A video synthesis method and system for high-fidelity real-time rendering

Abstract

The present invention discloses a high-fidelity real-time rendering video synthesis method and system, and relates to the technical field of video synthesis. The method of the present invention realizes high synchronization of audio and video through precise audio processing and precise provision of head and torso features. By utilizing advanced NeRF technology, the synthesized character movements are realistic, including lip movements, facial expressions and whole-body movements. Through the steps of timbre extraction, conversion, embedding and synthesis, the present invention can replicate the timbre characteristics of a specific speaker and provide personalized voice output. The method of the present invention has real-time rendering capability and is suitable for application scenarios requiring rapid response, such as live broadcast, video conferencing, etc., and supports the generation of high-resolution video streams and provides clear image details.

Description

Technical Field

[0001] The present invention relates to the technical field of video synthesis, and more specifically, to a high-fidelity real-time rendering video synthesis method and system. Background Art

[0002] At present, video synthesis technology mainly relies on traditional computer graphics methods and some preliminary machine learning techniques. These technologies perform well when processing static images and simple animations, but face many challenges when processing complex dynamic videos, especially portrait videos that require real-time rendering and high fidelity. First, existing technologies are often unable to cope with real-time rendering. In application scenarios that require fast response, such as live broadcast or real-time video communication, traditional methods may not be able to complete high-quality image rendering in a timely manner. Secondly, although some systems can achieve portrait extraction and synthesis, they usually require a lot of computing resources and complex settings, which limits their application on resource-constrained devices. Furthermore, existing technologies often have difficulty maintaining high-fidelity rendering effects when dealing with different lighting conditions, complex backgrounds, and dynamic expression changes. In addition, most systems focus on a single language or a specific type of video content, and lack adaptability to multilingual and diverse video content. Summary of the invention

[0003] In view of the above problems, the present invention provides a high-fidelity real-time rendering video synthesis system. The system adopts advanced deep learning algorithms, especially Neural Radiance Field (NeRF) technology, to quickly learn and simulate the portrait features in a specific video from a small number of samples. Through optimized data processing flow and real-time rendering technology, the system can achieve fast response and real-time rendering while maintaining high fidelity, significantly improving the efficiency and quality of video synthesis.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A high-fidelity real-time rendering video synthesis method comprises the following steps:

[0006] D1: Collect video material;

[0007] D2: Preprocess the collected video materials;

[0008] D3: Separate the audio from the video material and perform noise reduction and enhancement processing;

[0009] D4: Extract key frames from the video stream of the video material;

[0010] D5: Analyze the relationship between lip movements and speech through machine learning algorithms, and train a model to simulate the characteristics of a person's head;

[0011] D6: Obtain the trained head features and key frames, use computer vision technology to extract the key points of the character's torso, and train a model that perceives and simulates the character's torso movements;

[0012] D7: collect the audio signal to be synthesized;

[0013] D8: pre-process the audio signal to be synthesized;

[0014] D9: Obtain the head feature information obtained through training and transmit it to the inference renderer S8;

[0015] D10: Obtain the trained torso perception feature information and transmit it to the inference renderer S8;

[0016] D11: Real-time rendering is performed using advanced neural radiance field technology based on the audio signal to be synthesized, head feature information, and torso perception feature information.

[0017] Further, a high-fidelity real-time rendering video synthesis system includes a video collector S1: collecting target video samples in real time;

[0018] Video processor S2: receives the video data collected by video collector S1, decomposes it into audio stream and video stream, and obtains key frames based on the video stream;

[0019] Speech-to-lip action model S3: uses a generative adversarial network to convert audio signals into lip action sequences;

[0020] Head Trainer S4: Based on the 3D deformable model, Head Trainer S4 extracts head posture and expression features from key frames;

[0021] Trunk sensor S5: extracts the key points of the trunk by analyzing the trunk area of ​​the character in the key frame;

[0022] Trunk prediction model S6: Using a long short-term memory network, the trunk prediction model S6 predicts the movement of the trunk in future frames based on the output of the trunk sensor S5;

[0023] Audio input device S7: As an audio input interface, the S7 audio input device is responsible for receiving the user's audio input;

[0024] Inference Renderer S8: It is the core rendering module, which combines the outputs of the head trainer S4, the torso sensor S5, and the torso prediction model S7, and uses neural radiation field technology to generate realistic 3D video output.

[0025] Further, the key frames are obtained based on the video stream, specifically: the video stream is converted into multiple video image frames, all the video image frames are sorted in chronological order, two adjacent video image frames after sorting are packaged into a comparison image frame group, the action state variable value of the comparison image frame group is obtained, and the action state high variable value and the action state low variable value are set;

[0026] When the action state change value is greater than or equal to the action state high change value, the state of the comparison image frame group is marked as a high fluctuation state;

[0027] When the action state change value is less than or equal to the action state low change value, the state of the comparison image frame group is marked as a stable state;

[0028] When the action state change value is between the action state high change value and the action state low change value, the state of the comparison image frame group is marked as a low fluctuation state;

[0029] Get the key frame extraction value of the video stream and set the key frame extraction threshold. When the key frame extraction value is greater than or equal to the key frame extraction threshold, mark all video image frames as key frames. When the key frame extraction value is less than the key frame extraction threshold, extract a set proportion of video image frames from the video image frames and mark them as key frames.

[0030] Furthermore, the key frame extraction value of the video stream is obtained by the following method: the total number of contrast image frame groups in high fluctuation state is obtained and marked as TWS, all contrast image frame groups are sorted in chronological order, when the states of two adjacent contrast image frame groups after sorting are different, no corresponding processing is performed, when the states of two adjacent contrast image frame groups after sorting are the same, the number of continuous stability is increased once, the number of continuous stability is summed up, the total number of continuous stability is obtained and marked as RKM, and the formula is used The key frame extraction value GPY of the video stream is obtained, wherein a1 is the total number coefficient of high fluctuation state, and a2 is the total number coefficient of continuous stability.

[0031] Furthermore, the action state variation value of the comparison image frame group is obtained by the following method: the comparison image frame group is used as input data of the action state analysis model, and the action state variation value is obtained as output.

[0032] Furthermore, the motion state analysis model is obtained in the following manner: obtain multiple comparison image frame groups, use the comparison image frame groups as training data for the neural network model, assign motion state variable values ​​to the training data, the value range of the motion state variable values ​​is [0~10), divide the training data into a training set and a verification set according to a set ratio, perform neural network iterative training on the training set and the verification set, and obtain the motion state analysis model after the training is completed. The larger the value of the motion state variable value, the greater the change in the head state and torso state of the character subject in the two video image frames, and the smaller the value of the motion state variable value, the smaller the change in the head state and torso state of the character subject in the two video image frames.

[0033] Furthermore, the speech-to-lip movement model S3 includes a generator network that accepts audio features as conditional input and outputs corresponding lip movement parameters. At the same time, a discriminator network is used to evaluate the consistency of the generated lip movements with the real data. Through the adversarial process, the generator learns to generate realistic lip movements to achieve a high degree of synchronization with the input audio.

[0034] Furthermore, through supervised learning methods, the head trainer can simulate dynamic changes of the head, including head rotation, tilt and complex expression changes, and the head trainer S4 outputs three-dimensional posture information of the head.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The method of the present invention achieves high synchronization between audio and video through precise audio processing and accurate provision of head and torso features. By using advanced NeRF technology, the synthesized character movements are realistic, including lip movements, facial expressions and whole-body movements. Through steps such as timbre extraction, conversion, embedding and synthesis, the present invention can replicate the timbre characteristics of a specific speaker and provide personalized voice output. The method of the present invention has real-time rendering capabilities and is suitable for application scenarios that require rapid response, such as live broadcasts, video conferencing, etc. It also supports the generation of high-resolution video streams and provides clear image details.

[0037] 2. Through the key frame extraction method of the present invention, the video impact frames in the video stream can be quickly analyzed, and key frames with training value can be quickly selected according to different behavioral actions, thereby reducing the difficulty of key frame extraction and ensuring the reasonable extraction of key frames. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A flowchart of a high-fidelity real-time rendering video synthesis method;

[0039] Figure 2 A block diagram of a high-fidelity real-time rendering video synthesis system. DETAILED DESCRIPTION

[0040] Example 1

[0041] Reference Figure 1 , a high-fidelity real-time rendering video synthesis method, comprising the following steps:

[0042] D1: Video collector S1 uses a high-resolution camera, supports 4K resolution and high frame rate acquisition to ensure high-quality video materials. The collector has automatic focus and exposure adjustment functions, which can automatically adjust parameters under various lighting conditions to ensure the true restoration of image clarity and color.

[0043] D2: Video collector S1 unifies the format of the collected video materials and transmits them to video processor S2 after processing. Video collector S1 is also responsible for the pre-processing of video data, including frame rate conversion, resolution adjustment and color space correction, to ensure that the video data is suitable for subsequent video processing and analysis tasks.

[0044] D3: Video processor S2 uses FFmpeg to separate the audio from the video material and performs noise reduction and enhancement processing to improve the clarity and intelligibility of the voice. The processed audio signal is further transmitted to the head trainer S4.

[0045] D4: Video processor S2 simultaneously performs frame analysis on the video stream and extracts key frame images. These key frames include the main actions and expression changes of the characters, providing visual information for subsequent torso movement analysis and prediction.

[0046] D5: The head trainer S4 combines the speech-to-lip movement model S3 and the processed audio data, analyzes the relationship between lip movement and speech through a machine learning algorithm, and trains a model that can simulate the head features of a person. The model can generate corresponding lip movements and facial expressions based on the speech signal.

[0047] D6: The torso sensor S5 combines the head features trained by S4 and the key frame images provided by S2, uses computer vision technology to extract the key points of the character's torso, and trains a model that can perceive and simulate the character's torso movements.

[0048] D7: The audio input module S7 is responsible for collecting the audio signal to be synthesized by the user. This module uses a high-sensitivity microphone and high-quality audio acquisition hardware to ensure that clear and distortion-free audio data is captured. During the acquisition process, S7 can also perform real-time audio signal monitoring to ensure that the audio quality meets the requirements of subsequent processing.

[0049] D8: The collected audio signal is first pre-processed by the audio input device S7, which includes removing background noise, adjusting volume level, performing equalization, etc. to optimize the audio quality. The pre-processed audio signal is converted into a format suitable for processing by the inference renderer S8, such as standard sampling rate and bit depth, and then transmitted to the inference renderer S8.

[0050] D9: The head trainer S4 transmits the head feature information obtained through its training, including facial expressions, lip movements, and head postures, to the inference renderer S8. These feature information are obtained by analyzing and learning the head movements of the characters in the basic video, and can accurately reproduce the head movements of the characters during the rendering process.

[0051] D10: The torso sensor S5 transmits the torso perception feature information it has trained, such as the movements of the shoulders, waist and limbs, to the inference renderer S8. This information is obtained by analyzing the body language and movement patterns in the key frame images of the character video, and is used to generate realistic full-body movements during the rendering process.

[0052] D11: The inference renderer S8 combines the audio signal provided by S7, the head features provided by S4, and the torso perception features provided by S5, and uses advanced neural radiance field (NeRF) technology for real-time rendering. The inference renderer S8 synchronizes the audio signal with the visual features, generates dynamic lip movements and facial expressions, and renders natural and smooth torso and limb movements. Finally, the inference renderer S8 outputs a high-fidelity video stream that not only has high image quality, but also has natural and realistic character movements, and the lip movements are highly synchronized with the audio input.

[0053] Example 2

[0054] Reference Figure 2 , a high-fidelity real-time rendering video synthesis system, comprising: a video collector S1: responsible for collecting target video samples in real time with high definition. The video collector S1 supports at least 4K resolution and 60 frames per second (fps) acquisition capability, ensuring the original quality of video data. In addition, the video collector S1 has automatic exposure and autofocus functions to adapt to different shooting environments and conditions.

[0055] Video processor S2: Receives video data collected by video collector S1 and decomposes it into audio stream and video stream. The audio stream is processed by a noise reduction algorithm to remove background noise and enhance the voice signal. The video stream is segmented into key frames, and the subject of the person is extracted using deep learning-based image segmentation technology, and the facial and body areas of the person are identified.

[0056] The processing method for dividing the video stream into key frames is as follows: convert the video stream into multiple video image frames, sort all the video image frames in chronological order, pack the two adjacent video image frames after sorting into a comparison image frame group, obtain the action state change value of the comparison image frame group, set the action state high change value and the action state low change value, wherein the action state high change value is greater than the action state low change value, and the action state high change value and the action state low change value are both system preset thresholds, which can be modified according to actual needs.

[0057] When the action state change value is greater than or equal to the action state high change value, the state of the comparison image frame group is marked as a high fluctuation state.

[0058] When the action state change value is less than or equal to the action state low change value, the state of the comparison image frame group is marked as a stable state;

[0059] When the action state change value is between the action state high change value and the action state low change value, the state of the comparison image frame group is marked as a low fluctuation state.

[0060] Get the total number of contrast image frame groups in high fluctuation state and mark it as TWS, sort all contrast image frame groups in chronological order, when the states of two adjacent contrast image frame groups after sorting are different, no corresponding processing is performed, when the states of two adjacent contrast image frame groups after sorting are the same, increase the number of continuous stability by one, sum the number of continuous stability, get the total number of continuous stability and mark it as RKM, use the formula The key frame extraction value GPY of the video stream is obtained, where a1 is the total number coefficient of high fluctuation state, a2 is the total number coefficient of continuous stability, the value of a1 is 0.86, and the value of a2 is 0.71;

[0061] Set the key frame extraction threshold. When the key frame extraction value is greater than or equal to the key frame extraction threshold, all video image frames are marked as key frames. When the key frame extraction value is less than the key frame extraction threshold, a set proportion of video image frames are extracted from the video image frames and marked as key frames.

[0062] Through the key frame extraction method of the present invention, the video impact frames in the video stream can be quickly analyzed, and key frames with training value can be quickly selected according to different behavioral actions, thereby reducing the difficulty of key frame extraction and ensuring the reasonable extraction of key frames.

[0063] The action state variation value of the comparison image frame group is obtained by the following method: the comparison image frame group is used as input data of the action state analysis model, and the action state variation value is obtained as output;

[0064] The motion state analysis model is obtained by the following method: obtaining multiple contrast image frame groups, using the contrast image frame groups as training data of the neural network model, assigning motion state variable values ​​to the training data, the value range of the motion state variable values ​​is [0-10), dividing the training data into a training set and a validation set according to a set ratio, performing neural network iterative training on the training set and the validation set, and obtaining the motion state analysis model after the training is completed. The larger the value of the motion state variable value, the greater the change of the head state (such as facial expression, head rotation angle, etc.) and the torso state (such as the position of the shoulder) of the subject in the two video image frames, and the smaller the value of the motion state variable value, the smaller the change of the head state and torso state of the subject in the two video image frames.

[0065] Speech-to-lip action model S3: Generative adversarial networks (GANs) are used to convert audio signals into lip action sequences. The model consists of a generator network that accepts audio features as conditional input and outputs corresponding lip action parameters. Meanwhile, a discriminator network is used to evaluate the consistency of the generated lip actions with the real data. Through the adversarial process, the generator learns to generate realistic lip actions to achieve a high degree of synchronization with the input audio.

[0066] Head Trainer S4: Based on the 3D Morphable Model (3DMM), Head Trainer S4 extracts the head posture and expression features from the keyframes. Through supervised learning methods, Head Trainer can simulate the dynamic changes of the head, including head rotation, tilt, and complex expression changes. Head Trainer S4 outputs the 3D posture information of the head to provide accurate head dynamics for the subsequent rendering process.

[0067] Torso Sensor S5: By analyzing the torso region of the person in the video frame, it extracts the key points of the torso, such as the movement information of the shoulders, waist, and hips. This sensor uses computer vision techniques, such as skeleton key point detection, to capture and understand the dynamic changes of the torso.

[0068] Trunk prediction model S6: Using a long short-term memory network (LSTM), the trunk prediction model S6 predicts the movements of the trunk in future frames based on the output of the trunk sensor S5. The recursive nature of the LSTM enables it to memorize long-term temporal dependencies, thereby accurately predicting complex trunk motion sequences.

[0069] Audio input device S7: As the audio input interface of the system, the S7 audio input device is responsible for receiving the user's audio input. It converts analog signals into digital signals and provides an audio format suitable for processing by the speech-to-lip action model S3 through preprocessing steps such as sampling, quantization and encoding.

[0070] Inference Renderer S8: It is the core rendering module of the system. It combines the outputs of the head trainer S4, the torso sensor S5, and the torso prediction model S7, and uses the neural radiance field (NeRF) technology to generate realistic 3D video output. Inference Renderer S8 not only takes into account lip movements, head postures, and torso movements, but also simulates complex visual effects such as lighting, shadows, and reflections to produce high-fidelity rendering results.

[0071] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0072] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0073] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0074] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0075] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0076] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0077] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks or optical disks.

[0078] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A high-fidelity real-time rendering video synthesis method, characterized in that: The steps include: D1: Collect video material; D2: Preprocess the collected video materials; D3: Separate the audio from the video material and perform noise reduction and enhancement processing; D4: Extract key frames from the video stream of the video material; D5: Analyze the relationship between lip movements and speech through machine learning algorithms, and train a model to simulate the characteristics of a person's head; D6: Obtain the trained head features and key frames, use computer vision technology to extract the key points of the character's torso, and train a model that perceives and simulates the character's torso movements; D7: collect the audio signal to be synthesized; D8: pre-process the audio signal to be synthesized; D9: Obtain the head feature information obtained through training and transmit it to the inference renderer; D10: Obtain the trained torso perception feature information and transmit it to the inference renderer; D11: Real-time rendering using advanced neural radiation field technology based on the audio signal to be synthesized, head feature information, and torso perception feature information; Extract key frames from the video stream of the video material, specifically: convert the video stream into multiple video image frames, sort all the video image frames in chronological order, pack two adjacent video image frames after sorting into a comparison image frame group, obtain the action state variable value of the comparison image frame group, and set the action state high variable value and the action state low variable value; When the action state change value is greater than or equal to the action state high change value, the state of the comparison image frame group is marked as a high fluctuation state; When the action state change value is less than or equal to the action state low change value, the state of the comparison image frame group is marked as a stable state; When the action state change value is between the action state high change value and the action state low change value, the state of the comparison image frame group is marked as a low fluctuation state; Get the key frame extraction value of the video stream, set the key frame extraction threshold, when the key frame extraction value is greater than or equal to the key frame extraction threshold, mark all video image frames as key frames, when the key frame extraction value is less than the key frame extraction threshold, extract a set proportion of video image frames from the video image frames and mark them as key frames; The key frame extraction value of the video stream is obtained by the following method: the total number of contrast image frame groups in high fluctuation state is obtained and marked as TWS, all contrast image frame groups are sorted in chronological order, when the states of two adjacent contrast image frame groups after sorting are different, no corresponding processing is performed, when the states of two adjacent contrast image frame groups after sorting are the same, the number of continuous stability is increased once, the number of continuous stability is summed up, the total number of continuous stability is obtained and marked as RKM, and the formula is used. The key frame extraction value GPY of the video stream is obtained, wherein a1 is the total number coefficient of high fluctuation state, and a2 is the total number coefficient of continuous stability; The action state variation value of the comparison image frame group is obtained by the following method: the comparison image frame group is used as input data of the action state analysis model, and the action state variation value is obtained as output; The motion state analysis model is obtained in the following way: obtain multiple contrast image frame groups, use the contrast image frame groups as training data of the neural network model, assign motion state variable values ​​to the training data, the value range of the motion state variable values ​​is [0~10), divide the training data into a training set and a validation set according to a set ratio, perform neural network iterative training on the training set and the validation set, and obtain the motion state analysis model after the training is completed. The larger the value of the motion state variable value, the greater the change in the head state and torso state of the character subject in the two video image frames, and the smaller the value of the motion state variable value, the smaller the change in the head state and torso state of the character subject in the two video image frames.

2. A high-fidelity real-time rendering video synthesis system, applied to a high-fidelity real-time rendering video synthesis method according to claim 1, characterized in that: Including video collector: collect target video samples in real time; Video processor: receives video data collected by the video collector, decomposes it into audio stream and video stream, and obtains key frames based on the video stream; Speech-to-lip action model: Generative adversarial network is used to convert audio signals into lip action sequences; Head Trainer: Based on the 3D deformable model, the head trainer extracts the head posture and expression features from the key frames; Torso sensor: extracts the key points of the torso by analyzing the torso area of ​​the character in the keyframe; Trunk prediction model: Using a long short-term memory network, the trunk prediction model predicts the movement of the trunk in future frames based on the output of the trunk sensor; Audio input device: As an audio input interface, the audio input device is responsible for receiving the user's audio input; Inference Renderer: It is the core rendering module, which combines the outputs of the head trainer, torso sensor, and torso prediction model, and uses neural radiance field technology to generate realistic 3D video output.

3. The high-fidelity real-time rendering video synthesis system according to claim 2, characterized in that: The speech-to-lip movement model consists of a generator network that accepts audio features as conditional input and outputs corresponding lip movement parameters. At the same time, a discriminator network is used to evaluate the consistency of the generated lip movements with the real data. Through the adversarial process, the generator learns to generate realistic lip movements to achieve a high degree of synchronization with the input audio.

4. The high-fidelity real-time rendering video synthesis system according to claim 2, characterized in that: Through supervised learning methods, the head trainer can simulate dynamic changes of the head, including head rotation, tilt and complex expression changes, and output three-dimensional posture information of the head.

Images