CN116051687A - Synthetic image processing method, apparatus, computer device, and storage medium - Google Patents

Abstract

The present application relates to a composite image processing method, device, computer equipment, storage medium and computer program product. The method includes: acquiring a composite image, a mask of the foreground image, and a mask of the background image, and the first generator of the image processing model extracts light and shadow features of the foreground according to the mask of the composite image and the foreground image, and generates a first composite image; the second generator of the image processing model extracts the background light and shadow features according to the mask of the synthetic image and the background image, and generates a second composite image; based on the similarity between the foreground light and shadow features and the background light and shadow features, the first composite image and the second The composite images have similar light and shadow characteristics, and the first composite image and the second composite image are fused to obtain a harmonious target composite image, and the light and shadow are generated according to the light and shadow characteristics of the foreground image and the background image, and a harmonious target composite with light and shadow is obtained image.

Description

Composite image processing method, device, computer equipment and storage medium

Technical Field

The present application relates to the field of computer vision, and in particular to a synthetic image processing method, apparatus, computer equipment, storage medium and computer program product.

Background Art

Image synthesis, as a common image editing operation, can enable dataset augmentation for various applications and downstream vision tasks in entertainment, art, and business. For example, one can replace the background of a selfie and use image synthesis techniques to make the obtained image more realistic. Image synthesis can also be used in automatic advertising, which helps advertisers insert products in background scenes. In the process of image synthesis, a foreground target image is first cut from one image using an image segmentation or cutout technique, and then the obtained foreground image is pasted onto another image to obtain a composite image. But in such a composite image, its foreground and background may be taken under different conditions (e.g., weather, season, time of day, camera settings), so the foreground and background have different lighting characteristics, which makes the composite image look disharmonious.

To solve the above-mentioned problem of inharmonious composite images, it is necessary to adjust the appearance of the composite foreground according to the composite background to make it compatible with the composite background. Traditional image solution methods generally perform color transformation on the foreground to match the low-level color statistics between the foreground and the background. Due to the lack of high-level information, these methods limit the coordination performance, and the statistical data are mostly calculated manually, which is time-consuming. Recently, deep learning theory has achieved great success in computer vision tasks. With a huge network structure and a large number of training parameters, deep learning algorithms optimize the training network based on a large amount of labeled data, showing excellent performance in many fields, and are also used to solve the problem of inharmonious images.

The current deep model mainly adopts the encoder-decoder structure based on CNN, which uses an encoder to capture the contextual information of the synthesized image and a decoder to reconstruct the harmonized synthesized image. However, due to the inherent local inductive bias of CNN, the harmonization effect of the harmonized synthesized image is not ideal.

Summary of the invention

Based on this, it is necessary to provide a synthetic image processing method, device, computer equipment, computer-readable storage medium and computer program product that can harmonize synthetic images in order to solve the above technical problems.

In a first aspect, the present application provides a synthetic image processing method. The method comprises:

Get the composite image, the mask of the foreground image, and the mask of the background image;

Inputting the composite image and the mask of the foreground image into a first generator of an image processing model, extracting foreground light and shadow features from the mask of the foreground image, and obtaining a first composite image according to the foreground light and shadow features and the composite image;

Inputting the composite image and the mask of the background image into the second generator of the image processing model, extracting background light and shadow features from the mask of the background image, and obtaining a second composite image according to the background light and shadow features and the composite image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator;

The first composite image and the second composite image are fused to obtain a target composite image.

In one embodiment, the method of training the image processing model includes:

Acquire an image data set, the image data set comprising multiple groups of data, each group of data consisting of two triplet data, one of the triplet is an original synthetic image, a mask of a foreground image and a mask of a background image, and the other triplet is a real image, a real foreground illumination image and a real background illumination image;

Inputting the original synthesized image and the mask of the foreground image in the image data set into a first generator of the image processing model, extracting foreground light and shadow features from the mask of the foreground image, and obtaining a third synthesized image according to the foreground light and shadow features and the original synthesized image;

Inputting the original composite image and the mask of the background image into a second generator of an image processing model, extracting background light and shadow features from the mask of the background image, and obtaining a fourth composite image according to the background light and shadow features and the composite image;

fusing the third synthesized image and the fourth synthesized image to obtain a harmoniously processed image;

Calculating a model loss according to the harmonized image, the real image, the real foreground illumination image, and the real background illumination image;

The parameters of the first generator and the second generator are adjusted based on the constraint of the model loss to obtain a trained image processing model; the constraint of the loss function includes: the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator.

In one of the embodiments, the model loss includes: adversarial loss and illumination loss;

Calculating a model loss according to the harmonized image, the real image, the real foreground illumination image, and the real background illumination image includes:

Inputting the harmonized processed image and the real image into a discriminator of the image processing model, wherein the discriminator is used to judge the authenticity of the harmonized processed image and the real image;

Calculating adversarial loss based on the output of the discriminator;

The illumination feature loss is calculated according to the difference between the real foreground illumination image and the foreground illumination image of the harmonized image, and the difference between the real background illumination image and the background illumination image of the harmonized image.

In one of the embodiments, the model loss further includes: non-illumination feature loss;

The method further includes: calculating a non-illumination feature loss based on the non-illumination features of the same foreground image under different illumination conditions; wherein the goal of the non-illumination feature loss function is to minimize the difference in non-illumination features under different illumination conditions.

In one of the embodiments, the model loss further includes: perceptual loss;

The method further comprises:

Extracting a feature map of the harmonized image using a pre-trained neural network model;

Extracting a feature map of a real image corresponding to the harmonized image using a pre-trained neural network model;

The perceptual loss is determined based on the feature map of the harmonized image and the feature map of the real image corresponding to the harmonized image; the goal of the perceptual loss function is to minimize the difference between the feature map of the harmonized image and the feature map of the real image corresponding to the harmonized image.

In one embodiment, the model loss is a weighted sum of illumination feature loss, non-illumination feature loss, perceptual loss, adversarial loss, and absolute value deviation loss.

In a second aspect, the present application also provides a synthetic image processing device. The device comprises:

The composite image acquisition module is used to acquire a composite image, a mask of a foreground image, and a mask of a background image.

The first synthetic image acquisition module is used to input the synthetic image and the mask of the foreground image into the first generator of the image processing model, extract the foreground light and shadow features from the mask of the foreground image, and obtain the first synthetic image according to the foreground light and shadow features and the synthetic image.

The second synthetic image acquisition module is used to input the masks of the synthetic image and the background image into the second generator of the image processing model, extract background light and shadow features from the mask of the background image, and obtain a second synthetic image based on the background light and shadow features and the synthetic image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator.

The target composite image acquisition module is used to fuse the first composite image and the second composite image to obtain a target composite image.

In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

Get the composite image, the mask of the foreground image, and the mask of the background image;

Inputting the composite image and the mask of the foreground image into a first generator of an image processing model, extracting foreground light and shadow features from the mask of the foreground image, and obtaining a first composite image according to the foreground light and shadow features and the composite image;

Inputting the composite image and the mask of the background image into the second generator of the image processing model, extracting background light and shadow features from the mask of the background image, and obtaining a second composite image according to the background light and shadow features and the composite image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator;

The first composite image and the second composite image are fused to obtain a target composite image.

In a fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

Get the composite image, the mask of the foreground image, and the mask of the background image;

Inputting the composite image and the mask of the foreground image into a first generator of an image processing model, extracting foreground light and shadow features from the mask of the foreground image, and obtaining a first composite image according to the foreground light and shadow features and the composite image;

Inputting the composite image and the mask of the background image into the second generator of the image processing model, extracting background light and shadow features from the mask of the background image, and obtaining a second composite image according to the background light and shadow features and the composite image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator;

The first composite image and the second composite image are fused to obtain a target composite image.

In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

Get the composite image, the mask of the foreground image, and the mask of the background image;

Inputting the composite image and the mask of the foreground image into a first generator of an image processing model, extracting foreground light and shadow features from the mask of the foreground image, and obtaining a first composite image according to the foreground light and shadow features and the composite image;

Inputting the composite image and the mask of the background image into the second generator of the image processing model, extracting background light and shadow features from the mask of the background image, and obtaining a second composite image according to the background light and shadow features and the composite image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator;

The first composite image and the second composite image are fused to obtain a target composite image.

The above-mentioned synthetic image processing method, device, computer equipment, storage medium and computer program product obtain a synthetic image, a mask of a foreground image and a mask of a background image. The first generator of the image processing model extracts foreground light and shadow features according to the masks of the synthetic image and the foreground image, and generates a first synthetic image; the second generator of the image processing model extracts background light and shadow features according to the masks of the synthetic image and the background image, and generates a second synthetic image; based on the similarity of the foreground light and shadow features and the background light and shadow features, the first synthetic image and the second synthetic image have similar light and shadow features, the target synthetic image obtained by fusing the first synthetic image and the second synthetic image is harmonized, and light and shadow generation is performed on the synthetic image according to the light and shadow features of the foreground image and the light and shadow features of the background image, to obtain a harmonized target synthetic image with light and shadow.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a diagram showing an application environment of a synthetic image processing method according to an embodiment;

FIG2 is a schematic diagram of a flow chart of a synthetic image processing method in one embodiment;

FIG3 is a schematic diagram of a process of training an image processing model in one embodiment;

FIG4 is a network structure diagram of an image processing model in one embodiment;

FIG5 is a schematic diagram of a process for determining a perception loss in one embodiment;

FIG6 is a schematic flow chart of a synthetic image processing method in another embodiment;

FIG7 is a schematic diagram of a flow chart of a synthetic image processing method in one embodiment;

FIG8 is a block diagram of a synthetic image processing device in one embodiment;

FIG. 9 is a diagram showing the internal structure of a computer device in one embodiment.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

The synthetic image processing method provided in the embodiment of the present application can be applied to the application environment shown in FIG1. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. The terminal 102 uploads the synthetic image, and the server 104 obtains the synthetic image, the mask of the foreground image, and the mask of the background image through the network, and inputs the synthetic image and the mask of the foreground image into the first generator of the image processing model, extracts the foreground light and shadow features from the mask of the foreground image, and obtains the first synthetic image according to the foreground light and shadow features and the synthetic image; the synthetic image and the mask of the background image are input into the second generator of the image processing model, and the background light and shadow features are extracted from the mask of the background image, and the second synthetic image is obtained according to the background light and shadow features and the synthetic image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator; the first synthetic image and the second synthetic image are fused to obtain the target synthetic image. The terminal 102 may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart TVs, smart car-mounted devices, etc. The portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers.

In one embodiment, as shown in FIG2 , a synthetic image processing method is provided, the synthetic image processing method comprising the following steps:

Step 202, obtaining a composite image, a mask of a foreground image, and a mask of a background image.

Among them, the composite image is a composite image formed by extracting the target image from the foreground image and pasting the target image onto another background image. In a broad sense, a composite image is a composite image formed by grafting multiple visual elements from different images onto the same image.

The mask of the foreground image and the mask of the background image can be obtained by the mask operation of the composite image. The mask operation of the image refers to recalculating the value of each pixel in the image through the mask kernel operator. The mask kernel operator describes the influence of the domain pixel points on the new pixel value, and at the same time, the pixel points are weighted averaged according to the weight factor in the mask operator. The mask of the foreground image refers to the value of the pixels in the foreground image in the composite image, and the mask of the background image refers to the value of the pixels in the background image in the composite image.

Step 204, input the composite image and the mask of the foreground image to the first generator of the image processing model, extract the foreground light and shadow features from the mask of the foreground image, and obtain the first composite image according to the foreground light and shadow features and the composite image.

The image processing model is composed of a first generator and a second generator. The first generator is composed of an encoder and a decoder.

Specifically, the masks of the composite image and the foreground image are input into the encoder of the first generator, the foreground light and shadow features are extracted from the mask of the foreground image through the transformer module, and the decoder of the first generator outputs the first composite image based on the foreground light and shadow features and the composite image.

Step 206, input the masks of the composite image and the background image to the second generator of the image processing model, extract the background light and shadow features from the mask of the background image, and obtain a second composite image based on the background light and shadow features and the composite image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator.

The image processing model is composed of a first generator and a second generator. The second generator is mainly used for light and shadow generation. During the training process of the image processing model, the training is performed with the goal of obtaining similar foreground light and shadow features and background light and shadow features. Therefore, the foreground light and shadow features extracted by the first generator of the trained image processing model are similar to the background light and shadow features extracted by the second generator.

Specifically, the masks of the composite image and the background image are input to the encoder of the second generator of the image processing model, the encoder extracts the background light and shadow features from the mask of the background image, and the decoder of the second generator outputs the second composite image according to the background light and shadow features and the composite image. Since the foreground light and shadow features are similar to the background light and shadow features, the light and shadow features of the generated first composite image and the second composite image are similar.

Step 208: fuse the first composite image and the second composite image to obtain a target composite image.

Specifically, based on the first composite image and the second composite image, the foreground image and the background image are fused, and light and shadow are generated on the composite image according to the light and shadow characteristics of the foreground image and the light and shadow characteristics of the background image, so as to obtain a target composite image. Since the light and shadow characteristics of the first composite image and the second composite image are similar, the target composite image obtained by fusing the first composite image and the second composite image is harmonized.

In this embodiment, the foreground light and shadow features and the background light and shadow features are extracted by synthesizing the image, the mask of the foreground image, and the mask of the background image, and a first synthesized image is formed according to the foreground light and shadow features, and a second synthesized image is formed according to the background light and shadow features. Based on the similarity between the foreground light and shadow features and the background light and shadow features, the first synthesized image and the second synthesized image have similar light and shadow features, and the target synthesized image obtained by fusing the first synthesized image and the second synthesized image is harmonized, and light and shadow are generated on the synthesized image according to the light and shadow features of the foreground image and the light and shadow features of the background image, so as to obtain a harmonized target synthesized image with light and shadow.

In one embodiment, as shown in FIG3 , a synthetic image processing method for training an image processing model includes:

Step 302, obtaining an image data set, the image data set includes multiple groups of data, each group of data consists of two triplet data, one triplet is the original synthetic image, the mask of the foreground image and the mask of the background image, and the other triplet is the real image, the real foreground illumination image and the real background illumination image.

The image dataset is a publicly available large-scale synthetic image and harmonized dataset, IH Dataset. The image dataset includes multiple sets of data in the form of six-tuples, that is, each set of data consists of two triplets. One of the triplets is used as input data, including the original synthetic image, the mask of the foreground image, and the mask of the background image, and the other triplets are the real image, the real foreground illumination image, and the real background illumination image.

Step 304, input the masks of the original synthetic image and the foreground image in the image data set to the first generator of the image processing model, extract the foreground light and shadow features from the mask of the foreground image, and obtain the third synthetic image based on the foreground light and shadow features and the original synthetic image.

Step 306, input the original composite image and the mask of the background image to the second generator of the image processing model, extract the background light and shadow features from the mask of the background image, and obtain a fourth composite image based on the background light and shadow features and the composite image.

Specifically, the foreground light and shadow features and the background light and shadow features are extracted through the original composite image, the mask of the foreground image and the mask of the background image, and a third composite image is formed according to the foreground light and shadow features, and a fourth composite image is formed according to the background light and shadow features.

Step 308: merge the third synthesized image and the fourth synthesized image to obtain a harmonized image.

Step 310, calculating the model loss based on the harmonized image, the real image, the real foreground illumination image and the real background illumination image.

The model loss may include illumination feature loss (L _light ), non-illumination feature loss (L _nonlight ), perceptual loss (L _perceptual ), adversarial loss (L _adv ), and absolute value deviation function (L ₁ ).

Specifically, according to the harmoniously processed image, a foreground illumination image of the harmoniously processed image and a background illumination image of the harmoniously processed image are determined. According to the harmoniously processed image, the real image, the foreground illumination image of the harmoniously processed image, the background illumination image of the harmoniously processed image, the real foreground illumination image and the real background illumination image, the model loss is calculated.

Step 312, adjusting the parameters of the first generator and the second generator based on the constraints of the model loss to obtain a trained image processing model; the constraints of the loss function include: the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator.

Specifically, the image dataset is divided into 8:1:1, 80% is used as a training set, 10% is used as a validation set, and 10% is used as a test set. There is no overlap between the foreground objects of the training dataset and the test dataset. The training set in the image dataset is obtained, and the model loss is determined based on the real images in the training set and the harmoniously processed images obtained by fusion processing, and the ordinary parameters of the generator are adjusted according to the model loss. The validation set in the image dataset is obtained, and the harmoniously processed images after fusion processing are obtained according to the data in the validation set. The model loss is determined based on the real images in the validation set and the harmoniously processed images obtained by fusion processing, and the hyperparameters of the generator are adjusted according to the model loss to obtain multiple trained models. The test set in the image dataset is obtained, and multiple trained models are evaluated through the data of the test set, and the model with the highest score is used as the image processing model.

In this embodiment, a harmoniously processed image is obtained through an image data set, and a model loss is determined based on the harmoniously processed image, a true foreground illumination image, and a true background illumination image. Parameters of the first generator and the second generator are adjusted based on the model loss to obtain a trained image processing model.

In one embodiment, the model loss includes: adversarial loss and illumination loss; the model loss is calculated based on the harmonized image, the real image, the real foreground illumination image and the real background illumination image, including: inputting the harmonized image and the real image into the discriminator of the image processing model, the discriminator is used to judge the authenticity of the harmonized image and the real image; according to the output of the discriminator, the adversarial loss is calculated; according to the difference between the real foreground illumination image and the foreground illumination image of the harmonized image, and the difference between the real background illumination image and the background illumination image of the harmonized image, the illumination feature loss is calculated.

Specifically, the adversarial loss _Ladv is the adversarial loss between the generator and the discriminator in the adversarial structure. The discriminator expects the generator to strive to generate realistic images, and the discriminator strives to judge whether the image is real or fake. As shown in Figure 4, after the generator generates the harmonized image, the harmonized image and the real image are input into the discriminator of the image processing model. The discriminator judges the authenticity of the harmonized image and the real image, and outputs the probability that the harmonized image and the real image are "real". The adversarial loss is calculated based on the output.

The adversarial loss is

表示真实图像。Where D(i) is the probability that image i is “real”, Pic represents the target synthetic image,

Represents a real image.

Specifically, the illumination feature loss _Llight is used to represent the element illumination loss between the illumination of the harmonized image and the illumination of the corresponding real image.

表示目标合成图像的前景光照图像，

表示目标合成图像的背景光照图像。Pic_obj表示真实图像的前景光照图像，Pic_background表示真实图像的背景光照图像，

为欧几里得范数的平方，期望光照损失特征尽可能的小。in,

represents the foreground illumination image of the target composite image,

Represents the background illumination image of the target synthetic image. Pic _obj represents the foreground illumination image of the real image, and Pic _background represents the background illumination image of the real image.

is the square of the Euclidean norm, and it is expected that the illumination loss feature is as small as possible.

In this embodiment, the adversarial loss and the illumination feature loss are calculated by using the foreground illumination image of the harmonized image, the foreground illumination image of the real image, the background illumination image of the harmonized image, and the background illumination image of the real image. The adversarial loss is the adversarial loss between the generator and the discriminator in the adversarial structure. The illumination feature loss characterizes the illumination loss between the illumination of the harmonized image and the illumination of the real image. We expect the illumination loss function to be as small as possible so that the illumination loss between the illumination of the target synthetic image and the illumination of the real image is smaller.

In one embodiment, the model loss further includes: non-illumination feature loss;

The synthetic image processing method also includes: calculating the non-illumination feature loss according to the non-illumination features of the same foreground image under different illumination conditions; wherein the goal of the non-illumination feature loss function is to minimize the difference in non-illumination features under different illumination conditions.

Specifically, the non-lighting feature loss L _nonlight is based on the Retinex theory and introduces forced non-lighting feature matching to improve the accuracy of object re-lighting images, expecting that the same object has the same non-lighting (i.e., reflectivity) features under different lighting conditions.

和

是相同前景对象在不同光照条件下的非光照特征，Nnonlight是非光照特征F_nonlight中的元素数，期望非光照特征损失函数尽可能的小。in,

and

is the non-lighting feature of the same foreground object under different lighting conditions, Nnonlight is the number of elements in the non-lighting feature F _nonlight , and it is expected that the non-lighting feature loss function is as small as possible.

In this embodiment, the expression of the non-illumination loss function is given by the non-illumination features of the same foreground object under different illumination conditions and the number of elements in the non-illumination features. It is expected that the smaller the non-illumination feature loss function is, the closer the non-illumination (i.e., reflectivity) features of the same object under different illumination conditions are. In this embodiment, it is hoped that the non-illumination feature loss function is as small as possible so that the same object has closer non-illumination (i.e., reflectivity) features under different illumination conditions.

In one embodiment, the model loss further includes: perceptual loss; as shown in FIG5 , the synthetic image processing method further includes:

Step 502: extracting and harmonizing feature maps of the image using a pre-trained neural network model.

Specifically, the neural network model may be VGG16, which is pre-trained on ImageNet. The pre-trained neural network model is used to extract and harmonize the feature map of the processed image.

Step 504: Use the pre-trained neural network model to extract a feature map of the real image corresponding to the harmonized image.

Specifically, a real image corresponding to the harmonized processed image is determined, and a feature map of the real image is extracted through a pre-trained neural network model.

Step 506, determining the perceptual loss based on the feature map of the harmonized image and the feature map of the real image corresponding to the harmonized image; the goal of the perceptual loss function is to minimize the difference between the feature map of the harmonized image and the feature map of the real image corresponding to the harmonized image.

The perceptual loss function _iper is used to represent the semantic difference between the harmonized image extracted by the neural network model and the real image.

指真实图像。感知损失函数表征和谐处理和真实图像的接近程度，感知损失函数越小，输出的和谐处理图像越接近真实图像。MSE(A, B) is the mean square error of A and B, which is used for error analysis. _{V I} represents the feature map extracted by the pre-trained model, pic _R refers to the output harmonized image,

Refers to the real image. The perceptual loss function represents the closeness between the harmonized image and the real image. The smaller the perceptual loss function, the closer the output harmonized image is to the real image.

In this embodiment, the perceptual loss function is determined by the neural network model, the target synthetic image and the real image, and the perceptual loss function represents the closeness between the target synthetic image and the real image. We hope that the perceptual loss function is as small as possible so that the outputted harmonized processed image is as close to the real image as possible.

In one embodiment, the model loss is a weighted sum of illumination feature loss, non-illumination feature loss, perceptual loss, adversarial loss, and absolute value deviation loss.

The absolute value deviation loss, also known as the L1 norm loss, aims to minimize the sum of the absolute differences between the harmonized image and the true image.

表示真实图像。Among them, Pic means harmoniously processing the image,

Represents a real image.

Specifically, in the actual training process, different weights can be set for the illumination feature loss, non-illumination feature loss, perceptual loss, adversarial loss, and absolute value deviation loss according to the purpose of model training. When the same weights are set, the model loss can be expressed as L _total = L _light + L _nonlight + L _perceptual + L _adv + L ₁ . Among them, L _light is the illumination feature loss, L _nonlight is the non-illumination feature loss, L _perceptual is the perceptual loss, L _adv is the adversarial loss, and L ₁ is the absolute value deviation loss, that is, the classic L1 norm loss.

In the process of training the image processing model, the original synthetic image, the mask of the foreground image and the mask of the background image in the image dataset are used to obtain the harmoniously processed image. According to the harmoniously processed image, the real image, the real foreground illumination image and the real background illumination image, the illumination feature loss, the non-illumination feature loss, the perceptual loss, the adversarial loss and the absolute value deviation loss are calculated to obtain the weighted sum of the illumination feature loss, the non-illumination feature loss, the perceptual loss, the adversarial loss and the absolute value deviation loss, and the image processing model loss is determined. According to the image processing model loss, the parameters of the first generator and the second generator are adjusted to obtain the trained image processing model.

In this embodiment, the model loss is calculated through illumination feature loss, non-illumination feature loss, perception loss, adversarial loss and absolute value deviation loss.

In one embodiment, as shown in FIG6 , a synthetic image processing method is provided, comprising the following steps:

Step 602, obtaining an image data set, the image data set includes multiple groups of data, each group of data consists of two triplet data, one triplet is the original synthetic image, the mask of the foreground image and the mask of the background image, and the other triplet is the real image, the real foreground illumination image and the real background illumination image.

Step 604, designing a loss function, the loss function includes an illumination feature loss function, a non-illumination feature loss function, a perceptual loss function, an adversarial loss function, and an absolute value deviation function.

Step 606, input the masks of the original synthetic image and the foreground image in the image data set to the first generator of the image processing model, extract the foreground light and shadow features from the mask of the foreground image, and obtain the third synthetic image based on the foreground light and shadow features and the original synthetic image.

Step 608, input the original composite image and the mask of the background image to the second generator of the image processing model, extract the background light and shadow features from the mask of the background image, and obtain a fourth composite image based on the background light and shadow features and the composite image.

Step 610: merge the third synthesized image and the fourth synthesized image to obtain a harmonized image.

Step 612, calculating the model loss based on the harmonized image, the real image, the real foreground illumination image and the real background illumination image.

Step 614, adjusting the parameters of the first generator and the second generator based on the constraints of the model loss to obtain a trained image processing model; the constraints of the loss function include: the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator.

Step 616, input the mask of the composite image and the foreground image to the first generator of the image processing model, extract the foreground light and shadow features from the mask of the foreground image, and obtain the first composite image based on the foreground light and shadow features and the composite image.

The composite image is a composite image that needs to be harmonized.

Step 618, input the masks of the composite image and the background image to the second generator of the image processing model, extract the background light and shadow features from the mask of the background image, and obtain a second composite image based on the background light and shadow features and the composite image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator.

Step 620: fuse the first composite image and the second composite image to obtain a target composite image.

In this embodiment, as shown in Figure 7, the synthetic image processing method mainly includes four parts: obtaining an image data set, designing a loss function, training an image processing model, and performing harmonization processing on the synthetic image. First, the image processing model is trained using the image data set to obtain a trained image processing model, and then the synthetic image is harmonized according to the trained image processing model to obtain a target synthetic image.

It should be understood that, although the steps in the flowcharts involved in the above embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

Based on the same inventive concept, the embodiment of the present application also provides a synthetic image processing device for implementing the synthetic image processing method involved above. The implementation solution provided by the device to solve the problem is similar to the implementation solution recorded in the above method, so the specific limitations in one or more synthetic image processing device embodiments provided below can refer to the limitations of the synthetic image processing method above, and will not be repeated here.

In one embodiment, as shown in FIG8 , a synthetic image processing device is provided, including: a synthetic image acquisition module 802, a first synthetic image acquisition module 804, a second synthetic image acquisition module 806 and a target synthetic image acquisition module 808, wherein:

The composite image acquisition module 802 is used to acquire a composite image, a mask of a foreground image, and a mask of a background image.

The first composite image acquisition module 804 is used to input the composite image and the mask of the foreground image into the first generator of the image processing model, extract the foreground light and shadow features from the mask of the foreground image, and obtain the first composite image according to the foreground light and shadow features and the composite image.

The second synthetic image acquisition module 806 is used to input the masks of the synthetic image and the background image into the second generator of the image processing model, extract the background light and shadow features from the mask of the background image, and obtain the second synthetic image based on the background light and shadow features and the synthetic image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator.

The target composite image acquisition module 808 is used to fuse the first composite image and the second composite image to obtain a target composite image.

In one embodiment, a synthetic image processing method for training an image processing model includes:

The image dataset acquisition module is used to acquire an image dataset, which includes multiple groups of data, each group of data consists of two triplet data, one of which is the original synthetic image, the mask of the foreground image and the mask of the background image, and the other triplet is the real image, the real foreground illumination image and the real background illumination image.

The third synthetic image acquisition module is used to input the masks of the original synthetic image and the foreground image in the image data set into the first generator of the image processing model, extract the foreground light and shadow features from the mask of the foreground image, and obtain the third synthetic image based on the foreground light and shadow features and the original synthetic image.

The fourth composite image acquisition module is used to input the masks of the original composite image and the background image into the second generator of the image processing model, extract the background light and shadow features from the mask of the background image, and obtain the fourth composite image based on the background light and shadow features and the composite image.

The harmonized processed image acquisition module fuses the third synthesized image and the fourth synthesized image to obtain a harmonized processed image.

The model loss calculation module is used to calculate the model loss according to the harmonically processed image, the real image, the real foreground illumination image and the real background illumination image.

The image processing model acquisition module is used to adjust the parameters of the first generator and the second generator based on the constraints of the model loss to obtain a trained image processing model; the constraints of the loss function include: the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator.

In one embodiment, the model loss includes: adversarial loss and illumination loss; a model loss calculation module is used to input the harmonized image and the real image into the discriminator of the image processing model, and the discriminator is used to judge the authenticity of the harmonized image and the real image; according to the output of the discriminator, the adversarial loss is calculated; according to the difference between the real foreground illumination image and the foreground illumination image of the harmonized image, and the difference between the real background illumination image and the background illumination image of the harmonized image, the illumination feature loss is calculated.

In one embodiment, the model loss also includes: non-illumination feature loss; the synthetic image processing method also includes: a non-illumination feature loss calculation module, which is used to calculate the non-illumination feature loss based on the non-illumination features of the same foreground image under different illumination; wherein the goal of the non-illumination feature loss function is to minimize the difference in non-illumination features under different illumination conditions.

In one embodiment, the model loss also includes: perceptual loss; the synthetic image processing method also includes: a perceptual loss calculation module, which is used to use a pre-trained neural network model to extract a feature map of the harmoniously processed image; use the pre-trained neural network model to extract a feature map of the real image corresponding to the harmoniously processed image; determine the perceptual loss based on the feature map of the harmoniously processed image and the feature map of the real image corresponding to the harmoniously processed image; the goal of the perceptual loss function is to minimize the difference between the feature map of the harmoniously processed image and the feature map of the real image corresponding to the harmoniously processed image.

In one embodiment, the model loss calculation module is used to calculate the weighted sum of illumination feature loss, non-illumination feature loss, perceptual loss, adversarial loss and absolute value deviation loss.

Each module in the above-mentioned synthetic image processing device can be implemented in whole or in part by software, hardware or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each module.

In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be shown in FIG9 . The computer device includes a processor, a memory, and a network interface connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data related to image processing. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a synthetic image processing method is implemented.

Those skilled in the art will understand that the structure shown in FIG. 9 is merely a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

Get the composite image, the mask of the foreground image, and the mask of the background image;

Inputting the composite image and the mask of the foreground image into a first generator of the image processing model, extracting foreground light and shadow features from the mask of the foreground image, and obtaining a first composite image according to the foreground light and shadow features and the composite image;

Inputting the masks of the composite image and the background image into the second generator of the image processing model, extracting background light and shadow features from the mask of the background image, and obtaining a second composite image according to the background light and shadow features and the composite image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator;

The first composite image and the second composite image are fused to obtain a target composite image.

In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

An image data set is obtained, the image data set includes multiple groups of data, each group of data consists of two triplet data, one of which is an original synthetic image, a mask of a foreground image and a mask of a background image, and the other triplet is a real image, a real foreground illumination image and a real background illumination image; the masks of the original synthetic image and the foreground image in the image data set are input to a first generator of an image processing model, foreground light and shadow features are extracted from the mask of the foreground image, and a third synthetic image is obtained according to the foreground light and shadow features and the original synthetic image; the masks of the original synthetic image and the background image are input to a first generator of the image processing model; The second generator extracts background light and shadow features from the mask of the background image, and obtains a fourth synthetic image based on the background light and shadow features and the synthetic image; the third synthetic image and the fourth synthetic image are fused to obtain a harmoniously processed image; the model loss is calculated based on the harmoniously processed image, the real image, the real foreground lighting image and the real background lighting image; the parameters of the first generator and the second generator are adjusted based on the constraints of the model loss to obtain a trained image processing model; the constraints of the loss function include: the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator.

In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

The harmonized image and the real image are input into the discriminator of the image processing model, and the discriminator is used to judge the authenticity of the harmonized image and the real image; the adversarial loss is calculated according to the output of the discriminator; the illumination feature loss is calculated according to the difference between the real foreground illumination image and the foreground illumination image of the harmonized image, and the difference between the real background illumination image and the background illumination image of the harmonized image.

In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

The method also includes: calculating the non-illumination feature loss according to the non-illumination features of the same foreground image under different illumination conditions; wherein the goal of the non-illumination feature loss function is to minimize the difference in non-illumination features under different illumination conditions.

In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

A feature map of the harmonized image is extracted using a pre-trained neural network model; a feature map of the real image corresponding to the harmonized image is extracted using a pre-trained neural network model; a perceptual loss is determined based on the feature map of the harmonized image and the feature map of the real image corresponding to the harmonized image; the goal of the perceptual loss function is to minimize the difference between the feature map of the harmonized image and the feature map of the real image corresponding to the harmonized image.

In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

The model loss is the weighted sum of illumination feature loss, non-illumination feature loss, perceptual loss, adversarial loss, and absolute value deviation loss.

In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

Get the composite image, the mask of the foreground image, and the mask of the background image;

Inputting the composite image and the mask of the foreground image into a first generator of the image processing model, extracting foreground light and shadow features from the mask of the foreground image, and obtaining a first composite image according to the foreground light and shadow features and the composite image;

Inputting the masks of the composite image and the background image into the second generator of the image processing model, extracting background light and shadow features from the mask of the background image, and obtaining a second composite image according to the background light and shadow features and the composite image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator;

The first composite image and the second composite image are fused to obtain a target composite image.

In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

An image data set is obtained, the image data set includes multiple groups of data, each group of data consists of two triplet data, one of which is an original synthetic image, a mask of a foreground image and a mask of a background image, and the other triplet is a real image, a real foreground illumination image and a real background illumination image; the masks of the original synthetic image and the foreground image in the image data set are input to a first generator of an image processing model, foreground light and shadow features are extracted from the mask of the foreground image, and a third synthetic image is obtained according to the foreground light and shadow features and the original synthetic image; the masks of the original synthetic image and the background image are input to a first generator of the image processing model; The second generator extracts background light and shadow features from the mask of the background image, and obtains a fourth synthetic image based on the background light and shadow features and the synthetic image; the third synthetic image and the fourth synthetic image are fused to obtain a harmoniously processed image; the model loss is calculated based on the harmoniously processed image, the real image, the real foreground lighting image and the real background lighting image; the parameters of the first generator and the second generator are adjusted based on the constraints of the model loss to obtain a trained image processing model; the constraints of the loss function include: the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator.

In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

The harmonized image and the real image are input into the discriminator of the image processing model, and the discriminator is used to judge the authenticity of the harmonized image and the real image; the adversarial loss is calculated according to the output of the discriminator; the illumination feature loss is calculated according to the difference between the real foreground illumination image and the foreground illumination image of the harmonized image, and the difference between the real background illumination image and the background illumination image of the harmonized image.

In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

According to the non-illumination features of the same foreground image under different illumination conditions, the non-illumination feature loss is calculated; wherein the goal of the non-illumination feature loss function is to minimize the difference in non-illumination features under different illumination conditions.

In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

A feature map of the harmonized image is extracted using a pre-trained neural network model; a feature map of the real image corresponding to the harmonized image is extracted using a pre-trained neural network model; a perceptual loss is determined based on the feature map of the harmonized image and the feature map of the real image corresponding to the harmonized image; the goal of the perceptual loss function is to minimize the difference between the feature map of the harmonized image and the feature map of the real image corresponding to the harmonized image.

In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

The model loss is the weighted sum of illumination feature loss, non-illumination feature loss, perceptual loss, adversarial loss, and absolute value deviation loss.

In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the following steps:

Get the composite image, the mask of the foreground image, and the mask of the background image;

Inputting the composite image and the mask of the foreground image into a first generator of the image processing model, extracting foreground light and shadow features from the mask of the foreground image, and obtaining a first composite image according to the foreground light and shadow features and the composite image;

Inputting the masks of the composite image and the background image into the second generator of the image processing model, extracting background light and shadow features from the mask of the background image, and obtaining a second composite image according to the background light and shadow features and the composite image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator;

The first composite image and the second composite image are fused to obtain a target composite image.

In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

An image data set is obtained, the image data set includes multiple groups of data, each group of data consists of two triplet data, one of which is an original synthetic image, a mask of a foreground image and a mask of a background image, and the other triplet is a real image, a real foreground illumination image and a real background illumination image; the masks of the original synthetic image and the foreground image in the image data set are input to a first generator of an image processing model, foreground light and shadow features are extracted from the mask of the foreground image, and a third synthetic image is obtained according to the foreground light and shadow features and the original synthetic image; the masks of the original synthetic image and the background image are input to a first generator of the image processing model; The second generator extracts background light and shadow features from the mask of the background image, and obtains a fourth synthetic image based on the background light and shadow features and the synthetic image; the third synthetic image and the fourth synthetic image are fused to obtain a harmoniously processed image; the model loss is calculated based on the harmoniously processed image, the real image, the real foreground lighting image and the real background lighting image; the parameters of the first generator and the second generator are adjusted based on the constraints of the model loss to obtain a trained image processing model; the constraints of the loss function include: the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator.

In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

The harmonized image and the real image are input into the discriminator of the image processing model, and the discriminator is used to judge the authenticity of the harmonized image and the real image; the adversarial loss is calculated according to the output of the discriminator; the illumination feature loss is calculated according to the difference between the real foreground illumination image and the foreground illumination image of the harmonized image, and the difference between the real background illumination image and the background illumination image of the harmonized image.

In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

According to the non-illumination features of the same foreground image under different illumination conditions, the non-illumination feature loss is calculated; wherein the goal of the non-illumination feature loss function is to minimize the difference in non-illumination features under different illumination conditions.

In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

A feature map of the harmonized image is extracted using a pre-trained neural network model; a feature map of the real image corresponding to the harmonized image is extracted using a pre-trained neural network model; a perceptual loss is determined based on the feature map of the harmonized image and the feature map of the real image corresponding to the harmonized image; the goal of the perceptual loss function is to minimize the difference between the feature map of the harmonized image and the feature map of the real image corresponding to the harmonized image.

In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

The model loss is the weighted sum of illumination feature loss, non-illumination feature loss, perceptual loss, adversarial loss, and absolute value deviation loss.

Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims (10)

1. A synthetic image processing method, characterized in that the method comprises: Get the composite image, the mask of the foreground image, and the mask of the background image; Inputting the composite image and the mask of the foreground image into a first generator of an image processing model, extracting foreground light and shadow features from the mask of the foreground image, and obtaining a first composite image according to the foreground light and shadow features and the composite image; Inputting the composite image and the mask of the background image into the second generator of the image processing model, extracting background light and shadow features from the mask of the background image, and obtaining a second composite image according to the background light and shadow features and the composite image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator; The first composite image and the second composite image are fused to obtain a target composite image. 2. The method according to claim 1, characterized in that the method of training the image processing model comprises: Acquire an image data set, the image data set comprising multiple groups of data, each group of data consisting of two triplet data, one of the triplet is an original synthetic image, a mask of a foreground image and a mask of a background image, and the other triplet is a real image, a real foreground illumination image and a real background illumination image; Inputting the original synthesized image and the mask of the foreground image in the image data set into a first generator of the image processing model, extracting foreground light and shadow features from the mask of the foreground image, and obtaining a third synthesized image according to the foreground light and shadow features and the original synthesized image; Inputting the original composite image and the mask of the background image into a second generator of an image processing model, extracting background light and shadow features from the mask of the background image, and obtaining a fourth composite image according to the background light and shadow features and the composite image; fusing the third synthesized image and the fourth synthesized image to obtain a harmoniously processed image; Calculating a model loss according to the harmonized image, the real image, the real foreground illumination image, and the real background illumination image; The parameters of the first generator and the second generator are adjusted based on the constraint of the model loss to obtain a trained image processing model; the constraint of the loss function includes: the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator. 3. The method according to claim 2, characterized in that the model loss includes: adversarial loss and illumination loss; Calculating a model loss according to the harmonized image, the real image, the real foreground illumination image, and the real background illumination image includes: Inputting the harmonized processed image and the real image into a discriminator of the image processing model, wherein the discriminator is used to judge the authenticity of the harmonized processed image and the real image; Calculating adversarial loss based on the output of the discriminator; The illumination feature loss is calculated according to the difference between the real foreground illumination image and the foreground illumination image of the harmonized image, and the difference between the real background illumination image and the background illumination image of the harmonized image. 4. The method according to claim 3, characterized in that the model loss further comprises: non-illumination feature loss; The method further includes: calculating a non-illumination feature loss based on the non-illumination features of the same foreground image under different illumination conditions; wherein the goal of the non-illumination feature loss function is to minimize the difference in non-illumination features under different illumination conditions. 5. The method according to claim 4, characterized in that the model loss further comprises: perceptual loss; The method further comprises: Extracting a feature map of the harmonized image using a pre-trained neural network model; Extracting a feature map of a real image corresponding to the harmonized image using a pre-trained neural network model; The perceptual loss is determined based on the feature map of the harmonized image and the feature map of the real image corresponding to the harmonized image; the goal of the perceptual loss function is to minimize the difference between the feature map of the harmonized image and the feature map of the real image corresponding to the harmonized image. 6. The method according to claim 5 is characterized in that the model loss is a weighted sum of illumination feature loss, non-illumination feature loss, perceptual loss, adversarial loss and absolute value deviation loss. 7. A synthetic image processing device, characterized in that the device comprises: A composite image acquisition module, used to acquire a composite image, a mask of a foreground image, and a mask of a background image; a first synthetic image acquisition module, configured to input the synthetic image and the mask of the foreground image into a first generator of an image processing model, extract foreground light and shadow features from the mask of the foreground image, and obtain a first synthetic image according to the foreground light and shadow features and the synthetic image; A second synthetic image acquisition module, used for inputting the synthetic image and the mask of the background image into the second generator of the image processing model, extracting background light and shadow features from the mask of the background image, and obtaining a second synthetic image according to the background light and shadow features and the synthetic image; wherein the foreground light and shadow features extracted by the first generator of the image processing model are similar to the background light and shadow features extracted by the second generator; The target composite image acquisition module is used to fuse the first composite image and the second composite image to obtain a target composite image. 8. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, wherein the processor implements the steps of the method according to any one of claims 1 to 6 when executing the computer program. 9. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented. 10. A computer program product, comprising a computer program, characterized in that when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images