #171823 | #131416 |
| #615f54 | #765853 | #5c5a54 | #5a5752 | #d1d0cc | #625f55 | #5e5b53 | #585659 | #6b635e |
| #1f2223 | #2b2225 | #24292b | #383a3c | #f4f6f4 | #27272d | #1f232b | #20262f | #1f2526 |
| #292a3e | #352239 | #2b2b43 | #2b2c44 | #e8e7eb | #29293f | #2a283e | #242c4f | #2a3042 |
| #36322f | #42282d | #383232 | #4d4543 | #e9ebe5 | #342b2d | #302b31 | #333137 | #373b37 |
| #60607d | #86617a | #636978 | #545265 | #cecbd1 | #6c677a | #4c4d69 | #929dbc | #66747e |
| #71696f | #80636a | #786e76 | #736d73 | #dddeda | #676168 | #726879 | #767588 | #717171 |
| #323746 | #524153 | #383f4c | #373c4d | #b4b8bc | #474a5a | #414654 | #2f3956 | #35404f |
The present application adds a cross-connection residual structure to the producer's decoding network relative to the original WGAN network. The residual structure can promote gradient circulation, merge shallow features and alleviate gradient disappearance; the advantages are mainly as follows:
(1) Simplifying learning process
If the residual structure does not exist, the neural network learns the original signal, and the residual structure enables the neural network to learn the difference value of the signal, so that the learning process is simplified. The application proves that the residual structure is an effective network optimization mode and has good effect on relieving the problems of gradient disappearance and gradient explosion.
(2) Mitigating network degradation
As the number of network layers increases, network degradation problems occur, mainly due to the symmetry of the network. And the residual connection can break the symmetry of the network, so that the problem of network degradation is solved.
(3) Enhancing generalization capability of a network
The deep network comprising the residual structure can be regarded as the synthesis of shallow neural networks with different depths, and the removal of a certain network layer does not have great influence on the network performance, namely, the residual structure can enhance the generalization capability of the network.
Secondly, on the basis of the WGAN original loss function, the application increases the L2 norm term on the reconstruction loss. The method has the main advantages that the gradient is smoother, and the problems of gradient disappearance and gradient explosion can be relieved. When an abnormally large value of a certain weight in the network appears, the abnormal value is amplified by the square in the two norms, and the abnormal value can be forced to be reduced in the back propagation process. In addition, the L2 norm term can prevent the generation of overfitting.
By optimizing the network structure and the loss function, the WGAN has more excellent effect than before.
In summary, the application has the characteristics of rapid end-to-end generation, multiple generation attributes and good effect.
It will be understood that the application has been described in terms of several embodiments, and that various changes and equivalents may be made to these features and embodiments by those skilled in the art without departing from the spirit and scope of the application. In addition, many modifications may be made to adapt a particular situation or material to the teachings of the application without departing from the essential scope thereof. Therefore, it is intended that the application not be limited to the particular embodiment disclosed, but that the application will include all embodiments falling within the scope of the appended claims.
Claims (5)
1. The clothing attribute editing method based on the improved WGAN is divided into a training stage and a testing stage, wherein the training stage is optimized by adopting a supervised learning mode; the testing stage adopts the converged network to generate clothing attribute; the method is characterized by comprising the following steps of:
step 1: in the training stage, firstly, acquiring clothing construction clothing data sets of various styles, and then, labeling attributes of each clothing sample; the attributes comprise three primary granularity attributes of sleeve length, color and collar;
step 2: inputting training data with attribute labels to generate an countermeasure network for training;
training data X with labels 0 Inputting a generated countermeasure network Z for training; the generating countermeasure network Z consists of a generating network and a judging network; the generating network comprises a generator Genc and a decoder Gdec for learning attribute features of the garment from the training data and then reconstructing the input garment data X based on the learned feature expressions over a normal distribution 0 Obtaining a reconstructed picture X 1 The method comprises the steps of carrying out a first treatment on the surface of the The generation network calculates a reconstruction loss L based on the reconstructed picture and the original input sample, and the reconstruction loss function is shown as follows:
wherein ,，
the generating network optimizes the reconstruction effect of the generating network by minimizing the reconstruction loss L;
the discrimination network is composed of decoder Gdec and discriminationA discriminator C and a discriminator D; the decoder Gdec of the discriminating network shares weights with the decoder Gdec of the generating network, which reconstructs false clothing data X based on the generating network Z 2 Then X is taken up 2 Raw data X 0 Inputting the corresponding label to the discriminator D for discrimination;
the discriminator D learns the characteristics of the input data to discriminate that the input picture is a real sample X 0 Or reconstructed dummy sample X 2 Then optimizing through a cross entropy loss function; the discriminator C discriminates the input dummy sample data X with finer granularity 2 Which classes of attributes are generated and then compared with the real sample X 0 Attribute comparison of (2); optimizing through a cross entropy loss function, so that the attribute generated by the generator approximates to a real sample;
the discriminator D adopts a coarse grain discriminator to discriminate true and false generated data, and the discriminator C adopts a fine grain discriminator to discriminate false sample attributes;
adding a cross-connection residual structure to a decoding network of the generator on the basis of WAGN;
establishing an integral loss function, wherein the integral loss function is formed by reconstructing loss L, true and false discrimination loss0 and attribute loss att loss Three parts are composed, and the following formula is shown:
。
2. the garment property editing method based on improved WGAN as claimed in claim 1, wherein the generating countermeasure network in step 2 is composed of two parts of generating network and discriminating network; the generating network comprises a generator and a decoder for learning attribute features of the garment from training data, reconstructing input garment data based on the learned feature expressions on a normal distribution; the distinguishing network consists of a decoder and a distinguishing device; the decoder of the judging network shares weight with the decoder of the generating network, and the decoder of the judging network reconstructs false clothing data based on the generating network and then inputs the false clothing data and the original data label to the judging device for analysis and judgment; the generating network and the discriminating network maintain Nash equilibrium.
3. The improved WGAN-based garment property editing method of claim 1, wherein the three primary granularity properties of sleeve length, color, collar include 14 granularity properties of high collar, V-collar, round collar, red, orange, yellow, green, cyan, blue, violet, long sleeve, half sleeve, short sleeve, sleeveless.
4. The garment property editing method based on improved WGAN as claimed in claim 3, wherein said step 1 comprises the specific steps of:
firstly, collecting clothing construction clothing data sets X of various styles, and then labeling attributes of each clothing sample; the collar attribute labels are divided into a high collar, a V-shaped collar and a round collar; color attributes are marked as red, orange, yellow, green, cyan, blue and purple; the sleeve length attribute is marked as long sleeve, half sleeve, short sleeve and no sleeve.
5. The garment property editing method based on the improved WGAN as claimed in claim 1, wherein the weight of the generated network Z is frozen, then a garment picture without a property label is input, the improved network generates a new garment style with a new property, and the intelligent garment editing goal is achieved.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110871983.6A CN113793397B (en) | 2021-07-30 | 2021-07-30 | Garment attribute editing method based on improved WGAN |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110871983.6A CN113793397B (en) | 2021-07-30 | 2021-07-30 | Garment attribute editing method based on improved WGAN |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113793397A CN113793397A (en) | 2021-12-14 |
| CN113793397B true CN113793397B (en) | 2023-11-03 |
Family
ID=79181455
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110871983.6A Active CN113793397B (en) | 2021-07-30 | 2021-07-30 | Garment attribute editing method based on improved WGAN |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113793397B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114694012B (en) * | 2022-04-07 | 2023-10-20 | 哈尔滨理工大学 | Clothing identification method based on convolutional neural network |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110189253A (en) * | 2019-04-16 | 2019-08-30 | 浙江工业大学 | A kind of image super-resolution rebuilding method generating confrontation network based on improvement |
| CN111798369A (en) * | 2020-06-29 | 2020-10-20 | 电子科技大学 | Face aging image synthesis method for generating confrontation network based on circulation condition |
| CN111932444A (en) * | 2020-07-16 | 2020-11-13 | 中国石油大学(华东) | Face attribute editing method based on generation countermeasure network and information processing terminal |
| CN112529768A (en) * | 2020-12-04 | 2021-03-19 | 中山大学 | Garment editing and generating method based on generation countermeasure network |
| CN112837247A (en) * | 2021-04-06 | 2021-05-25 | 哈尔滨理工大学 | GAN image denoising algorithm fusing improved residual error network |
-
2021
- 2021-07-30 CN CN202110871983.6A patent/CN113793397B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110189253A (en) * | 2019-04-16 | 2019-08-30 | 浙江工业大学 | A kind of image super-resolution rebuilding method generating confrontation network based on improvement |
| CN111798369A (en) * | 2020-06-29 | 2020-10-20 | 电子科技大学 | Face aging image synthesis method for generating confrontation network based on circulation condition |
| CN111932444A (en) * | 2020-07-16 | 2020-11-13 | 中国石油大学(华东) | Face attribute editing method based on generation countermeasure network and information processing terminal |
| CN112529768A (en) * | 2020-12-04 | 2021-03-19 | 中山大学 | Garment editing and generating method based on generation countermeasure network |
| CN112837247A (en) * | 2021-04-06 | 2021-05-25 | 哈尔滨理工大学 | GAN image denoising algorithm fusing improved residual error network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113793397A (en) | 2021-12-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Lin et al. | Zstgan: An adversarial approach for unsupervised zero-shot image-to-image translation | |
| Matzen et al. | Streetstyle: Exploring world-wide clothing styles from millions of photos | |
| JP6890720B2 (en) | Image processing system | |
| US8983142B1 (en) | Programmatic silhouette attribute determination | |
| CN110909754B (en) | Attribute generation countermeasure network and matching clothing generation method based on same | |
| Yuan et al. | Attribute-aware generative design with generative adversarial networks | |
| CN112100908B (en) | Clothing design method for generating countermeasure network based on multi-condition deep convolution | |
| KR102045017B1 (en) | apparatus and method for generating designs based on images and text | |
| Xiang et al. | Clothing attribute recognition based on RCNN framework using L-Softmax loss | |
| Huang et al. | Clothing landmark detection using deep networks with prior of key point associations | |
| Zhao et al. | Clothing cosegmentation for shopping images with cluttered background | |
| Cychnerski et al. | Clothes detection and classification using convolutional neural networks | |
| Sattar et al. | Fashion is taking shape: Understanding clothing preference based on body shape from online sources | |
| CN113793397B (en) | Garment attribute editing method based on improved WGAN | |
| Yu et al. | Inpainting-based virtual try-on network for selective garment transfer | |
| KR20190128143A (en) | Apparatus and method for generating designs based on image and text | |
| US20200143454A1 (en) | Computer vision based methods and systems of universal fashion ontology fashion rating and recommendation | |
| Zhan et al. | Pose-normalized and appearance-preserved street-to-shop clothing image generation and feature learning | |
| CN111581410B (en) | Image retrieval method, device, medium and system thereof | |
| Sapna et al. | Recommendence and fashionsence: Online fashion advisor for offline experience | |
| Zhang et al. | Warpclothingout: A stepwise framework for clothes translation from the human body to tiled images | |
| Zhao et al. | Private‐encoder: Enforcing privacy in latent space for human face images | |
| Tango et al. | Anime-to-real clothing: Cosplay costume generation via image-to-image translation | |
| CN112446340A (en) | Pedestrian searching method and system combining pedestrian local feature and decorative attribute classification and storage medium | |
| Li et al. | Semi-supervised compatibility learning across categories for clothing matching |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |