Style migration method for content image object separation

Abstract

The invention discloses a style migration method for separating content image objects, which comprises the following steps of: extracting a style image and a content image of a target image, processing the style image by using a salient target detection module, and distinguishing a style main object from a background object; constructing a depth estimation mechanism, processing the content image through a depth estimation module, and obtaining depth information of the content image; on the basis of the style separation attribute of the convolutional feature channel, extracting the features of the style subject object, and designing an opacity sensitive loss function to obtain a stylized subject object migration result; in combination with the characteristic that the content background contains a complex object, based on defocusing and instance normalization, the definition and style information of the content background are adjusted to obtain a stylized background migration result; and finally, combining object migration results. When facing a content image with a disordered content main body, the method guides the style migration module to realize style migration through a depth estimation mechanism, and effectively solves the problem of disordered style migration results of a complex scene.

Description

Technical Field

[0001] The present invention belongs to the field of digital image processing and machine vision, and in particular relates to a style transfer method for separating content image objects. Background Art

[0002] Excellent works of art often provide information different from real photos. However, ordinary people may not be able to independently complete a work of art that satisfies themselves and others without long-term professional training. At the same time, the time and cost required to train a real artist are inestimable. Not only that, even for artists who are proficient in style painting, it takes a long time to complete a work of art. In order to efficiently convert real photos into artistic images, the image style transfer task came into being. Image style transfer aims to combine the content of a real photo with the style of an artwork to form a stylized photo that has both the content of the photo image and the style of the artwork. In the style transfer task, the image that provides content is called the content image, the image that provides style is called the style image, and the generated result is called the stylized image. Style transfer specifically refers to image style transfer. The development of style transfer can be divided into two stages: the first stage from the mid-1990s to 2016, its main feature is the use of mathematical models to simulate texture simulation. The second stage from 2016 to the present, its main feature is the use of deep learning and neural networks for style transfer. Compared with the two, the former is more traditional, while the latter incorporates new methods, so the first stage is called "traditional style transfer" and the second stage is called "neural style transfer". Traditional style transfer usually uses mathematical and signal processing techniques, such as texture synthesis, histogram matching, and filtering. These methods involve operations on pixels to achieve the purpose of simulating the desired style. For example, frequency domain filtering can be used to enhance or weaken certain frequency components of an image, thereby changing its appearance. The advantages of traditional style transfer are higher computing speed and lower resource usage, but they may not be able to capture more advanced artistic styles and textures.

[0003] Style transfer methods based on neural networks are more flexible and efficient. These methods use deep learning technology to learn and apply image styles. They train neural networks to capture the features of different artistic styles, and then apply these features to the input image to generate new images with the desired style. The advantage of neural style transfer is that it can better capture the details and complexity of artistic styles. The time and resource usage required will vary greatly depending on the network structure. There should not be a relationship between traditional style transfer and neural style transfer. On the contrary, the ideas of some current neural style transfer results come from traditional style transfer and digital image processing. Image style transfer has many application scenarios in real life, such as in environmental atmosphere rendering, font generation, font recognition, portrait editing, auxiliary design, photo restoration, virtual reality (VR) and augmented reality (AR). At the same time, style transfer, as the underlying task of computer vision, can assist in the development of other tasks such as adversarial sample research, image generation research, and domain adaptation. Whether from the perspective of practical application or scientific research, style transfer tasks have a wide range of applications. The current neural style transfer method still has some defects. An excellent work of art should have a sense of hierarchy in the description of the scene and retain a certain amount of white space in the space. However, current neural style transfer methods often ignore the characteristics of clear primary and secondary relationships and the combination of whole and scattered in works of art, resulting in unclear primary and secondary relationships in the stylized results when performing style transfer on images with cluttered content. To address the above issues, we also make up for the defect that style transfer methods cannot maintain consistency with the background characteristics of the original style image when performing style transfer tasks on certain style images. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a style transfer method for separating content image objects. When faced with a content image with a cluttered main body, the depth estimation mechanism is used to guide the style transfer module to achieve style transfer in a specific area, effectively solving the problem of cluttered style transfer results in complex scenes, and ensuring that the background part of the style transfer result is highly consistent with the background style of the original style image.

[0005] Technical solution: A style transfer method for separating content, image and object of the present invention comprises the following steps: Step 1: Extract the style image and content image of the target image, input the style image into the salient object detection module, and distinguish the main object from the background object in the style image; Step 2: Based on the characteristics that there are many and complex scenes and subjects in the content image, and the same subject is at the same depth, a depth estimation mechanism is constructed, and the content image is input into the depth estimation module. Then, according to the ambiguity of the photographic image taken from different camera models, a standard camera space transformation is constructed, and the depth information in the content image is estimated by using the standard camera transformation and destandardization operation, so as to separate the content image into an image of the region of interest and an image of the region of non-interest; Step 3: For the main object of the style image, based on its obvious texture features, the style features of the main object of the style image are extracted based on the style separation properties of the Fourier transform and deep convolution feature channels, and an opacity-sensitive loss function is designed to train the migration results of the main object to obtain the stylized main object migration results; Step 4: For the background objects in the style image, based on their single color characteristics and the fact that the non-interest objects contain more complex objects, the clarity and style information of the non-interest objects are adjusted based on background defocus and adaptive instance normalization to obtain the stylized background object migration result; Step 5: Merge the stylized subject object migration results and the stylized background object migration results to achieve object separation style migration of the target image.

[0006] Furthermore, step 1 specifically includes the following steps: Step 101: Use saliency detection network Style Image Processing to obtain the saliency map , used to represent the probability that each pixel value belongs to the subject, the formula is as follows: Step 102: The saliency map obtained in step 101 , separate the style image S into the main part With background part , the formula is as follows: in, Represents element-wise multiplication.

[0007] Furthermore, step 2 specifically includes the following steps: Step 201: input content image After wavelet decomposition, the low-frequency components are obtained. With high frequency components , the formula is as follows: in Represents wavelet transform, for high frequency components Apply the non-local mean method to reduce noise and obtain the high-frequency component after noise reduction , the formula is as follows: Among them, the weight Calculated based on the similarity of pixel blocks; The low frequency component The high frequency component after noise reduction Perform inverse wavelet transform to obtain the enhanced image : Step 202: The enhanced image obtained in step 101 is Convert to the standard camera space to eliminate the metric ambiguity caused by different camera models. The standard camera transformation is expressed as: in, represents the standard camera space transformation operation, Represents a standard camera model image; Step 203: In the standard camera space, the standard camera model image obtained in step 202 is Perform depth estimation and obtain a standardized depth map , the formula is as follows: in, represents the depth estimation model, Content image Depth information; Step 204: perform a normalization operation to reduce the error. The formula is as follows: in, represents the normalization operation, and Respectively represent depth information The minimum and maximum values ​​in ; Step 205: Based on the normalized depth information in step 204 , combined with user input, the original content image Divide into regions of interest and regions of no interest, create masks With the corresponding Channel Mask ; Step 206: Apply the four masks obtained in step 205 to the enhanced content image to obtain a content interest image. Images with non-interesting content ; Further, in step 205, the channel mask The formula is as follows: by Indicates the number of regions where the content image is divided, Represent the region of interest mask: Represents the corresponding interest region mask Channel Mask: Represents a mask of a non-interest region: Represents the mask corresponding to the non-interest region Channel Mask: Furthermore, in step 206, the content interest image Images with non-interesting content , the formula is as follows: in, represents element-wise multiplication, Represents a concatenation operation.

[0008] Furthermore, step 3 specifically includes the following steps: Step 301: the style subject object obtained in step 1 The content interest image obtained in step 2 Perform octave convolution to decompose it into high-frequency and low-frequency components. The formula is as follows: in, and Represent the high-frequency and low-frequency components of the content interest image, respectively. and They represent the high-frequency and low-frequency components of the style subject object respectively; Step 302: Use the style feature extractor E_{s} to extract the style features of the style subject object obtained in step 301, and predict the corresponding convolution kernel and bias through the kernel prediction network. The formula is as follows: in, and is the convolution and bias prediction network, represents the convolution kernel and bias predicted from the high-frequency style subject object, Represents the convolution kernel and bias predicted from the low-frequency style subject object; Step 303: The high-frequency and low-frequency components of the content interest image are combined with the convolution kernel and bias predicted in step 302 and are added to the generator. Perform style transfer to obtain the stylized subject object , the formula is as follows: Step 304: Calculate the loss function based on the stylized subject object obtained in step 303 And back propagate to update the convolution and bias prediction network and style transfer generator , the formula is as follows: in, Respectively represent the content loss function, style loss function, and opacity loss function, is a hyperparameter used to represent the content of the three loss functions in the total loss function.

[0009] Furthermore, in step 304, the content loss function It is used to ensure that the stylized image retains similar structure and details to the original content image. The formula is as follows: in, represents the pre-trained VGG feature extraction network; The style loss function It is used to ensure that the stylized image can capture and reproduce the texture and style characteristics of the style image. The formula is as follows: in, represents the Gram matrix, Indicates Layer feature map; The opacity loss function It is used to ensure the sensitivity of the style transfer network to the opacity information in the image, and to use positive and negative samples to improve the aesthetic effect of style transfer. The formula is as follows: in, Indicates the use of generator No. The high-frequency part of the layer output feature map, Indicates the use of generator No. The low-frequency part of the layer output feature map, The generator generates A stylized subject object. express of aisle, Indicates Positive samples, Indicates Negative samples.

[0010] Furthermore, step 4 specifically includes the following steps: Step 401: Use Gaussian blur to apply defocus processing to the content non-interest object to make it more closely match the single color characteristics of the style background object. The formula is as follows: in, The blurred content is not the object of interest. The original content is not the object of interest. is the standard deviation of the Gaussian blur, which controls the degree of blur; Step 402: The pre-trained neural network VGG 19 extracts the style background object features of the blurred content non-interest object in step 401. These features are used to guide the style transfer of the content non-interest object. The formula is as follows: in, is a background object from the style The features extracted from Represents the pre-trained feature extraction network VGG19; Step 403: Combining the features of the content non-interest object and the style background object, use adaptive instance normalization AdaIN to adjust the style of the content non-interest object obtained in step 402; in, and Respectively represent the mean and standard deviation calculation, Represents the characteristics of the style background object, Represents a stylized background object.

[0011] Furthermore, step 5 is specifically as follows: performing a final merging operation on the stylized subject object and the stylized background object, and the formula is as follows: in, A mask representing the subject object, represents the identity matrix, Represents the final style transfer result.

[0012] The present invention also discloses a computer device, comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0013] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. When faced with content images with cluttered main bodies, the present invention guides the style transfer module to implement style transfer for specific areas through a depth estimation mechanism, thereby effectively solving the problem of cluttered style transfer results in complex scenes.

[0014] 2. In the training of the style subject object transfer network, the style training loss function that takes the image alpha channel information into consideration is redesigned. This can not only guide the model to converge quickly, but also allow the model to consider the opacity information in the image, effectively solving the problem of difficulty in transferring PNG format images in the field of style transfer. In the training of the network, the difference between positive and negative samples is also used to train the network, achieving more effective style transfer.

[0015] 3. In terms of content image decomposition, the characteristic that the same subject is basically at the same depth is used to decompose the content image correlation matrix, and then the subject object area and the background object area are extracted. This can better distinguish the boundary between the subject and the background. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is the overall framework diagram of the model proposed by the present invention; Figure 2 is a multi-scale information extraction module of the present invention; Figure 3 The cross-scale information interaction method proposed by the present invention; Figure 4 Schematic diagram of the deep convolution gated forward propagation layer of the present invention; Figure 5 This is the final style transfer flow chart of the present invention; Figure 6 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0017] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0018] The present embodiment of the present invention discloses a style transfer method for separating content image objects. The overall framework of the model is as follows: Figure 1 As shown, the following steps are included: Step 1: Based on the fact that the background and main content in traditional Chinese art works are quite different, the style image is input into the salient object detection module to distinguish the main part from the background part of the style image, thereby achieving the purpose of separating different style parts in the style image.

[0019] Further, the step 1 is specifically as follows: Step 101: Use saliency detection network Style Image Processing to obtain the saliency map , used to represent the probability that each pixel value belongs to the subject, the formula is as follows: Step 102: Based on the saliency map obtained in step 101 , separate the style image S into the main part With background part , the formula is as follows: in, Represents element-wise multiplication.

[0020] Step 2: Based on the characteristics of the content image that there are many and complex scenes and subjects, and the same subject is basically at the same depth, a depth estimation mechanism is constructed. The content image is input into the depth estimation module, and then a standard camera space transformation is constructed according to the ambiguity of the photographic image taken by different camera models. The standard camera transformation and de-standardization operation are used to estimate the depth information in the content image to achieve the purpose of separating the content image into the image of the region of interest and the image of the region of non-interest. The process is as follows: Figure 1 shown.

[0021] Further, the step 2 is specifically as follows: Step 201: input content image After wavelet decomposition, the low-frequency components are obtained. With high frequency components , the formula is as follows: in Represents wavelet transform. For high frequency components Apply the non-local mean method to reduce noise and obtain the high-frequency component after noise reduction , the formula is as follows: Among them, the weight Calculated based on the similarity of pixel blocks.

[0022] The low frequency component The high frequency component after noise reduction Perform inverse wavelet transform to obtain the enhanced image Step 202: The enhanced image obtained in step 101 is Convert to the standard camera space to eliminate the metric ambiguity caused by different camera models. The standard camera transformation can be expressed as: in, represents the standard camera space transformation operation, Represents a standard camera model image.

[0023] Step 203: In the standard camera space, the standard camera model image obtained in step 202 is Perform depth estimation and obtain a standardized depth map , the formula is as follows: in, represents the depth estimation model, Content image depth information.

[0024] Step 204, further perform normalization operation to reduce the error. The formula is as follows: in, represents the normalization operation, and Respectively represent depth information The minimum and maximum values ​​in .

[0025] Step 205: Based on the normalized depth information in step 204 , combined with user input, the original content image Divide into regions of interest and regions of no interest, create masks With the corresponding Channel Mask , the formula is as follows: by Indicates the number of regions where the content image is divided, Represent the region of interest mask: Represents the corresponding interest region mask Channel Mask: Represents a mask of a non-interest region: Represents the mask corresponding to the non-interest region Channel Mask: Step 206: Apply the four masks obtained in step 205 to the enhanced content image to obtain a content interest image. Images with non-interesting content , the formula is as follows.

[0026] in, represents element-wise multiplication, Represents the splicing operation, the process is as follows Figure 2 shown.

[0027] Step 3: For the style subject object, based on its obvious texture characteristics, the style characteristics of the style subject object are extracted based on the style separation properties of the Fourier transform and deep convolution feature channels, and an opacity-sensitive loss function is designed to train the migration results of the subject object to obtain the stylized subject object migration results.

[0028] Further, the step 3 is specifically as follows: Step 301: the style subject object obtained in step 1 and the content interest image obtained in step 2 Perform octave convolution to decompose it into high-frequency and low-frequency components. The formula is as follows: in, and Represent the high-frequency and low-frequency components of the content interest image, respectively. and Respectively represent the high-frequency and low-frequency components of the style subject object.

[0029] Step 302: Use style feature extractor The style features of the style subject object obtained in step 301 are extracted, and the corresponding convolution kernel and bias are predicted by the kernel prediction network. The formula is as follows: in, and is the convolution and bias prediction network, represents the convolution kernel and bias predicted from the high-frequency style subject object, Represents the convolution kernel and bias predicted from the low-frequency style subject.

[0030] Step 303, the high-frequency and low-frequency components of the content interest image are combined with the convolution kernel and bias predicted in step 302 and are added to the generator Perform style transfer to obtain the stylized subject object , the formula is as follows: Step 304: Calculate the loss function based on the stylized subject object obtained in step 303 And back propagate to update the convolution and bias prediction network and style transfer generator , the formula is as follows: in, They represent content loss function, style loss function, and opacity loss function respectively. is a hyperparameter used to represent the content of the three loss functions in the total loss function.

[0031] Content loss It is used to ensure that the stylized image maintains similar structure and details to the original content image. The formula is as follows: in, Represents the pre-trained VGG feature extraction network.

[0032] Style Loss It is used to ensure that the stylized image can capture and reproduce the texture and style characteristics of the style image. The formula is as follows: in, represents the Gram matrix, Indicates Layer feature map.

[0033] Opacity loss It is used to ensure the sensitivity of the style transfer network to the opacity information in the image, and to use positive and negative samples to improve the aesthetic effect of style transfer. The formula is as follows: in, Indicates the use of generator No. The high-frequency part of the layer output feature map, Indicates the use of generator No. The low-frequency part of the layer output feature map, The generator generates A stylized subject object. express of aisle, Indicates Positive samples, Indicates negative samples, the process is as follows Figure 3 shown.

[0034] In step 4, based on the single color characteristics of the style background objects and the fact that the content non-interest objects contain more complex objects, the clarity and style information of the content non-interest objects are adjusted based on background defocus and adaptive instance normalization to obtain the stylized background object migration result.

[0035] Further, the step 4 is specifically as follows: Step 401, use Gaussian blur to apply defocus processing to the content non-interest object to make it more closely match the single color characteristics of the style background object. The formula is as follows: in, The blurred content is not the object of interest. The original content is not the object of interest. is the standard deviation of the Gaussian blur, controlling the degree of blur.

[0036] In step 402, the pre-trained neural network VGG 19 extracts style background object features from the blurred content non-interest object in step 401. These features are used to guide the style transfer of the content non-interest object. The formula is as follows: in, is a background object from the style The features extracted from Represents the pre-trained feature extraction network VGG19.

[0037] Step 403 , combining the features of the content non-interest object and the style background object, and using Adaptive Instance Normalization (AdaIN) to adjust the style of the content non-interest object obtained in step 402 .

[0038] in, and Respectively represent the mean and standard deviation calculation, Represents the characteristics of the style background object, Represents a stylized background object.

[0039] In step 4, based on the single color characteristics of the style background objects and the fact that the content non-interest objects contain more complex objects, the clarity and style information of the content non-interest objects are adjusted based on background defocus and adaptive instance normalization to obtain the stylized background object migration result.

[0040] Step 5: After obtaining the stylized subject object and the stylized background object, the original content image is used as a reference to normalize the stylized subject object migration result and the stylized background object migration result to reduce the style difference between the two, and finally merge them. The process is as follows: Figure 4 shown.

[0041] Further, the step 5 is specifically as follows: Step 501, perform a final merging operation on the stylized main object and the stylized background object, the formula is as follows: in, A mask representing the subject object, represents the identity matrix, Represents the final style transfer result. The process is as follows Figure 5 shown.

Claims

1. A style transfer method for separating content image objects, characterized in that: The steps include: Step 1: Extract the style image and content image of the target image, input the style image into the salient object detection module, and distinguish the main object from the background object in the style image; Step 2: Based on the characteristics that there are many and complex scenes and subjects in the content image, and the same subject is at the same depth, a depth estimation mechanism is constructed, and the content image is input into the depth estimation module. Then, according to the ambiguity of the photographic image taken from different camera models, a standard camera space transformation is constructed, and the depth information in the content image is estimated by using the standard camera transformation and destandardization operation, so as to separate the content image into an image of the region of interest and an image of the region of non-interest; Step 3: For the main object of the style image, based on its obvious texture features, the style features of the main object of the style image are extracted based on the style separation properties of the Fourier transform and deep convolution feature channels, and an opacity-sensitive loss function is designed to train the migration results of the main object to obtain the stylized main object migration results; Step 4: For the background objects in the style image, based on their single color characteristics and the fact that the non-interest objects in the content contain complex objects, the clarity and style information of the non-interest objects in the content are adjusted based on background defocus and adaptive instance normalization to obtain the stylized background object migration result; Step 5: Merge the stylized subject object migration results and the stylized background object migration results to achieve object separation style migration of the target image.

2. The style transfer method for separating content image objects according to claim 1, characterized in that: Step 1 specifically includes the following steps: Step 101: Use saliency detection network Style Image Processing to obtain the saliency map , used to represent the probability that each pixel value belongs to the subject, the formula is as follows: Step 102: The saliency map obtained in step 101 , separate the style image S into the main part With background part , the formula is as follows: in, Represents element-wise multiplication.

3. The style transfer method for separating content image objects according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 201: input content image After wavelet decomposition, the low-frequency components are obtained. With high frequency components , the formula is as follows: in Represents wavelet transform, for high frequency components Apply the non-local mean method to reduce noise and obtain the high-frequency component after noise reduction , the formula is as follows: Among them, the weight Calculated based on the similarity of pixel blocks; The low frequency component The high frequency component after noise reduction Perform inverse wavelet transform to obtain the enhanced image : Step 202: The enhanced image obtained in step 101 is Convert to the standard camera space to eliminate the metric ambiguity caused by different camera models. The standard camera transformation is expressed as: in, represents the standard camera space transformation operation, Represents a standard camera model image; Step 203: In the standard camera space, the standard camera model image obtained in step 202 is Perform depth estimation and obtain a standardized depth map , the formula is as follows: in, represents the depth estimation model, Content image Depth information; Step 204: perform a normalization operation to reduce the error. The formula is as follows: in, represents the normalization operation, and Represents depth information The minimum and maximum values ​​in ; Step 205: Based on the normalized depth information in step 204 , combined with user input, the original content image Divide into regions of interest and regions of no interest, create masks With the corresponding Channel Mask ; Step 206: Apply the four masks obtained in step 205 to the enhanced content image to obtain a content interest image. Images with non-interesting content .

4. The style transfer method for separating content image objects according to claim 3, characterized in that: In step 205, the channel mask The formula is as follows: by Indicates the number of regions where the content image is divided, Represent the region of interest mask: Represents the corresponding interest region mask Channel Mask: Represents a mask of a non-interest region: Represents the mask corresponding to the non-interest region Channel Mask: 。 5. The style transfer method for separating content image objects according to claim 3, characterized in that: In step 206, the content interest image Images with non-interesting content , the formula is as follows: in, represents element-wise multiplication, Represents a concatenation operation.

6. The style transfer method for separating content image objects according to claim 1, characterized in that: Step 3 specifically includes the following steps: Step 301: the style subject object obtained in step 1 and the content interest image obtained in step 2 Perform octave convolution to decompose it into high-frequency and low-frequency components. The formula is as follows: in, and Represent the high-frequency and low-frequency components of the content interest image, respectively. and They represent the high-frequency and low-frequency components of the style subject object respectively; Step 302: Use style feature extractor The style features of the style subject object obtained in step 301 are extracted, and the corresponding convolution kernel and bias are predicted by the kernel prediction network. The formula is as follows: in, and is the convolution and bias prediction network, represents the convolution kernel and bias predicted from the high-frequency style subject object, Represents the convolution kernel and bias predicted from the low-frequency style subject object; Step 303: The high-frequency and low-frequency components of the content interest image are combined with the convolution kernel and bias predicted in step 302 and are added to the generator. Perform style transfer to obtain the stylized subject object , the formula is as follows: Step 304: Calculate the loss function based on the stylized subject object obtained in step 303 And back propagate to update the convolution and bias prediction network and style transfer generator , the formula is as follows: in, Respectively represent the content loss function, style loss function, and opacity loss function, is a hyperparameter used to represent the content of the three loss functions in the total loss function.

7. The style transfer method for separating content image objects according to claim 6, characterized in that: In step 304, the content loss function It is used to ensure that the stylized image retains similar structure and details to the original content image. The formula is as follows: in, represents the pre-trained VGG feature extraction network; The style loss function It is used to ensure that the stylized image can capture and reproduce the texture and style characteristics of the style image. The formula is as follows: in, represents the Gram matrix, Indicates Layer feature map; The opacity loss function It is used to ensure the sensitivity of the style transfer network to the opacity information in the image, and to use positive and negative samples to improve the aesthetic effect of style transfer. The formula is as follows: in, Indicates the use of generator No. The high-frequency part of the layer output feature map, Indicates the use of generator No. The low-frequency part of the layer output feature map, The generator generates A stylized subject object. express of aisle, Indicates Positive samples, Indicates Negative samples.

8. The style transfer method for separating content image objects according to claim 1, characterized in that: Step 4 specifically includes the following steps: Step 401: Use Gaussian blur to apply defocus processing to the content non-interest object to make it more closely match the single color characteristics of the style background object. The formula is as follows: in, The blurred content is not the object of interest. The original content is not the object of interest. is the standard deviation of the Gaussian blur, which controls the degree of blur; Step 402: The pre-trained neural network VGG 19 extracts the style background object features of the blurred content non-interest object in step 401. These features are used to guide the style transfer of the content non-interest object. The formula is as follows: in, is a background object from the style The features extracted from Represents the pre-trained feature extraction network VGG19; Step 403: Combining the features of the content non-interest object and the style background object, use adaptive instance normalization AdaIN to adjust the style of the content non-interest object obtained in step 402; in, and Respectively represent the mean and standard deviation calculation, Represents the characteristics of the style background object, Represents a stylized background object.

9. The style transfer method for separating content image objects according to claim 8, characterized in that: Step 5 is as follows: perform a final merging operation on the stylized subject object and the stylized background object. The formula is as follows: in, A mask representing the subject object, represents the identity matrix, Represents the final style transfer result.

10. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method of claim 1.

Images