Communication method and related apparatus

Abstract

A communication method and a related apparatus, used for updating a model for processing image or video data, thereby reducing generative communication loss between a sending end and a receiving end. In the method, upon determining model update information, a communication apparatus updates a target model at an appropriate timing on the basis of a model update timing indicated by the model update information, so that upon successful configuration, the target model can also be updated at an appropriate timing, thereby avoiding model performance degradation caused by application scenario switching or updates in image clarity requirements, ensuring that image data or video data can be efficiently transmitted between communication apparatuses, and thus improving user experience during image browsing or video viewing.

Description

A communication method and related apparatus

[0001] This application claims priority to Chinese Patent Application No. 202411573689.7, filed on November 5, 2024, entitled "A Communication Method and Related Device", the entire contents of which are incorporated herein by reference. Technical Field

[0002] This application relates to the field of communications, specifically to a communication method and related apparatus. Background Technology

[0003] With the continuous upgrading of streaming media, higher resolution and faster transmission rates have become the main development directions for streaming media data transmission. Generative artificial intelligence (Generation AI) provides the technological foundation for more efficient data transmission.

[0004] By leveraging the feature extraction and feature recovery capabilities of artificial intelligence, at the sending end of audio and video data, a model is used to extract or compress features from the audio and video data, and then the feature extraction or compression result is sent to the receiving end. The receiving end then recovers the audio and video data based on the feature extraction or compression result.

[0005] However, once the model is successfully configured in this scenario, it will no longer be updated. The image or video data transmitted between the sending and receiving ends may have a large range of frames. Therefore, when the frame type is switched, there will be a performance drop due to the model's inability to adapt to the frame type switch, resulting in high generative communication loss during the transmission of images or videos. Summary of the Invention

[0006] This application provides a communication method and related apparatus for updating a model that processes image or video data, thereby reducing generative communication loss between the sender and receiver.

[0007] The first aspect of this application provides a communication method executed by a communication device, wherein the communication device can be a transmitting end (e.g., a user equipment), a receiving end (e.g., a base station or a server), an entity in the transmitting end (e.g., a processor, chip, or chip system), an entity in the receiving end (e.g., a processor, chip, or chip system), or the method can also be implemented by a logic module or software capable of implementing all or part of the functions of the transmitting end or the receiving end. In the first aspect and its possible implementations, the method is described as being executed by either the transmitting end or the receiving end. In this method, the transmitting end or the receiving end determines model update information and updates the target model at an appropriate time based on the model update information, wherein the model update information indicates the model update timing.

[0008] Based on the above scheme, after the communication device determines the model update information from the other end, it updates the target model at an appropriate time according to the model update timing indicated by the model update information. This allows the target model to be updated at appropriate times after successful configuration, avoiding performance degradation caused by application scenario switching or image clarity requirements. This ensures efficient transmission of image or video data between communication devices, thereby improving the user experience when browsing images or watching videos.

[0009] In one implementation example, the model update information may include a target duration (period). After determining the target duration, the communication device updates the target model at a first time point. The time interval between the first and second time points is the target duration, and the second time point is the update time of the target model, preceding the first time point. That is, the communication device updates the target model by determining the time interval between two model update times. It should be understood that the update time of the target model is the time when the model is trained to obtain the target model.

[0010] In one possible implementation of the first aspect, when updating the target model based on the target duration, the target dataset used for updating the target model can be pre-agreed upon. The target dataset is determined by identifying target dataset information, which includes at least one of the following: the target dataset itself, or the identifier of the target dataset. Furthermore, based on the target duration, the target model is updated using the target dataset information at the first moment.

[0011] Based on the above scheme, while sending the target duration to the peer device, the communication device can also send the target dataset information to the peer device, so that the communication device and the peer device can update the model simultaneously based on the same or related datasets. This effectively synchronizes the model update direction of the communication device and the peer device, enabling the communication device and the peer device to have better encoding and decoding effects, or feature extraction and feature restoration effects.

[0012] Furthermore, when the receiving end directly sends the target dataset to the sending end, the sending end only needs to store the relevant information of the target dataset, eliminating the need to store unnecessary datasets. This reduces the storage space occupied by the sending end, effectively freeing up storage space in scenarios where the sending end has limited storage capacity. Conversely, when the receiving end pre-sends candidate datasets to the sending end, during configuration, it only needs to send the identity identifier of the target dataset, effectively reducing the amount of data transmitted between nodes, saving network resources while improving transmission speed.

[0013] In one possible implementation of the first aspect, when the target model is a data recovery model, the communication device is the receiving end. The data recovery model is used to perform feature restoration or decompression of the data to be recovered, obtaining video data or image data. That is, when the receiving end device synchronizes the target duration with the sending end device, the receiving end device not only needs to update the data recovery model based on the target duration, but also needs to send the target duration and target dataset information to the sending end device so that the sending end device can synchronize with the receiving end device in updating the source compression model. Here, the target duration can be generated autonomously by the receiving end device or configured by technicians.

[0014] Based on the above scheme, when the update of the target model is triggered by the receiving device, the target duration and target dataset also need to be sent to the sending device to ensure that the sending device and the receiving device update the target model synchronously. This effectively synchronizes the model update timing of the communication device and the peer device, enabling the communication device and the peer device to have better encoding and decoding effects, or feature extraction and feature restoration effects.

[0015] In one possible implementation of the first aspect, when the target model is a source compression model, the communication device is the transmitter. The source compression model is used to extract or compress features from video or image data to obtain the data to be recovered. That is, when the transmitter synchronizes the target duration with the receiver device, the transmitter device first sends at least one reference duration to the base station based on its configuration information or actual operating conditions. After receiving the reference duration, the base station selects one duration from the reference durations as the target duration and sends the target duration to the transmitter device.

[0016] Based on the above scheme, the sending end triggers the determination of the target duration, enabling the base station to select the target duration from the reference duration as the update cycle of the source compression model. Since the sending end generates the reference duration based on the configuration and usage of the sending end equipment, the target duration determined in this way is closer to the actual usage requirements of the sending end equipment, which is beneficial to improving the compression efficiency and feature extraction efficiency of video or image data by the sending end.

[0017] In another implementation example, the model update information includes a model update signal. Upon determining the model update signal, the communication device immediately updates the target model. That is, the communication device determines the model update signal based on the specific circumstances and updates the target model after determining the model update signal.

[0018] In this implementation example, the update of the target model is triggered by the model update signal, making the update of the target model more flexible. The sending and receiving ends can more flexibly determine the model update signal according to the specific situation of data transmission, and trigger the update of the target model in a timely manner according to the data processing situation of the target model. This allows the solution to adjust the generation conditions of the model update signal according to different needs, thereby achieving flexible updates of the target model.

[0019] In one possible implementation of the first aspect, the model update signal is a signal generated when the distance between the first feature vector and the second feature vector is greater than a first threshold, wherein the first feature vector and the second feature vector are the results of feature extraction of image data or video data by the source compression model in adjacent time periods.

[0020] Based on the above scheme, the sending end analyzes whether the video content or image content sent by the video end has changed according to the feature vector generated by the source compression model. When a change occurs, the sending end actively generates a model update signal, so that the source compression model can be adaptively adjusted in a timely manner according to the output content of the sending end, ensuring that the source compression model is updated at the appropriate time, which helps to improve the compression performance of the source compression model.

[0021] In one possible implementation of the first aspect, the model update signal is a signal generated when the first damage degree or the second damage degree is greater than the second threshold, wherein the first damage degree is the damage degree of the image data generated by the data recovery model based on the data to be recovered, and the second damage degree is the damage degree of the video data generated by the data recovery model based on the data to be recovered.

[0022] Based on the above scheme, the receiving end analyzes the data to be recovered sent by the sending end to determine the degree of damage to the data. When the degree of damage to the data to be recovered exceeds a second threshold, a model update signal is generated. The data recovery model is then updated after generating this model update signal. By generating the model update signal based on the data to be recovered, the receiving end can promptly trigger updates to the data recovery model and the source compression model based on anomalies in the data to be recovered during communication. This also effectively eliminates the reduction in transmission efficiency caused by communication anomalies and data source anomalies.

[0023] In one possible implementation of the first aspect, the model update signal is a signal generated when the third damage degree is greater than the third threshold. The third damage degree is determined based on the similarity between the first image and the second image. The first image is an image obtained by recovering the first string using a data recovery model. The second image is an image that corresponds to the first string and is similar to the image corresponding to the data to be recovered.

[0024] Based on the above scheme, the receiving end uses the data recovery model to process the first string to obtain the first image, and determines the third damage degree based on the similarity between the first image and the second image. When the third damage degree is greater than the third threshold, an update model signal is generated to detect transmission anomalies and anomalies in the data recovery model. This helps to trigger the update of the data recovery model in a timely manner when anomalies occur in the data transmission path or when the data recovery model cannot meet the recovery requirements. This effectively avoids the situation where data transmission continues even when anomalies occur at the receiving end or in the data transmission path, thus improving the reliability of data transmission.

[0025] In one possible implementation of the first aspect, when the target model is a source compression model and the communication device is a transmitting device, the transmitting end can simultaneously send a first string, or a second string and a second image to the receiving end while sending the data to be recovered to the receiving end, wherein the transmitting end device receives a model update signal from the receiving end.

[0026] Based on the above scheme, with the correspondence between reference strings and reference images pre-configured at the receiving end and the sending end, the first string is transmitted from the sending end to the receiving end. The receiving end uses the data recovery model to process the first string to obtain the first image, and performs analysis based on the first image. This effectively reduces the amount of data that needs to be transmitted between communication devices and effectively saves network resources.

[0027] Alternatively, the sending end can directly send the first string and the second image to the receiving end without configuring the correspondence between the string and the image on both ends. This allows for immediate use and effectively eliminates the occupation of the receiving end's storage resources by the correspondence between the string and the image. In cases where the receiving end's storage resources are scarce, this can alleviate the pressure on the receiving end's storage resources.

[0028] In one possible implementation of the first aspect, when the communication device is the receiving end, the receiving end device generates a model update signal and sends the model update signal to the sending end device.

[0029] Based on the above scheme, after generating the model update signal, the receiving end sends the model update signal to the sending end. That is, after generating the model update signal, the receiving end sends the model update signal to the other end to ensure that the models of the sending end and the receiving end are updated at the same time, avoiding the situation where only one side updates the model.

[0030] In one possible implementation of the first aspect, before determining the model update information, it is also necessary to first determine the model to be trained and the preset dataset, and then use the preset dataset to train the model to be trained to obtain the target model.

[0031] Based on the above scheme, when training the source compression model, the transmitting end uses the model to be trained or the preset dataset sent by the receiving end to train the model and obtain the source compression model. This ensures that the source compression model trained by the transmitting end maintains a certain degree of consistency with the data recovery model of the receiving end, while also avoiding overfitting of the source compression model.

[0032] A second aspect of this application provides a communication device, comprising:

[0033] The processing unit is used to determine model update information, which indicates when to update the model.

[0034] The model training unit is used to update the target model based on the model update information. The target model includes a source compression model or a data recovery model. The source compression model is used to extract or compress features from video data or image data to obtain the data to be recovered. The data recovery model is used to restore features from the data to be recovered or decompress it to obtain video data or image data.

[0035] The beneficial effects shown in this aspect are similar to those of the first aspect or any possible implementation of the first aspect, and will not be repeated here.

[0036] A third aspect of this application provides a communication device, and a third aspect of this application provides a computer device including a processor and a memory, wherein the processor stores instructions, and when the instructions stored in the memory are executed on the processor, the method shown in the first aspect or any possible implementation of the first aspect is implemented.

[0037] The fourth aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a processor, implement the method shown in the first aspect or any possible implementation thereof.

[0038] The fifth aspect of this application provides a computer program product that, when executed on a processor, implements the method shown in the first aspect or any possible implementation of the first aspect.

[0039] The beneficial effects shown in any of the third to fifth aspects are similar to those of the first aspect or any possible implementation of the first aspect, and will not be repeated here. Attached Figure Description

[0040] Figure 1 is a schematic diagram of an architecture of the communication system provided in this application;

[0041] Figure 2 is a flowchart illustrating a communication method provided in this application;

[0042] Figure 3 is a schematic diagram of another communication method provided in this application;

[0043] Figure 4 is a schematic diagram of another communication method provided in this application;

[0044] Figure 5 is a schematic diagram of another communication method provided in this application;

[0045] Figure 6 is a schematic diagram of another communication method provided in this application;

[0046] Figure 7 is a schematic diagram of a communication device provided in this application;

[0047] Figure 8 is a schematic diagram of another structure of the communication device provided in this application. Detailed Implementation

[0048] This application provides a communication method and related apparatus for timely updating the model for processing image or video data, thereby improving the transmission efficiency of image or video data between devices.

[0049] First, some of the terms used in the application will be explained to help those skilled in the art better understand the solution provided in this application.

[0050] (1) Generative Adversarial Network (GAN): A GAN consists of a generator and a discriminator, which learn the distribution of the training data based on the generator and discriminator. The generator is a multi-layer perceptron network. The discriminator determines whether the input data comes from the generator or the training data; its output is the probability of the training data. Finally, the discriminator is trained to accurately determine the data source, and the generator is trained to make its generated data conform as closely as possible to the distribution of the training data. During training, the optimization of the generator and discriminator is performed alternately.

[0051] (2) Variational auto-encoders (VAE): VAE is an unsupervised neural network method consisting of an encoder and a decoder. VAE is used to discover models that represent some hidden states (incomplete, sparse, denoised, and contracted) of data.

[0052] Specifically, the input data is transformed into an encoded vector, where each dimension represents an attribute learned from the data. The encoder outputs a single value for each encoded dimension, and the decoder then receives these values ​​and attempts to recreate the original input.

[0053] VAE provides a probabilistic way to describe latent space observations. Therefore, instead of building an encoder that outputs a single value to describe each latent state attribute, the encoder describes the probability distribution of each latent attribute.

[0054] (3) Diffusion Model: The diffusion model is a generative model based on probability theory, originally derived from the diffusion process theory in physics, such as the diffusion process of ink in water. In the field of machine learning, this concept has been creatively applied to data generation tasks, especially image and sound synthesis. They reconstruct high-quality data samples from noise by simulating a gradual "diffusion" process from a data distribution to a simple noise distribution, and then learning the inverse process.

[0055] The diffusion model first defines a process of gradually transforming the data distribution into a Gaussian noise distribution (forward diffusion), which can be viewed as a series of steps that gradually add noise. Then, the model learns how to perform the inverse operation of this process, starting with pure noise and gradually "de-noising" through a series of inverse steps to finally generate samples that approximate the original data distribution (backward diffusion). This inverse process typically involves complex probability distribution estimation and must ensure that the generated samples have high fidelity and diversity.

[0056] The advantage of diffusion models is that they can theoretically approximate arbitrarily complex data distributions, and the generated samples often have higher quality and consistency.

[0057] (4) Autoregressive models: Data is generated by considering the sequence dependencies of elements in the data, such as PixelRNN and PixelCNN.

[0058] To facilitate a better understanding of the solution provided in this application, the application scenarios of the solution provided in this application are introduced below.

[0059] The solution provided in this application can be applied to communication systems, such as 5th generation mobile communication technology (5G) communication systems, new radio (NR) communication systems, long term evolution (LTE) communication systems, LTE frequency division duplex (FDD) communication systems, LTE time division duplex (TDD) communication systems, etc.

[0060] In one possible implementation, the solution provided in this application is applicable to the communication system shown in Figure 1.

[0061] The communication system includes a wireless access network. The wireless access network may include at least one wireless access network device and at least one terminal. The terminal connects wirelessly to the wireless access network device.

[0062] Optionally, it may also include a core network and the Internet. The wireless access network (WLAN) device connects to the core network wirelessly or via a wired connection. The core network device and the WLAN device can be independent physical devices, or the functions of the core network device and the logical functions of the WLAN device can be integrated into a single physical device. Alternatively, a single physical device can integrate some of the functions of the core network device and some of the functions of the WLAN device. Terminals and WLAN devices can be interconnected via wired or wireless connections.

[0063] Network equipment can be any device with wireless transceiver capabilities. This includes, but is not limited to: traditional macro base stations (eNBs) (evolved node Bs) in traditional Universal Mobile Telecommunications Systems (UMTS) / LTE wireless communication systems; micro base stations (eNBs) in heterogeneous network (HetNet) scenarios; baseband units (BBUs) and remote radio units (RRUs) in distributed base station scenarios; baseband pools (BBU pools) and RRUs in cloud radio access networks (CRAN) scenarios; gNBs in future wireless communication systems; base stations evolved from 3GPP systems; access nodes, wireless relay nodes, and wireless backhaul nodes in WiFi systems. Base stations can be: macro base stations, micro base stations, pico base stations, small cells, relay stations, or balloon stations, etc. Network equipment can also be servers, wearable devices, or vehicle-mounted devices, etc.

[0064] User equipment can be vehicle communication modules or other embedded communication modules, mobile phones, tablets, computers with wireless transceiver capabilities, virtual reality (VR) terminal devices, augmented reality (AR) terminal devices, wireless terminals in industrial control, tactile terminal devices, vehicle terminal devices, wireless terminals in autonomous driving, wireless terminals in remote medical care, wireless terminals in smart grids, wireless terminals in transportation safety, wireless terminals in smart cities, wireless terminals in smart homes, wearable terminal devices, and so on.

[0065] The network elements involved in this application are data network (DN), core network, access network (AN), and user equipment (UE).

[0066] Data Network (DN): Provides services such as carrier services, Internet access, or third-party services, including servers, which implement video source encoding, rendering, etc.

[0067] Network Open Function (NEF) element: Exposes the services and capabilities of 3GPP network functions to the AF, and also allows the AF to provide information to 3GPP network functions. The corresponding interface is the N33 interface.

[0068] Policy Control Function (PCF) network element: performs policy management for charging policies and QoS policies;

[0069] Application function (AF) network elements mainly convey the application side's requirements to the network side;

[0070] Access and Mobility Management (AMF) network elements: primarily perform mobility management, access authentication / authorization, and other functions. Additionally, they are responsible for transmitting user policies between the UE and the PCF; the N1 interface is the signaling plane interface between the UE and the AMF. Since the UE cannot directly interact with the core network, it needs to pass NAS (Non-Access Stratum) information through the AN; the N2 interface is the signaling plane interface through which the AMF requests resources from the AN to allocate for PDU sessions, etc.

[0071] Session Management Function (SMF) network element: performs session management functions such as UE IP address allocation, UPF selection, and billing and QoS policy control;

[0072] User Plane Function (UPF) network elements: Serving as the interface with the data network, they perform functions such as user plane data forwarding, session / flow-level billing statistics, and bandwidth limiting. The N3 interface is the interface between the RAN (Radio Access Network) and the UPF, primarily used for transmitting uplink and downlink user plane data between the 5G RAN and the UPF.

[0073] Access Network (AN): This can be any device with wireless transceiver capabilities. It includes, but is not limited to: evolved NodeBs or eNBs in LTE, base stations (gNodeBs or gNBs) or Transmission Reception Points (TRPs) in NR, base stations evolved from 3GPP, access nodes, wireless relay nodes, and wireless backhaul nodes in WiFi systems. Base stations can be: macro base stations, micro base stations, pico base stations, small cells, relay stations, or balloon stations, etc. Network equipment can also be servers, wearable devices, or vehicle-mounted equipment, etc.

[0074] Generative artificial intelligence (GAI) is a type of data that uses machine learning algorithms to generate new data samples that are similar to, but not exactly the same as, the training data.

[0075] With the continuous development of streaming media, higher resolution and faster transmission rates have become major development directions. Higher resolution inevitably comes with larger data transmission volumes, making it difficult to achieve both simultaneously with the same network quality.

[0076] To address this issue, high-definition video or image transmission based on super-resolution technology offers an innovative solution for the efficient transmission and processing of images or videos. Super-resolution technology utilizes generative artificial intelligence to perform machine learning at the data sending end, learning image features from large amounts of data. By transmitting these features, the data receiving end can then reconstruct the data based on these features, thus achieving image or video restoration.

[0077] In the transmission of high-definition video or images based on super-resolution technology, once the model is successfully configured, it is no longer updated. However, the image or video data transmitted between the data sender and receiver may show significant differences in detail. For example, the image transmitted at one moment might depict animals, while the image transmitted at the next moment might show a cityscape. If the data sender and receiver still use the model trained for images depicting animals, it will lead to significant communication impairments. The challenge lies in accurately identifying model performance degradation and updating the models at both the data sender and receiver to maintain efficient image or video transmission between them.

[0078] To address the above issues, this application proposes that the data receiving end or data sending end can update the target model based on the determined model update information. The target model is used for feature extraction, compression, feature restoration, or decompression of video or image data. The model update information can be preset or generated by the device according to specific circumstances. By using the model update information to determine the timing of the target model update, continuous updates can improve the target model's data encoding and decoding performance during data transmission.

[0079] Based on the above ideas, the solution provided in this application will be described below with reference to the accompanying drawings:

[0080] Please refer to Figure 2, which is a flowchart of a communication method provided in this application.

[0081] S210. Determine the model to be trained and the preset dataset;

[0082] The model to be trained is a generative model, and the preset dataset is images or videos.

[0083] In some possible implementations, the generative model can be a GAN model, a VAE model, an autoregressive model, or a transformer model, without limitation.

[0084] S220. Train the model to be trained using a preset dataset to obtain the target model;

[0085] The target model is a model used for feature extraction, compression, feature restoration, or decompression of video or image data.

[0086] S230, Determine model update information;

[0087] Among them, model update information is used to determine when to update the model.

[0088] Specifically, model update information can be a target duration, which enables the device to periodically update the model. For example, a target duration of 5 seconds means the target model will be updated every 5 seconds. Model update information can also be a model update signal, allowing the device to update the target model immediately upon receiving the signal. For example, the model update signal can be a string like "update modelS," where "modelS" represents the target model. It should be understood that this string is pre-agreed upon by the sender and receiver; the specific content is not limited here.

[0089] S240. Update the target model based on the model update information.

[0090] Specifically, based on model update information, the target model is updated at the appropriate model update time.

[0091] In one possible implementation, if the target duration is 5 seconds, the target model is updated 3 days after the target model training is completed; once the model update signal is determined, the target model is updated immediately.

[0092] In this embodiment, the communication device determines model update information and updates the target model at the time indicated by the model update information. This ensures the target model is updated at specific times, preventing performance degradation caused by models ceasing updates once successfully configured, and guaranteeing efficient transmission of image or video data between communication devices. In scenarios where the application environment is constantly changing and rapid updates of the target model are undesirable, a longer target duration can be set to reduce the model update frequency, thereby conserving the computing resources of the communication device.

[0093] In specific application scenarios, the solutions provided in this application can be implemented in different ways. The solutions provided in this application will be introduced below in conjunction with different implementing entities.

[0094] This application divides the implementing entities of the solution into two categories: the transmitting end and the receiving end. The transmitting end is the UE (User Equipment), and the receiving end is the base station or server. The transmitting end is a device that compresses or extracts features from image or video data, and the receiving end is a device that receives the data to be recovered. The data to be recovered is data that, after processing, can yield image or video data.

[0095] The solution provided in this application will be described below from the perspective that the model update is performed at the sending end. Based on the method shown in Figure 2 above, please refer to Figure 3, which is a schematic diagram of the communication method provided in this application.

[0096] S310. The sending end determines the model to be trained and the preset dataset;

[0097] In this context, at the sending end, at least one of the model to be trained and the preset dataset is sent from the receiving end to the sending end.

[0098] In one possible implementation, when the model to be trained is a model sent from the receiver to the sender, and the relevant information of the model to be trained is transmitted between the sender and the receiver, both the receiver and the sender can be configured with a model library, the form of which is shown in Table 1.

[0099] Table 1. Model Correspondence Table.

[0100] As shown in Table 1, the receiving and sending ends can pre-determine the correspondence between model identifiers and model types. When transmitting model types between the receiving and sending ends, this can be achieved by transmitting model identifiers. Transmitting model identifiers effectively transmits models, reducing the amount of data transmission required for transmitting the same model between the sending and receiving ends, saving network resources, and improving the efficiency of model transmission.

[0101] Similarly, the preset dataset can also be a dataset pre-stored at both the receiving and sending ends. Both ends also store an identifier for the preset dataset. When transmitting the preset dataset between the receiving and sending ends, this identifier is transmitted. This saves network resources used for transmitting the preset dataset and improves the efficiency of model transmission.

[0102] For example, the model to be trained is sent from the receiver to the sender, or the preset dataset is sent from the receiver to the sender.

[0103] It should be understood that the determination of certain information mentioned in this application can be implemented by receiving the information or by generating the information. In actual application, it should be set according to the specific scenario, and no restrictions are made here.

[0104] S320. The transmitting end uses a preset dataset to train the model to be trained and obtains the source compression model.

[0105] The source compression model is included in the target model. The source compression model is used to compress video data or image data or extract features.

[0106] In this embodiment, when training the source compression model, the transmitting end uses the model to be trained or a preset dataset sent by the receiving end to train the model and obtain the source compression model. This ensures that the source compression model trained by the transmitting end maintains a certain degree of consistency with the data recovery model of the receiving end, while also avoiding overfitting of the source compression model.

[0107] S330, The sending end receives the target duration sent by the receiving end, or the receiving end sends the target duration;

[0108] The target duration is included in the model update information, and the target duration is the model update cycle.

[0109] In one possible implementation, the receiving end is a server and the sending end is a UE.

[0110] In this scenario, after determining the target duration based on its configuration, the server sends the target duration to the base station via user plane signaling. For example, the server adds the target duration to the IP packet or GTP-U extension protocol header, and then transmits the target duration to the base station by sending the IP packet or GTP-U extension protocol message.

[0111] In this embodiment of the application, when the update of the target model is triggered by the receiving device, the target duration and target dataset also need to be sent to the sending device to ensure that the sending device and the receiving device update the target model synchronously. This effectively synchronizes the model update timing of the communication device and the peer device, enabling the communication device and the peer device to have better encoding and decoding effects, or feature extraction and feature restoration effects.

[0112] The base station then sends the target duration to the UE. For example, the base station can send the target duration to the UE through physical layer signaling, such as DCI; or it can send the target duration to the UE through higher layer signaling, such as MAC CE or RRC signaling. There are no restrictions on this.

[0113] In one possible implementation, the receiver is a base station and the transmitter is a UE.

[0114] In this scenario, after determining the target duration based on its configuration, the base station sends the target duration to the UE and the server via signaling messages. These signaling messages can be physical layer signaling; for example, the base station can send the target duration to the UE via physical layer signaling such as DCI. Alternatively, it can send the target duration to the UE via higher-layer signaling such as MAC CE or RRC signaling; there are no restrictions on this.

[0115] It should be understood that the sending end can be a base station or a server. The description here is only an example. In actual applications, the settings should be combined with the specific application scenario. There are no restrictions here.

[0116] S340, The sending end receives the target dataset information sent by the receiving end, or the receiving end sends the target dataset information;

[0117] The target dataset information includes at least one of the following: the target dataset, or the identity identifier of the target dataset.

[0118] Specifically, while sending the target duration to the sender, the receiving end can also send the target dataset information to the sender.

[0119] In one possible scenario, the receiver is the server and the sender is the UE.

[0120] For example, the server uses MAC CE or RRC to send candidate datasets to the UE and the base station; when sending candidate datasets, the server uses DCI to send the identity of the target dataset to the UE and the base station to instruct the UE or the base station to use the target dataset in the candidate dataset to update the source compression model, wherein the candidate dataset includes the target dataset.

[0121] Alternatively, the server may use MAC CE or RRC to send the target dataset to the UE and the base station to instruct the UE or the base station to use the target dataset to update the source compression model; no restrictions are imposed here.

[0122] In one possible scenario, the receiver is the base station and the transmitter is the UE.

[0123] For example, during the configuration phase, the base station uses MAC CE or RRC to send candidate datasets to the UE and server; during the application phase, it uses DCI to send the identity of the target dataset to the UE and server to instruct the UE or server to use the target dataset in the candidate dataset to update the source compression model.

[0124] Alternatively, the base station may use MAC CE or RRC to send the target dataset to the UE and server to instruct the UE or server to use the target dataset to update the source compression model; no restrictions are imposed here.

[0125] It should be understood that in specific application scenarios, the transmission of target dataset information and target duration can be achieved through the same signaling message or through different signaling messages; no restriction is imposed here.

[0126] In this embodiment, when the receiving end directly sends the target dataset to the sending end, the sending end only needs to store the relevant information of the target dataset, eliminating the need to store unnecessary datasets. This reduces the storage space occupied by the sending end, effectively freeing up storage space in scenarios where the sending end has limited storage space. Furthermore, when the receiving end pre-sends candidate datasets to the sending end, during configuration, it only needs to send the identity identifier of the target dataset to the sending end, effectively reducing the amount of data transmitted between nodes, saving network resources while improving transmission speed.

[0127] It should be understood that there is no explicit order between steps S330 and S340. In practical applications, the order should be determined based on the specific application scenario, and no restrictions are imposed here.

[0128] S350, The sending end updates the source compression model using the target dataset information at the first moment;

[0129] The time interval between the first and second moments is the target duration, and the second moment is the update time of the source compression model.

[0130] Specifically, the second moment is the moment when the source compression model is obtained. Taking the scheme shown in Figure 3 as an example, the second moment is the moment when step S320 is completed, and there is no restriction here.

[0131] For example, if the source compression model is obtained by training a preset model using a reference dataset, then the second time step is the end time of training the preset model using the reference dataset. That is, when the model obtained by updating the source compression model using the target dataset is called the updated source compression model, the time when the updated source compression model is updated again is taken as the second time step.

[0132] In this embodiment of the application, while the communication device sends the target duration to the peer device, it can also send the target dataset information to the peer device, so that the communication device and the peer device can update the model simultaneously based on the same or related datasets, effectively synchronizing the model update direction of the communication device and the peer device, and enabling the communication device and the peer device to have better encoding and decoding effects, or feature extraction and feature restoration effects.

[0133] S360: The receiving end updates the data recovery model using the target dataset information at the first moment.

[0134] The data recovery model is the model trained in the second time step, and there are no restrictions on its specific training method or the dataset used for training.

[0135] In this embodiment of the application, by setting a target duration, both the sending end and the receiving end can periodically update the target model, which effectively solves the problem that the target model cannot be updated again once it is determined, and effectively improves the transmission efficiency of image data or video data between devices.

[0136] Based on the method shown in Figure 2 above, Figure 3 above describes the situation where the receiver generates the target duration and sends the target duration to the sender. Next, we will introduce the situation where the sender generates the target duration in conjunction with Figure 4.

[0137] S410. The sending end determines the model to be trained and the preset dataset;

[0138] S420. The transmitting end uses a preset dataset to train the model to be trained and obtains the source compression model.

[0139] It should be understood that steps S410 and S420 are similar to the implementation of steps S310 and S320 shown in Figure 3 above. For details, please refer to steps S310 and S320 in Figure 3 above, which will not be repeated here.

[0140] S430, The transmitting end sends a reference duration to the base station, or the base station receives the reference duration;

[0141] The reference duration includes the target duration, which is included in the model update information and represents the model update cycle.

[0142] Specifically, the transmitter generates at least one reference duration according to its configuration and sends the reference duration to the base station.

[0143] S440, The base station selects the target duration from the reference duration;

[0144] Specifically, the base station selects any one of at least one reference duration as the target duration.

[0145] S450, The transmitting end receives the target duration sent by the base station, or the base station sends the target duration to the transmitting end;

[0146] Specifically, the target duration is sent to the sending end.

[0147] In one possible implementation, the base station sends the target duration to the server at the same time as sending the target duration, so that the server can periodically update the data recovery model deployed on the server according to the target duration. The data recovery model is used to recover compressed data or feature vectors into video data or image data.

[0148] S460, The sending end updates the source compression model in the first instant;

[0149] The time interval between the first and second moments is the target duration, and the second moment is the update time of the source compression model.

[0150] Specifically, the second moment is the moment when the source compression model is obtained. Taking the scheme shown in Figure 4 as an example, the second moment is the moment when step S420 is completed. It should be understood that obtaining the source compression model here means obtaining the source compression model by training the model to be trained.

[0151] In this embodiment, the sending end triggers the determination of the target duration, enabling the base station to select the target duration from the reference duration as the update cycle of the source compression model. Since the sending end generates the reference duration based on the configuration and usage of the sending end device, the target duration determined in this way is closer to the actual usage requirements of the sending end device, which is beneficial to improving the compression efficiency and feature extraction efficiency of the sending end for video or image data.

[0152] Optionally, step S470 may be included after step S440.

[0153] S470, the base station updates the data recovery model at the first moment.

[0154] The data recovery model is included in the target model and is used to recover compressed data or feature vectors into video data or image data.

[0155] In one possible implementation, step S470 is performed after the base station receives the signal sent by the transmitter indicating the target duration. Step S470 can also be performed after receiving the signal sent by the transmitter indicating that the source compression model update is complete. No limitation is made here.

[0156] In this embodiment, when the base station triggers the determination of a target duration at the transmitting end, the base station selects the target duration from the reference duration as the update period for the source compression model and the data recovery model. Furthermore, the base station also updates the data recovery model based on the target duration, enabling different devices to update their models simultaneously according to the update period. This effectively improves the efficiency of image and video data compression and decompression between different devices.

[0157] Based on the method shown in Figure 2 above, after the source compression model and data recovery model have been trained and the system is running normally, the model update information can be generated by either the sending end or the receiving end. The following will introduce the model update information generated by the sending end and the model update information generated by the receiving end respectively.

[0158] First, we will introduce the case where the model update information is generated by the sending end, as shown in Figure 5.

[0159] S510, The transmitting end generates a model update signal;

[0160] The model update signal is included in the model update information.

[0161] Specifically, in one possible implementation, the sending end can generate a model update signal when the video or image content transmitted by the sending end changes. For example, when the video content transmitted by the sending end changes from a face-related video to a street view-related video, the sending end generates a model update signal.

[0162] The transmitting end can analyze whether the video content has changed significantly based on the feature vectors output by the source compression model. For example, when the distance between the first feature vector and the second feature vector output by the source compression model is greater than a first threshold, the transmitting end generates a model update signal. The source compression model outputs the first feature vector after it outputs the second feature vector, and the second feature vector is the average of the feature vectors output by the source compression model in the first time period.

[0163] In this embodiment, the sending end analyzes whether the video content or image content sent by the video end has changed based on the feature vector generated by the source compression model. When a change occurs, the sending end actively generates a model update signal, so that the source compression model can be adaptively adjusted in a timely manner according to the output content of the sending end, ensuring that the source compression model is updated at the appropriate time, which helps to improve the compression performance of the source compression model.

[0164] S520, The sending end updates the source compression model;

[0165] In one possible implementation, the transmitter pre-stores multiple datasets, selects the dataset that is most similar to the image corresponding to the first feature vector from the multiple datasets, and uses the dataset that is most similar to the image corresponding to the first feature vector to update the source compression model.

[0166] S530, The transmitting end sends a model update signal to the receiving end;

[0167] When the receiving end is a base station, the sending end sends the model update signal to the base station through UCI signaling messages, MAC CE signaling messages, or UAI signaling messages.

[0168] When the receiving end is a server, the sending end first sends the model update signal to the base station via UCI signaling message, MAC CE signaling message or UAI signaling message, and then the base station forwards the model update signal to the server.

[0169] S540, receiver updates data recovery model.

[0170] It should be understood that the data set used by the receiving end to update the data recovery model is collected in a similar way to the data set obtained by the sending end to update the source compression model in step S520 above, and will not be described again here.

[0171] In this embodiment, the update of the target model is triggered by a model update signal, making the update of the target model more flexible. The sending end and the receiving end can more flexibly determine the model update signal according to the specific situation of data transmission, and trigger the update of the target model in a timely manner according to the processing of data by the target model. This allows the solution to adjust the generation conditions of the model update signal according to different needs, thereby achieving flexible updating of the target model.

[0172] It should be understood that there is no explicit order between steps S520 and S530 and S540. The description here is only an example. In actual application, step S520 can be implemented after step S510, and step S540 can be implemented after step S530. There are no restrictions here.

[0173] The following describes the case where the model update information is generated by the receiving end; please refer to Figure 6.

[0174] S610: The sending end sends the data to be recovered to the receiving end;

[0175] The data to be recovered is the data from the input data recovery model.

[0176] Optionally, in some possible scenarios, the sending end sends the data to be recovered and a first string to the receiving end, or the sending end sends the data to be recovered, the first string, and a second image to the receiving end. The first string corresponds to the second image, and the second image is similar to the image or video frame corresponding to the data to be recovered. The first string is a sequence of random codes used to uniquely identify an image; no restrictions are placed here.

[0177] S620, The receiver generates a model update signal;

[0178] The model update signal is included in the model update information.

[0179] The receiver generates model update signals based on two main approaches: Approach 1: After processing the data to be recovered using the data recovery model, the receiver analyzes the data to determine whether a model update is needed; Approach 2: While receiving the data to be recovered, the receiver also receives the first string. After processing the first string, the receiver analyzes the processing result to determine whether a model update is needed.

[0180] The two approaches will be introduced below.

[0181] Approach 1:

[0182] When the data transmitted at the sending end is image data, the model update signal is the signal generated at the receiving end when the first impairment degree is greater than the second threshold. The first impairment degree is the impairment degree of the image data generated by the data recovery model based on the data to be recovered. The impairment degree can be a value obtained by weighting calculation based on semantic similarity, image clarity, and whether there is mosaic in the image. The data to be recovered is the data sent from the sending end to the receiving end.

[0183] When the data transmitted at the sending end is video data, the model update signal is the signal generated at the receiving end when the second impairment degree is greater than the second threshold. The second impairment degree is the impairment degree of the video data generated by the data recovery model based on the data to be recovered. The impairment degree can be a value obtained by weighting calculations based on semantic similarity, image clarity, and the presence of mosaic in the image. The data to be recovered is the data sent from the sending end to the receiving end.

[0184] It should be understood that after generating the model update signal, the receiving end not only sends the model update signal to the sending end, but also updates the data to recover the model after generating the model update signal.

[0185] In this embodiment, the receiving end analyzes the data to be recovered sent by the sending end to determine the degree of damage to the data to be recovered. When the degree of damage to the data to be recovered exceeds a second threshold, a model update signal is generated. The data recovery model is then updated after generating the model update signal. By generating the model update signal based on the data to be recovered, the receiving end can promptly trigger updates to the data recovery model and the source compression model based on anomalies in the data to be recovered during communication. This also effectively eliminates the reduction in transmission efficiency caused by communication anomalies and data source anomalies.

[0186] Approach 2:

[0187] The sending end can also send a first string to the receiving end when sending the data to be recovered. After receiving the first string, the receiving end processes it using the data recovery model to obtain the first image. Then, it calculates a third damage value based on the similarity between the first and second images. When the third damage value is greater than a third threshold, it generates a model update signal and sends it back to the sending end. For example, the receiving end can obtain the third damage value by weighting the image clarity of the first image, the presence of mosaic in the image, and the similarity between the first and second images.

[0188] For example, the correspondence between the first string and the second image is a pre-defined correspondence between the receiver and the transmitter. The UE, base station, and server can pre-store the correspondence between reference images and reference strings, where the reference image includes the second image and the reference string includes the first string. The reference string can be a sequence of random codes used to uniquely identify an image. Each image in the reference image has different characteristics; for example, the reference image may include a driving scene, a scene containing animals, and a scene containing a football field. The transmitter only needs to send the first string to the receiver to determine the update and upgrade of the data recovery model at the receiver, thereby triggering an update of the source model.

[0189] It should be understood that after generating the model update signal, the receiving end not only sends the model update signal to the sending end, but also updates the data to recover the model after generating the model update signal.

[0190] In this embodiment, the receiving end uses a data recovery model to process the first string to obtain a first image, and determines a third damage degree based on the similarity between the first image and the second image. When the third damage degree is greater than a third threshold, an update model signal is generated to detect transmission anomalies and anomalies in the data recovery model. This helps to trigger an update of the data recovery model in a timely manner when anomalies occur in the data transmission path or when the data recovery model cannot meet the recovery requirements. This effectively avoids the situation where data transmission continues even when anomalies occur at the receiving end or in the data transmission path, thus improving the reliability of data transmission.

[0191] In this embodiment of the application, when the correspondence between the reference string and the reference image is pre-configured at the receiving end and the sending end, the first string is transmitted from the sending end to the receiving end. The receiving end uses the data recovery model to process the first string to obtain the first image and performs analysis based on the first image, which effectively reduces the amount of data that needs to be transmitted between communication devices and effectively saves network resources.

[0192] For example, the sending end can also send a second image to the receiving end at the same time as sending the first string, so that the receiving end can perform an analysis on whether the model needs to be updated based on the first string and the second image after receiving them. This will not be elaborated here.

[0193] It should be understood that after generating the model update signal, the receiving end not only sends the model update signal to the sending end, but also updates the data to recover the model after generating the model update signal.

[0194] In this embodiment, the sending end directly sends the first string and the second image to the receiving end, without the need to configure the correspondence between the string and the image at the sending end and the receiving end. It is ready to use immediately after sending, effectively eliminating the occupation of the receiving end's storage resources by the correspondence between the string and the image. When the receiving end's storage resources are tight, it can alleviate the pressure on the receiving end's storage resources.

[0195] S630, receiver-side data recovery model;

[0196] It should be understood that the way the receiving end updates the data recovery model is similar to step S540 in Figure 5 above. Please refer to Figure 5 above for details, which will not be repeated here.

[0197] S640, The receiver sends a model update signal to the transmitter;

[0198] When the receiving end is a base station, the receiving end sends model update signals to the sending end through physical layer signaling, such as using DCI signaling messages, MEC CE signaling messages, or RRC signaling messages to send model update signals to the sending end.

[0199] When the receiving end is a server, the receiving end first adds the model update signal to the IP packet or GTP-U extended protocol header, and sends the model update signal to the base station. Then, the base station sends the model update signal to the sending end through physical layer signaling. There are no restrictions here.

[0200] S650, The sending end updates the source compression model.

[0201] It should be understood that the way the sending end updates the source compression model is similar to step S520 in Figure 5 above. Please refer to the description in Figure 5 above for details, which will not be repeated here.

[0202] It should be understood that there is no explicit order between steps S630 and S640 and S650. Steps S630 to S650 can be performed after step S620, and step S640 can be performed before step S650. There are no restrictions here.

[0203] In this embodiment of the application, after generating the model update signal, the receiving end sends the model update signal to the sending end, that is, after generating the model update signal, the receiving end sends the model update signal to the other end, so as to ensure that the models of the sending end and the receiving end are updated at the same time, and to avoid the situation where only one party updates the model.

[0204] The communication method provided in this application has been described above. The communication device provided in this application will be described below with reference to the accompanying drawings. Please refer to the accompanying drawings.

[0205] Please refer to Figure 7, which is a schematic diagram of a communication device 700 provided in an embodiment of this application.

[0206] Processing unit 710 is used to determine model update information, which indicates the timing of model updates;

[0207] The model training unit 720 is used to update the target model based on the model update information. The target model includes a source compression model or a data recovery model. The source compression model is used to extract or compress features from video data or image data to obtain data to be recovered. The data recovery model is used to restore features from the data to be recovered or decompress it to obtain video data or image data.

[0208] Optionally, the processing unit 710 is specifically used to determine the target duration, and the model update information includes the target duration;

[0209] The model training unit 720 is specifically used to update the target model at the first time step. The time interval between the first and second time steps is the target duration, and the second time step is the time step for updating the target model.

[0210] Optionally, the processing unit 710 is further configured to determine target dataset information, which includes at least one of the following: target dataset, or the identity identifier of the target dataset;

[0211] The model training unit 720 is specifically used to update the target model using the target dataset information at the first moment.

[0212] Optionally, the target model is a data recovery model, and the communication device also includes a transmitting unit 730 for transmitting target duration and target dataset information.

[0213] Optionally, the target model is a source compression model, and the transmitting unit 730 is also used to transmit a reference duration, which includes the target duration;

[0214] The processing unit is specifically used to receive the target duration.

[0215] Optionally, the target model is a source compression model, and the communication device 700 also includes a receiving unit 740 for receiving a reference duration, which includes the target duration.

[0216] The processing unit 710 is also used to select a target duration from the reference duration;

[0217] The transmitting unit 730 is also used to transmit the target duration.

[0218] Optionally, the processing unit 710 is specifically used to determine the model update signal, and the model update information includes the model update signal;

[0219] The model training unit 720 is specifically used to update the target model after determining the model update signal.

[0220] Optionally, the model update signal is a signal generated when the distance between the first feature vector and the second feature vector is greater than a first threshold, where the first feature vector and the second feature vector are the results of feature extraction by the target model in adjacent time periods.

[0221] Optionally, the model update signal is a signal generated when the first damage degree or the second damage degree is greater than the second threshold. The first damage degree is the damage degree of the image data generated by the data recovery model based on the data to be recovered, and the second damage degree is the damage degree of the video data generated by the data recovery model based on the data to be recovered.

[0222] Optionally, the model update signal is a signal generated when the third damage degree is greater than the third threshold. The third damage degree is determined based on the similarity between the first image and the second image. The first image is the image obtained by recovering the first string using the data recovery model. The second image is an image that corresponds to the first string and is similar to the image corresponding to the data to be recovered.

[0223] Optionally, the target model is a source compression model, and the processing unit 710 is specifically used to receive model update signals;

[0224] The sending unit 730 is also used to send a first string, or a first string and a second image.

[0225] Optionally, the processing unit 710 is specifically used to generate model update signals;

[0226] The transmitting unit 730 is also used to transmit model update signals.

[0227] Optionally, the processing unit 710 is also used to determine the model to be trained and the preset dataset;

[0228] The model training unit 720 is also used to train the model to be trained using a preset dataset to obtain the target model.

[0229] Please refer to Figure 8, which is a structural schematic diagram of a communication device provided in an embodiment of this application.

[0230] As shown in Figure 8, the communication device 800 includes a bus 803, a memory 804, a processor 805, and a communication interface 806. The processor 805, the memory 804, and the communication interface 806 communicate with each other via the bus 803. The communication device 800 can be a server or a terminal device. It should be understood that this application does not limit the number of processors and memories in the communication device 800.

[0231] Bus 803 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, only one line is used in Figure 8, but this does not indicate that there is only one bus or one type of bus. Bus 803 can include pathways for transmitting information between various components of the communication device 800 (e.g., memory 804, processor 805, communication interface 806).

[0232] The processor 805 may include any one or more processors such as a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP).

[0233] The memory 804 may include volatile memory, such as random access memory (RAM). The processor 805 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0234] The memory 804 stores executable program code to implement the communication method. That is, the memory 804 stores instructions for executing the communication method provided in the embodiments of this application.

[0235] The communication interface 806 uses transceiver units such as, but not limited to, network interface cards and transceivers to enable communication between the communication device 800 and other devices or communication networks.

[0236] The communication device 800 is used to perform the operations performed by the sending end or the receiving end in the foregoing embodiments to implement the communication method provided in the embodiments of this application.

[0237] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0238] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0239] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0240] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0241] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A communication method, characterized in that, include: Determine model update information, which indicates the timing of model updates; The target model is updated based on the model update information. The target model includes a source compression model or a data recovery model. The source compression model is used to extract or compress features from video data or image data to obtain data to be recovered. The data recovery model is used to restore features from the data to be recovered or decompress it to obtain video data or image data.

2. The method according to claim 1, characterized in that, The determined model update information includes: The target duration is determined, and the model update information includes the target duration. Updating the target model based on the model update information includes: The target model is updated at a first time point, and the time interval between the first time point and the second time point is the target duration, where the second time point is the update time of the target model.

3. The method according to claim 2, characterized in that, Before updating the target model at the first moment, the method further includes: Determine target dataset information, wherein the target dataset information includes at least one of the following: the target dataset, or the identity identifier of the target dataset; The step of updating the target model at the first moment includes: The target model is updated using the target dataset information at the first moment.

4. The method according to claim 3, characterized in that, The target model is the data recovery model, and the method further includes: Send the target duration and the target dataset information.

5. The method according to claim 2, characterized in that, The target model is the source compression model. Before determining the target duration, the method further includes: Send a reference duration, wherein the reference duration includes the target duration; The determination of the target duration includes: Receive the target duration.

6. The method according to claim 2, characterized in that, The target model is the source compression model. Before determining the target duration, the method further includes: Receive a reference duration, wherein the reference duration includes the target duration; The determination of the target duration includes: Select the target duration from the reference duration; The method further includes: Send for the target duration.

7. The method according to claim 1, characterized in that, The determined model update information includes: Determine the model update signal, wherein the model update information includes the model update signal; The updating of the target model based on the model update information includes: After determining the model update signal, the target model is updated.

8. The method according to claim 7, characterized in that, The model update signal is generated when the distance between the first feature vector and the second feature vector is greater than a first threshold. The first feature vector and the second feature vector are the results of feature extraction of video data or image data by the source compression model in adjacent time periods.

9. The method according to claim 7, characterized in that, The model update signal is generated when either the first damage degree or the second damage degree is greater than the second threshold. The first damage degree is the damage degree of the image data generated by the data recovery model based on the data to be recovered, and the second damage degree is the damage degree of the video data generated by the data recovery model based on the data to be recovered.

10. The method according to claim 7, characterized in that, The model update signal is a signal generated when the third damage degree is greater than the third threshold. The third damage degree is determined based on the similarity between the first image and the second image. The first image is an image obtained by recovering the first string using the data recovery model. The second image is an image that corresponds to the first string. The second image is similar to the image corresponding to the data to be recovered.

11. The method according to claim 10, characterized in that, The target model is the source compression model, and the determined model update signal includes: Receive the model update signal; Before determining the model update signal, the method further includes: Send the first string, or the first string and the second image.

12. The method according to any one of claims 8 to 11, characterized in that, The determined model update signal includes: Generate the model update signal; After determining the model update signal, the method further includes: Send the model update signal.

13. The method according to any one of claims 1 to 11, characterized in that, Before determining the model update information, the method further includes: Determine the model to be trained and the pre-set dataset; The target model is obtained by training the model to be trained using the preset dataset.

14. A communication device, characterized in that, include: A processing unit is used to determine model update information, wherein the model update information indicates the timing of model updates; The model training unit is used to update the target model based on the model update information. The target model includes a source compression model or a data recovery model. The source compression model is used to extract or compress features from video data or image data to obtain data to be recovered. The data recovery model is used to restore features from the data to be recovered or decompress it to obtain video data or image data.

15. A communication device, characterized in that, Includes a processor, which is coupled to a memory; The memory stores instructions that, when executed on the processor, cause the communication device to perform the method of any one of claims 1 to 13.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a processor, cause the method of any one of claims 1 to 13 to be implemented.

17. A computer program product, characterized in that, When the computer program product is executed on a computer, the method of any one of claims 1 to 13 is implemented.

Images