An automatic generation system, method, computer device and readable storage medium for entertainment pictures

Abstract

The present invention discloses an automatic generation system, method, computer device and readable storage medium for entertainment pictures. The system includes a model library construction module, a picture attribute library establishment module and a picture automatic generation module. The model library construction module is configured to be able to construct a detection model. The picture attribute library establishment module is configured to be able to use the constructed model to perform label prediction on each picture. Each picture and the respective labels obtained by model prediction thereof constitute the picture attribute library. The picture automatic generation module is configured to be able to automatically generate corresponding entertainment pictures. The method of the present invention has high efficiency and good effect, and improves the reusability of excellent designs by professionals.

Description

Technical Field

[0001] The present invention specifically relates to a system, method, computer device and readable storage medium for automatically generating entertainment pictures. Background Art

[0002] In the past, the content on the WEB was all created by humans. With the development of Internet technology and deep learning technology, machine-generated content (MGC) has gradually replaced user-generated content (UGC) or professionally produced content (PGC) in some fields. However, currently, MGC still requires manual supervision in some steps. Therefore, in most cases, UGC or PGC still dominates.

[0003] Traditional entertainment pictures in the field require professional artists and designers to manually collect a large number of picture, text and other materials, and create them using professional image processing software such as PS. This process is very laborious and time-consuming. At the same time, the designs of professional artists have not been effectively formed into templates and cannot be reused by others. Summary of the Invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a system, method, computer device and readable storage medium for automatically generating entertainment pictures.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An entertainment picture automatic generation system includes a model library construction module, a picture attribute library establishment module and a picture automatic generation module, and the model library construction module, the picture attribute library establishment module and the picture automatic generation module are connected in sequence;

[0007] The model library construction module is configured to be able to construct a detection model. The picture attribute library establishment module is configured to be able to use the constructed model to perform label prediction on each picture. Each picture and its respective labels obtained through model prediction constitute the picture attribute library. The picture automatic generation module is configured to be able to automatically generate corresponding entertainment pictures.

[0008] An entertainment picture automatic generation method uses the above-mentioned entertainment picture automatic generation system, and includes the following steps:

[0009] (1) Construct a model library;

[0010] (2) Establish a picture attribute library;

[0011] (3) Automatically generate entertainment pictures.

[0012] Further, in step (1), basic models for face detection, face pose estimation, human pose estimation, human integrity judgment, star recognition, age recognition, and human segmentation are constructed.

[0013] Further, in step (1), the construction of the human pose estimation and human integrity judgment model:

[0014] (1.1) Prepare pictures of people.

[0015] (1.2) Use the open-source YOLOv5 model trained on the COCO dataset to perform human detection on the pictures in step (1.1), remove pictures with multiple people, and only keep pictures of single people.

[0016] (1.3) Use the trained open-source AlphaPose model to perform human key point detection on the pictures obtained in step (1.2), and automatically pre-annotate the human pose and integrity according to the human key point information. The human pose is divided into four categories: sitting, standing, lying, and others, and the integrity is divided into four categories: headshot, half-body photo, full-body photo, and others.

[0017] (1.4) Manually check and modify the pre-annotated results in step (1.3) to obtain a dataset for model training and testing.

[0018] (1.5) Use the multi-task framework to train the EfficientNet model on the training set obtained in step (1.4). The final loss function of the model is L total = λ1L pose + λ2L completion , where L pose represents the loss function of human pose judgment, and L completion represents the loss function of human integrity judgment. Both L pose and L completion adopt cross-entropy loss. λ1 and λ2 are the weight parameters of L pose and L completion respectively, and λ1 = λ2 = 1 is set. The cross-entropy loss function is:

[0019] where N is the number of samples, C is the number of categories, is the one-hot encoded representation of the label of the i-th sample, refers to the probability values of each label of the predicted i-th sample, and k = 0 ~ C - 1;

[0020] During the model training process, the parameter settings are as follows: Batch size is 32, the initial learning rate is 0.01, weight decay is 1e -5 , and momentum is 0.9. The model can be trained until convergence.

[0021] (1.6) During prediction, input the picture into the EfficientNet trained in step (1.5) to obtain the human pose prediction value and the human integrity prediction value respectively.

[0022] Further, in step (1), the construction of the star recognition model:

[0023] (1.1) Prepare star pictures;

[0024] (1.2) Manually screen and filter the star pictures prepared in step (1.1);

[0025] (1.3) Train a face recognition model on the dataset obtained in step (1.2), using ResNet50 + CircleLoss;

[0026] (1.4) Select several pictures at different angles for each star from the dataset obtained in step (1.2), use the face detection model CenterFace to extract the face regions in these pictures, and input them into the model trained in step (1.3) to extract features X = {x1,..., x 10}, and then calculate the average value of these features to obtain the features of each star;

[0027] (1.5) During prediction, for each incoming picture, first perform face detection to obtain the face region, then input it into the model trained in step (1.3) to extract features, and then calculate the Euclidean distance between this feature and each feature in the star feature library obtained in step (1.4). The star corresponding to the feature with the smallest Euclidean distance is the possible candidate star. If the smallest Euclidean distance is less than the preset threshold, the match is successful, and the name of the corresponding star is returned; otherwise, it means that the picture does not match any star in the library.

[0028] Further, in step (1), the construction of the human segmentation model:

[0029] (1.1) The training data uses the open-source Supervisely_Person dataset and the manually annotated picture set; (1.2) Train the DeepLabv3+ model with ResNet50 as the backbone on the dataset in step (1.1), and the loss function uses the weighted cross-entropy loss function; the expression of the weighted cross-entropy loss function is:

[0030]

[0031] where, N is the number of samples, C is the number of classes, is the one-hot encoding representation of the label of the i-th sample, is the probability value of each label of the i-th predicted sample, k = 0 to C-1, α k =[α0,..., α C-1 are the weight parameters for each label category;

[0032] During the model training process, the parameter settings are as follows: the Batch size is 16, the initial learning rate is 0.01, the weight decay is 1e -5 , and the momentum is 0.9; the model can be trained until convergence;

[0033] (1.3) During prediction, each image is input into the trained DeepLabv3+ model in step (1.2) to predict the human mask.

[0034] Furthermore, the automatic generation of entertainment pictures in step (3) specifically includes: template selection, obtaining meta information from the database, selecting picture materials, and generating pictures.

[0035] Furthermore, the process of establishing the picture attribute library in step (2): using the model in step (1) to perform label prediction on each picture respectively, each picture and the respective labels obtained by its model prediction constitute the star picture attribute library, and its attribute types include name, age, human pose, face frame, face pose, and human integrity.

[0036] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it can implement the steps in the above-mentioned automatic generation method of entertainment pictures.

[0037] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can implement the steps in the above-mentioned automatic generation method of entertainment pictures.

[0038] The beneficial effects of the present invention are:

[0039] (1) By using structured data such as the knowledge base in the entertainment field, combining with the template library designed by a large number of professionals, and using deep learning technologies such as face detection, face pose estimation, human pose judgment, human integrity judgment, star recognition, age recognition, human segmentation, and image processing technologies, the present invention automatically generates pictures related to the entertainment field such as movie posters, stills of characters, star comparisons, star event timelines, star age timelines, and views and evaluations of big Vs. Compared with the manual production by professional graphic designers and designers, the method of the present invention has high efficiency (generally, it can automatically generate pictures related to the entertainment field within 2 seconds), good effects, low labor costs, greatly reduces the threshold of people's picture design, and improves the reusability of excellent designs by professionals.

[0040] (2) In the present invention, applying MGC in the field of entertainment picture generation can automatically generate pictures such as movie posters, still characters, star comparisons, star event timelines, star age timelines, opinions and evaluations of big Vs, etc. Description of the Drawings

[0041] Figure 1 It is a flowchart of the construction of the model.

[0042] Figure 2 It is a flowchart of model training, construction of the star feature library, and star recognition.

[0043] Figure 3 It is a schematic diagram of an automatically generated movie poster.

[0044] Figure 4 It is a spliced picture of still characters automatically generated.

[0045] Figure 5 It is a picture of the automatically generated star age axis.

[0046] Figure 6 It is a schematic diagram of a computer device.

[0047] Figure 7 It is a schematic diagram of the opinion of a big V.

[0048] Figure 8 It is a schematic diagram of training the EfficientNet model using the multi-task framework. Detailed Embodiments

[0049] The technical solution of the present invention will be further described in detail below with reference to the drawings. It should be noted that the detailed embodiments are only detailed descriptions of the present invention and should not be regarded as limitations of the present invention.

[0050] Embodiment 1

[0051] An entertainment picture automatic generation system includes a model library construction module, a picture attribute library establishment module, and a picture automatic generation module;

[0052] The model library construction module is configured to be able to construct a detection model. The picture attribute library establishment module is configured to be able to use the constructed detection model to perform label prediction on each picture. Each picture and the respective labels obtained by model prediction thereof constitute the picture attribute library. The picture automatic generation module is configured to be able to automatically generate corresponding entertainment pictures.

[0053] The attribute types in the picture attribute library include name, age, body pose, face frame, face pose (in three directions of Yaw, Pitch, and Roll), and body integrity.

[0054] The entertainment pictures may be movie poster pictures, still photo character splicing pictures, big V opinion pictures, star comparison pictures, star event timeline pictures, star age timeline pictures, and so on.

[0055] An automatic generation method for entertainment pictures, using the above-mentioned automatic generation system for entertainment pictures, includes the following steps:

[0056] (1) Construct a model library;

[0057] (2) Establish a picture attribute library;

[0058] (3) Automatically generate entertainment pictures.

[0059] In some preferred ways, for the generation tasks of pictures such as movie posters, still photo characters, star comparisons, star event timelines, star age timelines, big V opinions and evaluations, etc., basic models such as face detection, face pose estimation, human pose judgment, human integrity judgment, star recognition, age recognition, and human segmentation are constructed.

[0060] In some preferred ways, in the basic model library construction module, the following model structures are adopted, such as CenterFace, FSA-Net, HRNet, EfficientNet, ResNet50+CircleLoss, DeepLabv3+, etc. Among them, the face detection and face pose estimation directly use the models trained on the open-source standard datasets, and the human pose judgment, human integrity judgment, star recognition, age recognition, and human segmentation models are all retrained on the data of the real scene.

[0061] The construction process of the model is as Figure 1 shown. For each task, select a suitable model structure, evaluate it on the data of the real scene. If the effect requirements are met, directly deploy the model; otherwise, manually annotate the data of the real scene, prepare the training dataset, and then retrain the model on these data. After meeting the effect requirements, deploy the model. For example, star recognition, this scenario is a special scenario, and the effects of general open-source models are not very satisfactory. Manual annotation is to manually annotate the pictures, such as judging human integrity, and manually judge which one of the headshot, half-body photo, full-body photo, and others each picture belongs to.

[0062] In some preferred ways, the human face pose is divided into the turning direction (Yaw), nodding direction (Pitch), and tilting direction (Roll). Further, the turning direction is divided into five directions: left, left-middle, middle, right-middle, and right; the nodding direction is divided into three directions: looking down, straight ahead, and looking up; the tilting direction is divided into three directions: left, middle, and right; the human body integrity is divided into four categories: headshot, half-body photo, full-body photo, and others; the human body pose is divided into four categories: sitting, standing, lying, and others.

[0063] The data preparation and model training of models such as human body pose judgment, human body integrity judgment, star recognition, age recognition, and human body segmentation will be described in detail below. The main process is as Figure 1 shown.

[0064] (I) Human Body Pose Judgment and Human Body Integrity Judgment Model

[0065] (1.1) Prepare a large number of human pictures, such as crawling them from search engines like Baidu.

[0066] (1.2) Use the open-source YOLOv5 model trained on the COCO dataset to detect humans in the pictures in step (1.1), remove pictures with multiple people, and only keep pictures of single individuals.

[0067] (1.3) Use the open-source AlphaPose model trained on the COCO dataset to detect human key points in the pictures obtained in step (1.2), and automatically pre-annotate the human body pose and integrity according to the human key point information. The pose is divided into four categories: sitting, standing, lying, and others, and the integrity is divided into four categories: headshot, half-body photo, full-body photo, and others.

[0068] (1.4) Manually check and modify the pre-annotated results in step (1.3) to obtain a dataset for model training and testing. In actual implementation, a total of 25,000 pictures were annotated, of which 20,000 were used as the training set and the remaining 5,000 were used as the test set.

[0069] (1.5) Use the multi-task framework to train the EfficientNet model on the training set obtained in step (1.4), as Figure 8 shown. The final loss function of the model is L total = λ1L pose + λ2L completion , where L pose represents the loss function of human body pose judgment, L completion represents the loss function of human body integrity judgment, and both L pose and L completion adopt cross-entropy loss. λ1 and λ2 are the weights of L pose and Lcompletion The weight parameters, in actual implementation, set λ1 = λ2 = 1;

[0070] The cross-entropy loss function is:

[0071] where N is the number of samples, C is the number of classes, is the one-hot encoded representation of the label of the i-th sample, refers to the probability values of each label of the predicted i-th sample, k = 0 to C - 1.

[0072] In this embodiment, during the model training process, the parameter settings are as follows: Batch size is 32, the initial learning rate is 0.01, weight decay is 1e -5 , and momentum is 0.9. The model can be trained until convergence.

[0073] (1.6) During prediction, input each incoming picture into the EfficientNet trained in step (1.5) to obtain the human pose prediction value and the human integrity prediction value respectively.

[0074] Using the EfficientNet model trained by the present invention for human pose judgment, the accuracy rate is 96%, and using the EfficientNet model trained by the present invention for human integrity judgment, the accuracy rate is 98%.

[0075] (2) Celebrity recognition model

[0076] (2.1) Prepare a large number of celebrity pictures, such as crawling from search engines like Baidu or scraping from some dedicated websites;

[0077] (2.2) Manually screen and filter the celebrity pictures prepared in step (2.1), and only retain the pictures of a certain celebrity's face (for example, when training the model to recognize celebrity A, only retain the pictures of celebrity A's face), and select pictures with different poses. In actual implementation, finally, 600 pictures are retained for each celebrity;

[0078] (2.3) Train a face recognition model on the dataset obtained in step (2.2) (in actual implementation, use ResNet50 + CircleLoss, and other models can also be used), where there are 500 pictures of each celebrity in the training set and 100 pictures of each celebrity in the test set;

[0079] (1.4) For the dataset obtained in step (1.2), several pictures at different angles are selected for each star (in actual implementation, 10 pictures are selected for each star). The face detection model CenterFace is used to extract the face regions in these pictures, and they are input into the model trained in step (1.3) to extract features X = {x1,..., x 10}, and then the average value of these features is calculated to obtain the features of each star; different angles refer to: the left - right orientation of the face, looking down and up, and turning the head, etc.

[0080] (1.5) During prediction, for each incoming picture, first perform face detection to obtain the face region, then input it into the model trained in step (1.3) to extract features, and then calculate the Euclidean distance between this feature and each feature in the star feature library obtained in step (1.4). The star corresponding to the feature with the smallest Euclidean distance is the possible candidate star. If the smallest distance is less than a preset threshold (in actual implementation, this threshold is set to 0.8), then the matching is successful and the name of the corresponding star is returned; otherwise, it means that the picture does not match any star in the library.

[0081] (III) Age recognition model

[0082] (1.1) The training dataset uses the open - source AFAD and MegaAge - Asian datasets;

[0083] (1.2) Pre - process the data in step (1.1): First, use the CenterFace model to perform face detection and face key - point prediction on the pictures, then align the face based on the detected face key - points, and then crop and resize the face region to obtain a 224x224 face picture;

[0084] (1.3) Train the DLDL - v2 (ThinAgeNet) model on the dataset in step (1.2);

[0085] (1.4) During prediction, input each incoming picture into the ThinAgeNet model trained in step (1.3) for age prediction;

[0086] (IV) Human body segmentation model

[0087] (1.1) The training data uses the open - source Supervisely_Person dataset and 5000 manually annotated pictures;

[0088] (1.2) Train the DeepLabv3+ model with ResNet50 as the backbone on the dataset in step (1.1), and the loss function uses weighted cross - entropy loss;

[0089] The expression of weighted cross-entropy is as follows:

[0090]

[0091] Among them, N is the number of samples, C is the number of categories, is the one-hot encoded representation of the label of the i-th sample, is the probability value of each label of the predicted i-th sample, k = 0 to C-1, α k = [α0,..., α C-1 are the weight parameters of each label category.

[0092] In this embodiment, during the model training process, the parameter settings are as follows: the batch size is 16, the initial learning rate is 0.01, the momentum is 0.9, and the weight decay is 1e -5 . The model can be trained until convergence.

[0093] (1.3) During prediction, each picture is input into the trained DeepLabv3+ model in step (1.2) to predict the human mask.

[0094] In the present invention, the MIoU of the trained DeepLabv3+ model is 0.96, which can better predict the human mask.

[0095] In the present invention, the models adopted are all conventional models in the prior art, and no improvement is made to the model structure itself.

[0096] In some preferred ways, in the process of establishing the picture attribute library in step (2): for each photo in the database, the model in step (1) is used to predict the labels of each picture respectively. Each picture and the respective labels predicted by the model constitute the star picture attribute library, and its attribute types include name, age, human pose, face box, face pose (Yaw, Pitch, and Roll in three directions), and human integrity, as shown in the following table.

[0097] Table 1 Attribute types in the star picture attribute library

[0098]

[0099] In some preferred embodiments, the automatic generation of entertainment pictures in step (3) includes: template selection, obtaining meta information from the database, selecting picture materials, and generating pictures. The templates are designed by designers and / or collected online. The selection of picture materials includes picture filtering and obtaining candidate pictures. Picture filtering means filtering out pictures that do not meet the requirements of the template, and each template is different. For example, there are many pictures of stars, but only a half-body picture facing forward may be needed. In this case, pictures that meet this requirement need to be selected from the star pictures.

[0100] The following specifically describes the process of picture generation in the system.

[0101] (1) Movie poster generation:

[0102] The user needs to select a poster template and input the movie name. The system automatically extracts the meta information related to the movie from the entertainment domain knowledge base, including the movie cover, director, screenwriter, lead actors, movie type, release time, running time, etc. Then the system performs corresponding rendering according to the template selected by the user. An example of automatically generating a movie poster is as follows Figure 3 shown. The entertainment domain knowledge base refers to various film and television and entertainment-related information crawled from the Internet, which is structured and then stored in the database. In this embodiment, it takes about 200 ms to automatically generate a movie poster.

[0103] (2) The steps for generating a spliced picture of stills of characters are as follows:

[0104] (1) According to the TV drama name input by the user, the system automatically extracts the list of lead actors of the TV drama from the entertainment domain knowledge base;

[0105] (2) According to the template selected by the user, extract pictures that meet the template requirements from the star picture attribute library; Taking the following Figure 4 template as an example, the template requires that the face of the person in the middle faces forward, the face of the person on the left faces right, the face of the person on the right faces left, and upper body pictures of the human body are needed. The requirement for the face orientation can be determined by the values of the face pose (Yaw, Roll, and Pitch in three directions), and the upper body pictures of the human body can be screened according to the value of the human body integrity. The human body integrity is Completion in the above table, and the Completion value includes headshot, half-body shot, full-body shot, and others.

[0106] (3) Render the selected pictures according to the template. The automatically generated spliced picture of stills of characters is as follows Figure 4 shown. In this embodiment, it takes about 1 s to automatically generate a spliced picture of stills of characters.

[0107] (3) The steps for generating a star age axis picture are as follows:

[0108] (1) Extract the star's picture from the star image attribute library according to the star name input by the user;

[0109] (2) Sort the pictures by age from youngest to oldest. If there are multiple pictures of the same age, select one at random.

[0110] (3) Arrange the pictures in step 2 in sequence according to the template selected by the user to generate the star age axis picture. An example of a generated star age axis picture is as follows Figure 5 In this embodiment, it takes about 800ms to automatically generate the star age axis image.

[0111] (IV) The steps for generating the opinion picture of a big V are as follows:

[0112] (1) Extract the celebrity’s picture from the celebrity picture attribute library based on the celebrity’s name input by the user;

[0113] (2) Apply the trained DeepLabv3+ model to the image selected in step (1) to segment the large V human body and obtain the large V human body mask;

[0114] (3) Based on the template selected by the user and the input content of the big V opinion, combined with the big V body mask obtained in step (2), the big V opinion picture is rendered; an example of automatically generating a big V opinion picture is as follows Figure 7 In this embodiment, it takes about 500ms to automatically generate the opinion picture of the influencer.

[0115] Example 2, refer to the attached Figure 6 .

[0116] In this embodiment, a computer device 100 is provided, including a memory 102, a processor 101, and a computer program 103 stored in the memory 102 and executable on the processor 101. When the processor 101 executes the computer program 103, the steps of the method for automatically generating entertainment pictures provided in the above-mentioned embodiment 1 can be implemented.

[0117] Example 3

[0118] In this embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the method for automatically generating entertainment pictures provided in the above embodiments can be implemented.

[0119] In this embodiment, the computer program may be the computer program in Embodiment 2.

[0120] In this embodiment, the computer-readable storage medium can be executed by the computer device in Example 2.

[0121] Those of ordinary skill in the art can understand that implementing all or part of the processes in the above embodiments can be achieved by instructing relevant hardware through a computer program. This 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 methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided by the present invention can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0122] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as falling within the scope described in this specification.

[0123] The above-described embodiments merely represent several implementation manners of the present invention. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of protection. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be subject to the appended claims.

Claims

1. An automatic generation system for entertainment pictures, characterized in that, It includes a model library construction module, a picture attribute library establishment module, and a picture automatic generation module; The system includes: (1) Construct a model library, which includes face detection, face pose estimation, human pose estimation, human integrity judgment, celebrity recognition, age recognition, and human segmentation models; Among them, the construction of the human pose estimation and human integrity judgment model includes: (1.1a) Prepare pictures of people; (1.2a) Use the open-source YOLOv5 model to perform human detection on the pictures in step (1.1a); (1.3a) Use the trained open-source AlphaPose model to perform human key point detection on the pictures obtained in step (1.2a), and automatically pre-label the human pose and integrity according to the human key point information. Among them, the human pose is divided into four categories: sitting, standing, lying, and others, and the integrity is divided into four categories: headshot, half-body photo, full-body photo, and others; (1.4a) Check and modify the pre-labeled results in step (1.3a) to obtain a data set for model training and testing; (1.5a) Train the EfficientNet model using the multi-task framework on the training set obtained in step (1.4a), and the final loss function of the model is L total = λ1L pose + λ2L completion , where L pose represents the loss function for human pose judgment, and L completion represents the loss function for human integrity judgment. Both L pose and L completion use the cross-entropy loss function. λ1 and λ2 are the weight parameters of L pose and L completion respectively, and λ1 = λ2 = 1 is set; The cross-entropy loss function is as follows: ; Among them, N is the number of samples, C is the number of categories, is the onehot encoding representation of the label of the i-th sample, is the predicted probability value of each label of the i-th sample, k = 0 ~ C-1; (1.6a) Input the pictures into the trained EfficientNet in step (1.5a) to obtain human pose prediction values and human integrity prediction values respectively; The construction of the celebrity recognition model includes: (1.1b) Prepare pictures of celebrities; (1.2b) Manually screen and filter the pictures of celebrities prepared in step (1.1b); (1.3b) Train a face recognition model on the data set obtained in step (1.2b), using ResNet50+CircleLoss; (1.4b) For each star, select several pictures from different angles from the dataset obtained in step (1.2b). Use the face detection model CenterFace to extract the face regions in these pictures, and input them into the model trained in step (1.3b) to extract features X = {x1,..., x 10}, calculate the average value of these features to obtain the features of each star; (1.5b) For each incoming picture, first perform face detection to obtain the face area, and then input it into the model trained in step (1.3b) to extract features. Calculate the Euclidean distance between this feature and each feature in the celebrity feature library obtained in step (1.4b). The celebrity corresponding to the feature with the smallest Euclidean distance is the possible candidate celebrity. If the smallest Euclidean distance is less than the preset threshold, the match is successful, and the name of the corresponding celebrity is returned. Otherwise, it means that the picture does not match any celebrity in the library; (2) Establish a picture attribute library: Use the constructed model to perform label prediction on each picture. Each picture and its various labels predicted by the model constitute the picture attribute library; (3) Automatically generate entertainment pictures.

2. The automatic generation system of entertainment pictures according to claim 1, characterized in that, In step (1), the construction of the human segmentation model: (1.1c) The training data uses the open-source Supervisely_Person data set and the manually annotated picture set; (1.2c) Train the DeepLabv3+ model with ResNet50 as the backbone on the data set in step (1.1c), and the loss function uses the weighted cross-entropy loss function; The expression of the weighted cross-entropy loss function is as follows: ; Where N is the number of samples, C is the number of categories, is the one-hot encoded representation of the label of the i-th sample, are the probability values of each label of the predicted i-th sample, k = 0 to C-1, α k = [α0,..., α c-1 ] are the weight parameters for each label category; (1.3c) Input each picture into the trained DeepLabv3+ model in step (1.2c) to predict the human mask.

3. The automatic generation system of entertainment pictures according to claim 1, characterized in that, Step (3) for automatically generating entertainment pictures specifically includes: template selection, obtaining meta information from the database, picture material selection, and picture generation.

4. The automatic generation system of entertainment pictures according to claim 1, characterized in that, Process of step (2) for establishing the picture attribute library: Use the model in step (1) to perform label prediction on each picture respectively. Each picture and its respective labels obtained through model prediction constitute the star picture attribute library, and its attribute types include name, age, human body posture, face box, face posture, and human body integrity.

Images