Intelligent generation method of mixed style clothing pattern based on deep learning

Abstract

The invention discloses a deep learning-based mixed style clothing pattern intelligent generation method, which comprises the following steps of S1, collecting clothing sample data with patterns, and obtaining initial pattern data; s2, detecting and cutting target pattern data in the initial pattern data and preprocessing the target pattern data to obtain processed original pattern data; s3, carrying out pattern and background labeling on the original pattern data; s4, training a pattern segmentation model and a rapid style migration model; s5, inputting the original pattern data into the pattern segmentation model to obtain a mask pattern; and S6, inputting the original pattern data and the mask patterns into the rapid style migration models of different styles together, and selecting different mask patterns to mask different element positions to obtain patterns of different styles or mixed styles. The method has the advantages that concrete patterns are efficiently generated, and the requirement for style diversification is met.

Description

Technical Field

[0001] This invention relates to a method for intelligent pattern generation, and more particularly to a method for intelligent generation of hybrid style clothing patterns based on deep learning. Background Technology

[0002] Clothing patterns are unique textures and shapes on the surface of garments, combining decorative and practical functions. Traditional clothing pattern design relies on experienced designers using various design software and hand-drawing techniques. While this produces high-quality patterns, it requires a significant amount of design time and has a low conversion rate. The digital age demands that pattern design be more accessible, efficient, and diverse. Therefore, finding efficient and low-cost ways to acquire clothing patterns and provide inspiration for pattern designers has become a critical bottleneck that needs to be overcome.

[0003] Many scholars have studied clothing patterns or related patterns, but most studies focus on pattern generation based on function design. For example, Chinese patent CN105512446A, published on 2016-04-20, entitled "A Method for Generating Butterfly Patterns Based on Iterative Functions," discloses that "the generation of butterfly patterns is divided into three steps: basic elements, iteration, and whisker generation using mathematical functions, and finally, a pattern with fractal characteristics is automatically generated under the mathematical model of iterative functions." Chinese patent CN114332268A, published on 2022-09-16, entitled "A Method for Generating Tu Ethnic Embroidery Patterns Based on L-System," discloses that "Tu ethnic embroidery pattern rules are automatically generated based on L-System. According to the rules of the patterns, corresponding position constraint functions are defined, and the generated various rule patterns are combined to automatically generate Tu ethnic embroidery patterns." While the methods described above can all automatically generate patterns and produce targeted designs, they all suffer from the following shortcomings: ① They require prior function design, which consumes significant time in function selection and design, resulting in low pattern generation efficiency and high cost. ② The methods have limited generalization ability and cannot generate all patterns. ③ Most studies focus on abstract patterns, making it difficult to automatically and quickly generate representational patterns.

[0004] The clothing style transfer system and method disclosed in publication number CN113160033A, published on July 23, 2021, includes "an image information processing unit, an image feature extraction unit, a style transfer generation unit, and a generated image enhancement unit; the image information processing unit is used to select and collect images of clothing of the target style and clothing of the style to be converted; the image feature extraction unit is used to process the style features of the target clothing and its own style features; the style transfer generation unit is used to perform corresponding style transfer and fusion on the clothing of the style to be converted; the generated image enhancement unit is used to generate a semantic segmentation map of the clothing image of the style to be converted, and use the semantic segmentation map to separate the deformed background and restore the background of the transferred image. Using the style of one type of clothing, a new style of clothing defined by the shape and outline of the other type is rendered and generated."

[0005] Although the aforementioned documents disclose algorithms for style transfer and semantic segmentation and their application in the clothing field, the semantic segmentation algorithm in the documents is used to segment clothing outlines, and the style transfer algorithm is used to transfer clothing styles. The algorithms are relatively complex, have a large amount of computation, long processing time, and low efficiency. Moreover, the pattern in the aforementioned documents can only obtain one style by transferring the style from A to B, and cannot obtain a pattern with mixed styles, which cannot meet the problem of diverse design needs of clothing styles. Summary of the Invention

[0006] The purpose of this invention is to provide a method for intelligently generating hybrid style clothing patterns based on deep learning. This invention features efficient generation of representational patterns, meeting diverse style requirements.

[0007] The technical solution of this invention: A method for intelligently generating hybrid style clothing patterns based on deep learning, comprising the following steps:

[0008] S1. Collect patterned clothing sample data to obtain initial pattern data;

[0009] S2. Detect the target pattern data in the initial pattern data and preprocess it to obtain the processed original pattern data;

[0010] S3. Label the original pattern data for the pattern itself and the background, and determine the pattern area that needs to be segmented;

[0011] S4. Use semantic segmentation algorithm to train pattern segmentation model, and use fast style transfer algorithm to train fast style transfer models of different styles.

[0012] S5. Input the original pattern data into the trained pattern segmentation model to obtain the mask image, and enhance the brightness and contrast of the mask image.

[0013] S6. Combine the original pattern data and mask. Figure 1 The input is fed into a pre-trained fast style transfer model with different styles, and different mask images are selected to mask different element positions to obtain patterns of different styles or mixed styles.

[0014] In the aforementioned method for intelligent generation of hybrid style clothing patterns based on deep learning, in step S2, the detection and cropping of the target pattern is based on the YOLOv3 pattern detection algorithm. The pattern detection model is trained using YOLOv3, and the clothing pattern is cropped using the pattern detection model. The preprocessing includes sharpening to improve the clarity of the pattern outline.

[0015] In the aforementioned method for intelligent generation of hybrid style clothing patterns based on deep learning, in step S5, the mask image includes pattern elements and background elements of the target transfer.

[0016] In the aforementioned method for intelligent generation of hybrid style clothing patterns based on deep learning, in step S4, the semantic segmentation algorithm adopts the DeepLabV3+ algorithm, and the backbone feature extraction network of the DeepLabV3+ algorithm adopts MobileNetV2.

[0017] In the aforementioned method for intelligent generation of hybrid style clothing patterns based on deep learning, in step S4, the loss network of the fast style transfer model uses a VGG16 model with the fully connected layers of VGG16 removed and the convolutional layers retained.

[0018] In the aforementioned method for intelligent generation of hybrid style clothing patterns based on deep learning, step S4 includes a fast style transfer model comprising an ink painting style transfer model, an oil painting style transfer model, and a pixel style transfer model.

[0019] In the aforementioned method for intelligent generation of hybrid style clothing patterns based on deep learning, the ink painting style transfer model includes content loss and style loss, with a content loss weight value α = 1e5 and a style loss weight value β = 5e10.

[0020] In the aforementioned method for intelligent generation of hybrid style clothing patterns based on deep learning, the oil painting style transfer model includes content loss and style loss, with a content loss weight value α = 1e5 and a style loss weight value β = 2e10.

[0021] In the aforementioned method for intelligent generation of hybrid style clothing patterns based on deep learning, the pixel style transfer model includes content loss and style loss, with a content loss weight value α = 1e5 and a style loss weight value β = 1e10.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] This invention utilizes a YOLOv3-based pattern detection and cropping algorithm to extract figurative patterns from clothing. By combining two deep learning algorithms—semantic segmentation and fast style transfer—a pattern segmentation model and several different fast style transfer models are trained. The pattern is input into the pattern segmentation model to obtain a mask image. The pattern and mask image are then input into the fast style transfer model, and different mask images are selected to mask different element positions, resulting in clothing patterns of different styles. This method boasts high processing efficiency and strong generalization performance. The semantic segmentation algorithm is used to segment the outline and background of figurative patterns into different blocks, employing the lighter MobilNetV2 network for faster processing. The fast style transfer algorithm is used for style transformation of modular patterns, preserving the original content while changing the style to obtain different style models. This ensures that the main body of the figurative pattern is preserved without distortion, achieving the goal of rapid pattern transfer. It also produces better generation results for figurative patterns.

[0024] By pre-training the fast style transfer algorithm, different style models can be obtained, which facilitates rapid transfer of large batches of styles when called later.

[0025] By selecting and recombining different models, styles can be recombined and combined, resulting in mixed-style patterns and single-style multi-style patterns. The generated patterns are more diverse and have higher variety, improving the generation quality of complex and varied figurative patterns and meeting the diverse design needs of clothing styles.

[0026] This invention provides design options, allowing for the training of different models based on changes in required patterns. It also incorporates interactive design to provide designers with inspiration or pattern design, significantly improving pattern design efficiency and enhancing the interactivity between clothing pattern design and customers. This meets the demands of the digital age for pattern design to be popularized, efficient, diverse, and low-cost.

[0027] This invention uses three styles as the base styles, and selects different styles to recombine them when generating the final pattern, thus achieving a mixed effect of multiple styles. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the generation process of Embodiment 1 of the present invention;

[0029] Figure 2 This is a schematic diagram of the pattern data extraction process in Embodiment 1 of the present invention;

[0030] Figure 3 This is a schematic diagram of the style transfer model structure for clothing patterns in Example 1;

[0031] Figure 4These are style transfer results for three style loss parameters;

[0032] Figure 5 This is a graph showing the subjective qualitative evaluation scoring results;

[0033] Figure 6 This is a style transfer result diagram of Example 2;

[0034] Figure 7 This is a style transfer result diagram of Example 3. Detailed Implementation

[0035] The present invention will be further described below with reference to embodiments, but these embodiments are not intended to limit the scope of the invention.

[0036] Example 1:

[0037] like Figure 1 As shown, a method for intelligently generating hybrid style clothing patterns based on deep learning, taking cat-themed clothing patterns as an example, includes the following steps:

[0038] S1. Data Acquisition:

[0039] S101. Collect clothing sample data with cat-like patterns through web crawling, text retrieval, and Kaggle to obtain initial pattern data;

[0040] S2. Extract pattern data:

[0041] S201. A pattern detection algorithm based on YOLOv3 is used to train a pattern detection model. The pattern detection model is used to organize, detect and cut out clothing patterns, and extract cat pattern data from the initial pattern data.

[0042] S202. The cat image data is sharpened and standardized using the Unsharp Mask (USM) sharpening algorithm to improve the clarity of the image outline, resulting in high-resolution original cat image data. The specific process is as follows: Figure 2 As shown,

[0043] S3: Annotate the pattern data:

[0044] High-resolution original cat image data were labeled using Labelme to mark the outline and background of the cat images, separating the cat images themselves from the background and determining the image regions that need to be segmented for subsequent semantic segmentation.

[0045] S401. Constructing a DeepLabV3+ clothing pattern segmentation model:

[0046] A semantic segmentation algorithm is used to train a clothing pattern segmentation model, resulting in a well-trained model. The model is constructed as follows: the DeepLabV3+ algorithm is used as the semantic segmentation algorithm. The high accuracy of DeepLabV3+ in segmenting images is utilized to segment the constituent elements of the clothing pattern, obtaining mask images of the relevant constituent elements. This prepares for subsequent style transfer of various constituent elements.

[0047] The backbone feature extraction network of the DeepLabV3+ algorithm uses MobileNetV2.

[0048] MobileNetV2 replaces ordinary convolution with depthwise separable convolution, combining depthwise separable convolution with an inverse residual unit structure. This significantly reduces the computational parameters of the network while maintaining high accuracy, meeting the needs of fast and accurate detection and segmentation when the model is embedded in the system.

[0049] In the initial data processing, 1500 labeled images were randomly divided into training and test sets in a 9:1 ratio for model training.

[0050] The semantic segmentation algorithm uses PyTorch as the deep learning framework, and the semantic segmentation data is labeled in Pascalvoc format.

[0051] The test set of 1500 labeled original cat pattern images was imported into a clothing pattern segmentation model using MobileNetV2 as the backbone feature extraction network. The downsampling factor was set to 16, and the input image size was 512×512 pixels. The segmentation results are shown in the table below.

[0052] Table 1. Semantic segmentation results of cat-themed clothing patterns.

[0053]

[0054] The results above show that using MobileNetV2 as the backbone network for segmenting cat-like patterns yields excellent segmentation results, with a shallow network depth, few parameters, and fast response speed.

[0055] S402. Constructing a fast style transfer model:

[0056] The fast style transfer algorithm is used to train fast style transfer models of different styles to obtain trained fast style transfer models.

[0057] Fast style transfer models are one of the style transfer models based on convolutional neural networks. The model's generative network is usually a deep residual network. The first part is downsampling for image feature extraction, the middle part is a residual network module, then upsampling, and the end is a convolutional layer, which finally outputs the generated image.

[0058] The loss network of the fast style transfer model uses a VGG16 model that removes the fully connected layers and retains the convolutional layers. This reduces the number of parameters in the fully connected layers, avoids wasting computational resources, and speeds up the computation process.

[0059] The fast style transfer algorithm uses PyTorch as its deep learning framework.

[0060] The original pattern is processed by an image conversion network f w The image is transformed into a transfer pattern. Then, a loss network is used to calculate the content loss between the transfer pattern and the original pattern, as well as the style loss between the transfer pattern and the style pattern. The transfer pattern is further normalized to train the image transfer network f. w The purpose, such as Figure 3 As shown.

[0061] Content loss refers to using Euclidean distance to compare the differences between the transferred pattern and the original pattern; style loss refers to the loss of information such as pattern color and texture. By adjusting the parameters, the proportion and magnitude of the loss are controlled, so that the image transfer results can produce different effects.

[0062] In the fast style transfer model, the style transfer effect is adjusted by setting the content loss weight value α and the style loss weight value β. The total loss is the sum of the two losses and their respective weights. Different style patterns correspond to different parameter settings. By adjusting and controlling the magnitude and proportion of the two style losses, different transfer patterns can be obtained.

[0063] Different style patterns are collected and input into a fast style transfer algorithm for training, resulting in several fast style transfer models with different styles. In this invention, the fast style transfer models used include an ink painting style transfer model, an oil painting style transfer model, and a pixel art style transfer model.

[0064] This invention selects three image styles: ink painting style, oil painting style, and pixel art style. For each style, a fast style transfer model is set with three sets of parameters: α = 1e5, β = 1e10; α = 1e5, β = 2e10; and α = 1e5, β = 5e10. Other parameters are set with Epoch = 200 and learning rate = 0.001. The input image is 256×256 pixels. Content loss testing is conducted, and the transferred image and model are saved every 50 batches to observe the transfer effect. Style testing is performed using cat-like images, and the results are as follows. Figure 4As shown.

[0065] As can be seen from the image, the (d) parameter setting for the ink wash style produces the best pattern effect. The characteristic of the ink wash style is the use of different proportions of water and ink to create various light and shadow effects, and the rich variations in ink density, lightness, and clarity, resulting in a unique artistic effect in the composition. The ink wash pattern under parameter (d) possesses the brushstroke feel of a stippling technique in painting, with colors appearing to spread from the center. Furthermore, the outline of the cat pattern transfers the rhythmic beauty of ink wash lines, some lines being strong and straight, while others are soft and flowing. The patterns under the other two parameters, however, only transfer the colors of the ink wash style, resulting in a poorer textural effect.

[0066] Oil painting style emphasizes the expression and temperament of the figures, and the painting will show the special texture and feel of oil paint. Because oil paint has strong color coverage, light and light colors are painted last, making the light colors the most rich. Observing the migration results of oil painting style under three parameters, the oil painting style with parameter (c) has a clear texture and is also consistent with oil painting style in terms of light color expression.

[0067] Pixel art is similar to mosaic art; it is composed of basic dots and color blocks arranged in a regular pattern, characterized by a high degree of abstraction of the image content. Among the three transfer results, (a) is the most aesthetically pleasing. All three results exhibit the characteristic of regular color block composition, but (a) shows a distinct color block abstraction, differentiating between the face and body. The body is abstracted by the main blue-purple color blocks, while the facial features retain their colors while being abstracted by color blocks. The other two are relatively over-transferred, and in (d), the main subject of the pixel art pattern is no longer clear.

[0068] Objective Quantitative Approach: Structural similarity is used to measure the similarity between the transferred and original patterns. An image is defined as a combination of three different factors: brightness, contrast, and structure. The mean is used as an estimate of brightness, the standard deviation as an estimate of contrast, and the covariance as a measure of structural similarity. The numerical range is 0 to 1; a higher value indicates a higher similarity between the two images. The calculation formula is as follows:

[0069]

[0070] Where x and y represent two images, μ x μ represents the mean of x. y σ represents the mean of y; x σ represents the standard deviation of x. y σ represents the standard deviation of y. xy Let c1 and c2 be the covariance of x and y; c1 and c2 are constants, and it is the existence of constants that can avoid the case where the denominator is 0.

[0071] The measurement results are shown in Table 2:

[0072] Table 2 Objective Quantitative Evaluation Results

[0073]

[0074] Therefore, in the ink painting style fast style transfer model, the content loss weight value α = 1e5 and the style loss weight value β = 5e10; in the oil painting style fast style transfer model, the content loss weight value α = 1e5 and the style loss weight value β = 2e10; in the pixel style transfer model, the content loss weight value α = 1e5 and the style loss weight value β = 1e10.

[0075] Subjective evaluation: The quality and effect of the transferred pattern were assessed based on human visual perception. Ten textile and apparel industry professionals, 32 textile and apparel students, and ten individuals with prior work or study experience in the textile and apparel sector were selected for subjective evaluation.

[0076] The scores are based on two aspects: the overall aesthetic appeal of the design and the style transfer effect. The scores range from 1 to 5, corresponding to five levels: very bad, bad, average, good, and very good. The higher the score, the better the effect.

[0077] Evaluation results as follows Figure 5 As shown in the figure, the scoring results indicate that the overall subjective evaluation of the experiment is almost consistent with the objective evaluation results.

[0078] Based on the above personal analysis, objective quantitative evaluation, and subjective qualitative evaluation results, the following conclusions are drawn: In the rapid style transfer model for ink painting style, the content loss weight value α = 1e5 and the style loss weight value β = 5e10; in the rapid style transfer model for oil painting style, the content loss weight value α = 1e5 and the style loss weight value β = 2e10; and in the pixel art style transfer model, the content loss weight value α = 1e5 and the style loss weight value β = 1e10.

[0079] S5. Call the trained pattern segmentation model, input the original cat pattern data into the trained pattern segmentation model, obtain a mask image including the pattern part elements and background part elements of the target transfer, and enhance the brightness and contrast of the mask image.

[0080] S6: Call the trained fast style transfer model to combine the original cat pattern data with the mask. Figure 1 The input is fed into different pre-trained style transfer models. Then, the desired style is selected, and the pattern transfer part is customized using a mask image. Different mask images are selected to mask different element positions to obtain clothing patterns of different styles or mixed styles, so that multiple styles can appear on one pattern.

[0081] Example 2:

[0082] S1. Collect clothing sample data with car patterns through web crawling, text retrieval, and Kaggle to obtain initial car pattern data;

[0083] S201. A pattern detection algorithm based on YOLOv3 is used to train a pattern detection model using YOLOv3. The pattern detection model is used to organize, detect, and cut out clothing patterns, and to extract car pattern data from the initial car pattern data.

[0084] S202. The car pattern data is sharpened and standardized using the USM sharpening algorithm to improve the clarity of the pattern outline and obtain the processed high-definition original car pattern data.

[0085] S3. Label the high-definition original car pattern data using Labelme to mark the outline and background of the car pattern, separate the car pattern itself from the background, and determine the pattern area that needs to be segmented.

[0086] S401. The pattern segmentation model is trained using the DeepLabV3+ algorithm with MobileNetV2 as the backbone feature extraction network, PyTorch is used as the deep learning framework, and Pascal VOC format is used for annotation.

[0087] S402. Train fast style transfer models of different styles using the fast style transfer algorithm. The fast style transfer model uses PyTorch as the deep learning framework. The loss network is the VGG16 model with the fully connected layers removed and the convolutional layers retained.

[0088] During training, 1500 labeled original car pattern data were collected to train the pattern segmentation model, and the resulting mask image was enhanced in terms of brightness and contrast.

[0089] Fast style transfer models include ink painting style transfer model, oil painting style transfer model and pixel style transfer model.

[0090] The content loss weight value of the ink painting style migration model is α = 1e5, and the style loss weight value is β = 5e10.

[0091] The content loss weight value of the oil painting style transfer model is α = 1e5, and the style loss weight value is β = 2e10.

[0092] The content loss weight value α = 1e5 and the style loss weight value β = 1e10 for the pixel style transfer model.

[0093] S5. Call the trained pattern segmentation model, input the original car pattern data into the trained pattern segmentation model, obtain a mask image including the car elements and background elements of the target transfer, and enhance the brightness and contrast of the mask image.

[0094] Table 3. Semantic segmentation results of car clothing patterns

[0095]

[0096]

[0097] S6. Call the trained fast style transfer model to transfer the original car pattern data and mask. Figure 1 The input is fed into pre-trained ink painting style transfer model, oil painting style transfer model, and pixel art style transfer model. Depending on the requirements, different mask images are selected to mask different element positions. Custom patterns are used to transfer car elements or background elements, transforming one car pattern into nine different car patterns. This method is efficient, offers diverse elements, and allows for pattern selection space. The result is as follows: Figure 6 As shown. Figure 6 In the free combination pattern of Chinese elements, the first row from left to right consists of pixel-style fruit patterns with the original background, ink-wash style fruit patterns, and oil painting style fruit patterns; the second row from left to right consists of ink-wash style background + pixel-style fruit pattern combination, pixel-style background + ink-wash style fruit pattern combination, and pixel-style background + oil painting style fruit pattern combination; the third row from left to right consists of oil painting style background + pixel-style fruit pattern combination, oil painting style background + ink-wash style fruit pattern combination, and ink-wash style background + oil painting style fruit pattern combination.

[0098] Example 3:

[0099] S1. Collect sample data of clothing with fruit patterns through web crawling, text retrieval, and Kaggle to obtain initial fruit pattern data;

[0100] S201. A pattern detection algorithm based on YOLOv3 is used to train a pattern detection model using YOLOv3. The pattern detection model is used to organize, detect, and cut out clothing patterns, and extract fruit pattern data from the initial fruit pattern data.

[0101] S202. The fruit pattern data is sharpened and standardized using the USM sharpening algorithm to improve the clarity of the pattern outline, resulting in high-definition original fruit pattern data.

[0102] S3. Label the high-resolution original fruit image data using Labelme to mark the outline and background of the fruit image, separate the fruit image itself from the background, and determine the image area that needs to be separated.

[0103] S401. The pattern segmentation model is trained using the DeepLabV3+ algorithm with MobileNetV2 as the backbone feature extraction network, PyTorch is used as the deep learning framework, and Pascal VOC format is used for annotation.

[0104] S402. Train fast style transfer models of different styles using the fast style transfer algorithm. The fast style transfer model uses PyTorch as the deep learning framework. The loss network is the VGG16 model with the fully connected layers removed and the convolutional layers retained.

[0105] During training, 1500 labeled original fruit pattern data were collected to train the pattern segmentation model, and the resulting mask image was obtained.

[0106] Fast style transfer models include ink painting style transfer model, oil painting style transfer model and pixel style transfer model.

[0107] The content loss weight value of the ink painting style migration model is α = 1e5, and the style loss weight value is β = 5e10.

[0108] The content loss weight value of the oil painting style transfer model is α = 1e5, and the style loss weight value is β = 2e10.

[0109] The content loss weight value α = 1e5 and the style loss weight value β = 1e10 for the pixel style transfer model.

[0110] S5. Call the trained pattern segmentation model, input the original fruit pattern data into the trained pattern segmentation model, obtain a mask image including the fruit elements and background elements of the target transfer, and enhance the brightness and contrast of the mask image.

[0111] Table 4. Semantic segmentation results of fruit clothing patterns

[0112]

[0113] S6. Call the trained fast style transfer model to combine the original fruit pattern data and the mask. Figure 1 The input is fed into pre-trained ink painting style transfer model, oil painting style transfer model, and pixel art style transfer model. Depending on the requirements, different mask images are selected to mask different element positions. Custom patterns are used to transfer fruit elements or background elements, transforming one fruit pattern into nine fruit patterns. This method is efficient, offers diverse elements, and allows for pattern selection space. The result is as follows: Figure 7 As shown. Figure 7In the free combination pattern of Chinese elements, the first row from left to right consists of pixel-style car patterns with the original background, ink-wash style car patterns, and oil painting style car patterns; the second row from left to right consists of ink-wash style background + pixel-style car pattern combination, pixel-style background + ink-wash style car pattern combination, and pixel-style background + oil painting style car pattern combination; the third row from left to right consists of oil painting style background + pixel-style car pattern combination, oil painting style background + ink-wash style car pattern combination, and ink-wash style background + oil painting style car pattern combination.

Claims

1. A deep learning-based hybrid style garment pattern intelligent generation method, characterized in that: The method comprises the following steps: S1, collecting patterned garment sample data to obtain initial pattern data; S2, detecting a target pattern in the initial pattern data and preprocessing to obtain processed original pattern data; S3, labeling the original pattern data for patterns and backgrounds to determine the pattern area to be segmented; S4, training a pattern segmentation model using a semantic segmentation algorithm and training a fast style transfer model of different styles using a fast style transfer algorithm; S5, inputting the original pattern data into the trained pattern segmentation model to obtain a mask image and enhancing the brightness and contrast of the mask image; S6, inputting the original pattern data and the mask image into the trained fast style transfer model of different styles, selecting different mask images to mask different element positions, and obtaining patterns of different styles or mixed styles.

2. The hybrid style garment pattern intelligent generation method based on deep learning according to claim 1, characterized in that: In step S2, the target pattern detection and cutting is a pattern detection algorithm based on YOLOv3, a pattern detection model is trained using YOLOv3, and the garment pattern is cut by the pattern detection model; the preprocessing includes sharpening processing to improve the contour clarity of the pattern.

3. The hybrid style garment pattern intelligent generation method based on deep learning according to claim 1, characterized in that: In step S5, the mask image includes target transferred pattern part elements and background part elements.

4. The hybrid style garment pattern intelligent generation method based on deep learning according to claim 1, characterized in that: In step S4, the semantic segmentation algorithm uses a DeepLabV3+ algorithm, and the backbone feature extraction network of the DeepLabV3+ algorithm uses MobileNetV2.

5. The hybrid style garment pattern intelligent generation method based on deep learning according to claim 1, characterized in that: In step S4, the loss network of the fast style transfer model uses a VGG16 model with the fully connected layer removed and the convolutional layer retained.

6. The hybrid style garment pattern intelligent generation method based on deep learning according to claim 1, characterized in that: In step S4, the fast style transfer model includes an ink style transfer model, an oil painting style transfer model, and a pixel style transfer model.

7. The hybrid style garment pattern intelligent generation method based on deep learning according to claim 6, characterized in that: The ink style transfer model includes content loss and style loss, the content loss weight value α = 1e5, and the style loss weight value β = 5e10.

8. The hybrid style garment pattern intelligent generation method based on deep learning according to claim 6, characterized in that: The oil painting style transfer model includes content loss and style loss, the content loss weight value α = 1e5, and the style loss weight value β = 2e10.

9. The hybrid style garment pattern intelligent generation method based on deep learning according to claim 6, characterized in that: The pixel style transfer model includes content loss and style loss, the content loss weight value α = 1e5, and the style loss weight value β = 1e10.

Images