An image processing method, system, device and medium based on a satellite communication network

Abstract

The application relates to an image processing method, system, device and medium based on a satellite communication network. The method buffers an original byte stream through a ring buffer to resist network fluctuations, scans and extracts a key field of a packet header based on a starting identifier and performs verification to ensure data integrity and sequence; dynamically calls a decoder compatible with standard and non-standard image data according to a data format identifier; inputs an image and imaging metadata into a pre-trained physical information neural network, the network introduces a physical residual loss term calculated based on a differentiable atmospheric radiation transmission model in training to constrain the output and improve visual quality and physical authenticity; and finally, standardizes and encodes an enhanced image and encapsulates and multicasts the same. The method realizes reliable reception, ordered recombination and multi-format decoding on an unreliable satellite link, and generates a high-quality image with better visual quality and higher physical accuracy for distribution to multiple users.

Description

Technical Field

[0001] This invention belongs to the field of satellite communication and remote sensing image processing technology, and in particular relates to an image processing method, system, device and medium based on satellite communication network. Background Technology

[0002] With the deepening of space remote sensing and deep space exploration activities, satellites have become an irreplaceable platform for acquiring visual information from outer space. Satellite communication networks bear the critical task of transmitting image data acquired by remote payloads back to Earth. However, this data link itself and its backend processing system face a series of inherent challenges. In the data transmission stage, the high latency, unstable bandwidth, and susceptibility to interference from the space environment of satellite links cause frequent delays, out-of-order delivery, and even loss of data packets during transmission. Existing ground systems often rely on data retransmission requests to ensure reliable reception, but in scenarios such as one-way transmission or deep space communication where the sending end cannot cooperate, this mechanism fails, resulting in incomplete or disordered image data, affecting the feasibility of subsequent processing.

[0003] During the data parsing phase, spaceborne cameras often use custom compression or encoding formats to output data for specific scientific objectives or to save bandwidth. This makes it impossible for general-purpose ground-based image processing software (such as decoding libraries based on standard JPEG or H264 formats) to directly recognize and interpret these data packets, creating a barrier of "data visible but unusable," which severely restricts the real-time processing and analysis of the data.

[0004] A more critical and in-depth issue lies in the image quality enhancement stage. Current image enhancement techniques commonly used in systems, such as histogram equalization and contrast stretching, are essentially general digital signal processing methods. They primarily focus on improving the subjective visual effect of images, completely ignoring the fact that satellite imaging is a process governed by strict physical laws. Image degradation, such as haze caused by atmospheric scattering, uneven illumination due to changes in the sun's angle, and sensor noise, all correspond to well-defined physical models such as the atmospheric radiative transfer equation. While blindly increasing visual contrast, traditional enhancement methods are prone to distorting or destroying the quantitative information contained in the data that is directly related to the actual physical properties of the Earth's surface (such as reflectivity), and may even introduce artifacts that violate physical laws. This can not only lead to misunderstandings in visual interpretation but also cause subtle and fundamental errors in subsequent quantitative remote sensing inversion and scientific analysis.

[0005] Therefore, the industry urgently needs a comprehensive satellite image processing method that can systematically address the entire chain of issues, from unreliable transmission and heterogeneous data parsing to physical realism enhancement. Summary of the Invention

[0006] Therefore, it is necessary to provide an image processing method, system, device, and medium based on satellite communication networks to address the aforementioned technical problems.

[0007] In a first aspect, this application provides an image processing method based on a satellite communication network, including:

[0008] S1. Write the raw byte stream from the satellite communication network into a circular buffer to obtain buffered byte data; extract the data packet header field by scanning the buffered byte data based on the start identifier; wherein, the data packet header field includes the data packet sequence number, data payload length, timestamp, camera type identifier, data format identifier and check code;

[0009] S2. Perform integrity verification on the corresponding data payload based on the checksum in the data packet header field to obtain the data packet that has passed the verification; sort and reassemble the data packet sequence number and timestamp in the data packet that has passed the verification to obtain an ordered sequence of valid image data packets.

[0010] S3. Based on the data format identifier of each data packet in the ordered valid image data packet sequence, call the corresponding decoding library to decode the data payload and obtain the decoded raw pixel data; bind and encapsulate the raw pixel data with the timestamp, camera type identifier and associated imaging metadata in the corresponding data packet header field to obtain a standardized image data object;

[0011] S4. Input the image data object into the pre-trained physical information neural network model, and obtain the preliminary enhanced image through forward propagation calculation of the physical information neural network model; wherein, during the training process of the physical information neural network model, the network parameters are optimized by calculating the total loss function including the physical residual loss term and backpropagating, and the pre-trained physical information neural network model is obtained; wherein, the physical residual loss term is calculated based on the differentiable atmospheric radiative transfer model, and is used to constrain the network output to conform to the physical laws of satellite imaging;

[0012] S5. Standardize and encode the initially enhanced image to generate a standard format image stream; encapsulate the standard format image stream into network data packets according to the real-time transmission protocol, and send the network data packets to the predetermined receiving end via multicast.

[0013] Secondly, this application also provides an image processing system based on a satellite communication network for implementing the method described in the first aspect, the system comprising:

[0014] The data stream preprocessing module is used to write the raw byte stream from the satellite communication network into a circular buffer to obtain buffered byte data; by scanning the buffered byte data based on the start identifier, the data packet header fields are extracted; the data packet header fields include data packet sequence number, data payload length, timestamp, camera type identifier, data format identifier, and checksum;

[0015] The data packet verification and reassembly module is used to perform integrity verification on the corresponding data payload based on the check code in the data packet header field to obtain the verified data packet; and to sort and reassemble the data packet sequence number and timestamp in the verified data packet to obtain an ordered and valid image data packet sequence.

[0016] The image decoding and encapsulation module is used to call the corresponding decoding library to decode the data payload according to the data format identifier of each data packet in the ordered valid image data packet sequence, so as to obtain the decoded raw pixel data; the raw pixel data is bound and encapsulated with the timestamp, camera type identifier and associated imaging metadata in the corresponding data packet header field to obtain a standardized image data object;

[0017] The physical enhancement module is used to input image data objects into a pre-trained physical information neural network model. Through forward propagation calculation of the physical information neural network model, a preliminary enhanced image is obtained. During the training process of the physical information neural network model, the total loss function including the physical residual loss term is calculated and the network parameters are optimized by backpropagation to obtain the pre-trained physical information neural network model. The physical residual loss term is calculated based on a differentiable atmospheric radiative transfer model and is used to constrain the network output to conform to the physical laws of satellite imaging.

[0018] The image transmission optimization module is used to standardize and encode the initially enhanced image to generate a standard format image stream; it encapsulates the standard format image stream into network data packets according to the real-time transmission protocol and sends the network data packets to the predetermined receiving end via multicast.

[0019] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement an image processing method based on a satellite communication network as described in the first aspect.

[0020] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an image processing method based on a satellite communication network as described in the first aspect.

[0021] The aforementioned image processing method, system, device, and medium based on satellite communication networks construct a complete automated processing chain from receiving the raw byte stream from the satellite communication network to outputting an enhanced and standardized image. This chain employs a circular buffer to smooth network fluctuations, structured parsing and verification based on start identifiers and key fields to handle packet loss and out-of-order delivery while ensuring data integrity, dynamic decoding based on data format identifiers to ensure compatibility with standard and non-standard image data, and inputting image data and imaging metadata into a pre-trained physical information neural network model for forward propagation calculations. The key innovation of this network model lies in its training process incorporating physical residual loss calculated based on a differentiable atmospheric radiative transfer model as a constraint. This ensures that while improving image visual quality, it strictly adheres to the physical laws of satellite imaging. Finally, the initially enhanced image output by the model is standardized, encoded, and encapsulated for multicast forwarding. Ultimately, this achieves the technical effect of reliably receiving and orderly reassembling image data on unreliable satellite links, flexibly decoding various proprietary formats, and generating high-quality standardized image data with superior visual effects and greater physical accuracy for distribution to multiple users. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating an image processing method based on a satellite communication network provided by the present invention;

[0024] Figure 2 This is a schematic diagram of the process of training a physical information neural network model in one optional embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of the structure of an image processing system based on a satellite communication network provided by the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] Example 1:

[0028] refer to Figure 1The document presents a flowchart illustrating an image processing method based on a satellite communication network, which includes the following steps:

[0029] S1. Write the raw byte stream from the satellite communication network into a circular buffer to obtain buffered byte data; extract the data packet header field by scanning the buffered byte data based on the start identifier; the data packet header field includes the data packet sequence number, data payload length, timestamp, camera type identifier, data format identifier, and checksum.

[0030] Specifically, the raw byte stream from the satellite communication network is written to a circular buffer to obtain buffered byte data. This step is designed to address the inherent high latency, bandwidth fluctuations, and bursty data transmission characteristics of satellite communication links. The circular buffer's unique memory structure decouples data reception from subsequent processing. The circular buffer employs a contiguous circular memory block architecture. Its technical principle lies in using independent read and write pointers to achieve concurrent read and write operations, avoiding read / write conflicts and data overwriting risks caused by contiguous addresses in traditional linear buffers. In implementation, the circular buffer's memory allocation uses a dynamic memory pool mechanism, adaptively adjusting the total buffer capacity based on the peak rate and average packet size of satellite data transmission, ensuring stable data storage capabilities even during sudden changes in link bandwidth. The reception of the raw byte stream is handled by the image access module. This module adapts to the transmission protocols supported by the satellite communication network, dynamically adjusting the receive window size using a sliding window flow control mechanism to match the link transmission rate and buffer processing capacity. Simultaneously, mutexes and atomic operations ensure synchronized updates of read and write pointers, guaranteeing thread safety in a multi-threaded environment.

[0031] The technique involves scanning the buffered byte data based on a start identifier to extract packet header fields. This technique relies on a pre-defined, fixed feature sequence to accurately identify packet boundaries. The start identifier must meet the requirements of uniqueness and anti-interference. It is determined through a random sequence generation algorithm combined with a collision detection mechanism, ensuring a very high probability of no repetition with the byte combinations in the satellite image data payload, thus avoiding misidentification. The scanning operation starts from the starting address of the buffered byte data and performs sliding comparisons in single-byte increments. When a continuous byte sequence of a pre-defined length is detected that perfectly matches the start identifier, the current position is marked as the header start point. Subsequently, each header field is extracted sequentially according to a pre-defined field definition order. The structure of the data packet header fields strictly adheres to standardized data transmission specifications. The data packet sequence number identifies the transmission order of data packets and is stored in an unsigned integer format to ensure cross-platform compatibility. The data payload length field specifies the number of bytes in the subsequent data payload, providing a length basis for data parsing. The timestamp field accurately records the specific time of satellite imaging, providing a time reference for subsequent data packet sorting. The camera type identifier field uses a binary encoding format, distinguishing different imaging devices on the satellite through different encoding values, providing a basis for subsequent decoding library selection and image processing parameter configuration. The data format identifier field specifies the compression or encoding format of the data payload, establishing a mapping relationship with the decoding library. The checksum field is used for data integrity verification, providing a basis for subsequent verification steps. During field extraction, a boundary detection mechanism ensures that the extraction range of each field does not exceed the preset length. Simultaneously, the extracted field data undergoes format conversion and validity verification to exclude illegal field data, ensuring the stability of subsequent processing steps.

[0032] S2. Perform integrity verification on the corresponding data payload based on the checksum in the data packet header field to obtain the data packet that has passed the verification; sort and reassemble the data packet sequence number and timestamp in the data packet that has passed the verification to obtain an ordered sequence of valid image data packets.

[0033] Specifically, the integrity of the corresponding data payload is verified based on the checksum in the packet header field to obtain a verified data packet. The technical principle is to use a verification algorithm to calculate the characteristic value of the data payload and compare it with the preset checksum in the packet header to detect problems such as bit flipping, byte loss, or misalignment caused by link interference during data transmission. In specific implementation, the verification algorithm adopts the Cyclic Redundancy Check (CRC) mechanism. Its calculation process follows the polynomial division principle, treating the data payload as a binary polynomial and performing a modulo-2 operation with a preset generator polynomial to obtain a fixed-length verification result; the corresponding formula is: ,in This represents the binary polynomial corresponding to the data payload. Indicates will Shift left Bit( (the degree of the generating polynomial). This represents the preset generator polynomial, and mod represents modulo 2 division. In this formula, The data payload is composed of each byte in sequence, and each byte corresponds to a coefficient of a polynomial. The selection must meet error correction capability requirements and be determined through rigorous algorithm screening; the calculation results... This is the checksum, which is compared bit by bit with the checksum field in the data packet header. If they match perfectly, the data payload is considered error-free during transmission and is marked as a verified data packet. If there is a discrepancy, the data packet is considered invalid, discarded, and a verification failure log is recorded. The log contains relevant packet identification information and the reason for the failure, providing a basis for troubleshooting link faults. To improve verification efficiency, a dedicated instruction set is used to accelerate the verification algorithm when hardware support is available. Simultaneously, a data block verification mechanism is used to divide large data payloads into multiple smaller blocks for parallel verification, shortening the verification operation time and meeting real-time processing requirements.

[0034] This method uses data packet sequence numbers and timestamps to sort and reassemble data packets to obtain an ordered and valid sequence of image data packets. Its purpose is to solve the problem of out-of-order data packets caused by delays and interference in satellite link transmission, and to restore the original transmission order of image data. The sorting operation employs a stable merge sort algorithm, which has the advantage of handling large-scale data sorting and maintaining high sorting efficiency even with large data volumes, while keeping time complexity optimized. During the sorting process, the data packet sequence number is the primary sorting criterion. Sequence numbers are continuously assigned according to the data packet generation order at the satellite transmitter, and the receiver determines the logical order of the data packets through the incrementing relationship of the sequence numbers. The timestamp serves as an auxiliary sorting criterion to handle special scenarios such as missing or abnormal sequence numbers. When a discontinuous sequence number is detected, the timestamps of adjacent data packets are extracted and the difference is calculated. If the timestamp difference is within a preset threshold range, it is determined that the out-of-order sequence is caused by transmission delay, and the corresponding data packet is temporarily stored in a temporary buffer, awaiting the reception of the data packet with the missing sequence number before sorting. If the timestamp difference exceeds the preset threshold, it is determined that the missing data packet has been lost, and a data completion mechanism is initiated. The data completion mechanism performs interpolation based on the imaging features of adjacent data packets. By extracting texture, grayscale, and other feature information from the data payloads of adjacent data packets, it uses an interpolation algorithm to generate complete data corresponding to missing data packets. This ensures the continuity of the reconstructed effective image data packet sequence, providing a complete data foundation for subsequent decoding and image processing. During the sorting and reconstructing process, a data packet index table is established to record the sorting status and storage address of each data packet, improving data access efficiency. Simultaneously, a timeout retransmission detection mechanism is used to identify missing data packets that have not been received for an extended period, preventing processing blockage caused by waiting for missing data packets.

[0035] S3. Based on the data format identifier of each data packet in the ordered valid image data packet sequence, call the corresponding decoding library to decode the data payload and obtain the decoded raw pixel data; bind and encapsulate the raw pixel data with the timestamp, camera type identifier and associated imaging metadata in the corresponding data packet header field to obtain a standardized image data object.

[0036] Specifically, based on the data format identifier of each data packet in the ordered sequence of valid image data packets, the corresponding decoding library is invoked to decode the data payload and obtain the decoded raw pixel data. The key technology in this step lies in establishing a mapping relationship between data format identifiers and decoding libraries, and implementing on-demand calling of the decoding library through a dynamic library loading mechanism. The mapping relationship between data format identifiers and decoding libraries is managed through a configuration file, which records the decoding library paths, interface function names, and parameter requirements corresponding to various data format identifiers. During implementation, this configuration file is loaded and a mapping table is established during system startup. When a decoding operation is required, the mapping table is queried based on the data format identifier of the current data packet to obtain the corresponding decoding library information. The decoding library is then loaded into memory through a dynamic link library loading function, and the decoding interface functions in the library are called to decode the data payload. For data payloads in standard encoding formats, mature open-source decoding libraries are used for decoding. For example, standard formats for still images use dedicated still image decoding libraries, and standard formats for moving images use dedicated moving image decoding libraries. These open-source decoding libraries have been proven through long-term practice and have high decoding efficiency and stability. For non-standard compression or encoding formats customized by satellite cameras, specially developed dedicated decoding libraries are used. These decoding libraries are customized for the compression algorithms and data structures of the corresponding encoding formats to ensure the accuracy and completeness of the decoding process.

[0037] The specific implementation of the decoding process must follow the decoding specifications of the corresponding encoding format. Taking a custom compression format as an example, the decoding library first decompresses the data payload, recovers the original data by reversing the compression algorithm in the encoding process, and then performs format conversion, converting the decompressed data stream into a standard pixel data format. During the decoding process, key parameters are dynamically adjusted. For example, the imaging parameters of the corresponding camera are obtained based on the camera type identifier in the data packet header, and these imaging parameters are input into the decoding library as decoding parameters to ensure that the decoded original pixel data accurately reflects the real scene of the satellite imaging. After decoding, the original pixel data is validated. Invalid data is excluded by detecting indicators such as the range of pixel values ​​and data length. At the same time, errors that occur during the decoding process are captured and processed, and error information is recorded through the log module to ensure the stable operation of the system.

[0038] The process involves binding and encapsulating raw pixel data with the corresponding timestamp, camera type identifier, and associated imaging metadata from the data packet header to obtain standardized image data objects. The technology lies in establishing a unified data object model to achieve the integration and standardized representation of multi-source data. The standardized image data object adopts an object-oriented design, comprising a data storage area and a metadata area. The data storage area stores the decoded raw pixel data, using contiguous memory space to store pixel values ​​in row-major order, supporting common pixel data formats. The metadata area stores various bound auxiliary information, including the timestamp and camera type identifier extracted from the data packet header, and imaging metadata obtained from the satellite imaging system. This imaging metadata includes key information such as satellite orbit parameters, imaging attitude, focal length, and exposure time, providing crucial parameter support for subsequent image processing. During encapsulation, a data association mechanism establishes a mapping relationship between the raw pixel data and various metadata types, ensuring rapid retrieval of auxiliary information corresponding to the image data in subsequent processing. Simultaneously, data serialization technology converts the standardized image data object into a storable and transmittable binary format, and data compression algorithms compress the metadata area, reducing data storage space and improving data transmission efficiency. After encapsulation, the image data object is subjected to integrity verification to ensure that the original pixel data and metadata are complete and error-free, providing standardized and high-quality data input for subsequent image processing.

[0039] S4. Input the image data object into the pre-trained physical information neural network model, and obtain the preliminary enhanced image through forward propagation calculation of the physical information neural network model; wherein, during the training process of the physical information neural network model, the network parameters are optimized by calculating the total loss function including the physical residual loss term and backpropagating, and the pre-trained physical information neural network model is obtained; wherein, the physical residual loss term is calculated based on the differentiable atmospheric radiative transfer model and is used to constrain the network output to conform to the physical laws of satellite imaging.

[0040] Specifically, the image data object is input into a pre-trained physical information neural network model. The model's forward propagation computes the initial enhanced image. The key technology here is integrating the physical laws of satellite imaging into the neural network model, achieving image enhancement that combines visual appeal with physical realism. The physical information neural network model's structure is based on a deep learning encoder-decoder architecture. The encoder consists of multiple convolutional layers, batch normalization layers, and activation function layers. It extracts multi-scale features from the image data by progressively increasing the convolutional kernel size and the number of channels, performing progressive feature learning from low-level features (such as edges and textures) to high-level features (such as semantic information). The decoder uses transposed convolutional layers to upsample the feature maps, gradually restoring the image's spatial resolution. Simultaneously, a skip connection mechanism fuses the features extracted by the encoder at each level with the corresponding feature maps from the decoder, preserving the image's detailed information. The model's input is the raw pixel data from the standardized image data object. The input data is first normalized, mapping pixel values ​​to a preset range to improve the model's training stability and convergence speed.

[0041] During the forward propagation of the model, the original pixel data sequentially undergoes feature extraction by the encoder and feature reconstruction by the decoder, ultimately outputting a preliminary enhanced image. The forward propagation computation process follows the signal transmission rules of neural networks. Each convolutional layer performs convolution operations with the input feature map through a convolution kernel, followed by data standardization processing by a batch normalization layer after a bias term. Subsequently, a nonlinear transformation is introduced through an activation function to enhance the model's feature representation capability. The transposed convolutional layer enlarges the feature map size through inverse convolution operations, while skip connections fuse the high-resolution features of the encoder with the low-resolution features of the decoder through feature map concatenation operations, supplementing image detail information and improving the clarity of the enhanced image.

[0042] The training process of the physical information neural network model is crucial to ensuring its performance. It involves guiding the optimization of network parameters through a total loss function that includes a physical residual loss term, ensuring that the model output conforms to the physical laws of satellite imaging. The formula for calculating the total loss function is: ,in Represents the total loss function. Represents the pixel-level loss term. Represents the physical residual loss term. Represents the regularization loss term. , , These are the weight coefficients for each loss term, used to balance the impact of different loss terms on model training.

[0043] The mean squared error loss function is used to measure the difference in pixel values ​​between the model's output image and the true label image. Its calculation formula is as follows: ,in This represents the total number of pixels in the image. The first image representing the real label image pixel value, The first image represents the output image of the model. Each pixel value is a loss term that ensures the enhanced image is visually consistent with the real scene. The calculation is based on a differentiable atmospheric radiative transfer model, which describes the physical process of electromagnetic wave propagation in the atmosphere. Its differentiability is achieved by transforming the physical parameters in the model into learnable variables, allowing the model to optimize these parameters through backpropagation, thereby constraining the network output to conform to physical laws. The calculation formula is ,in This represents a differentiable atmospheric radiative transfer model. This represents the input image to the model. Physical parameters representing atmospheric radiative transfer models (such as atmospheric optical thickness, aerosol scattering coefficient, etc.). This represents the model's output image. This loss term calculates the difference between the model's output and the prediction results of the atmospheric radiative transfer model, ensuring that the enhanced image does not violate the physical laws of satellite imaging. Represents the L2 norm; An L2 regularization term is used to suppress model overfitting; its calculation formula is as follows: ,in This represents the set of all learnable parameters in the model. This represents a single parameter value. By penalizing the model parameters, overfitting caused by excessively large parameter values ​​can be avoided.

[0044] During model training, a training dataset containing satellite image data and corresponding real labels is first constructed. Preprocessing operations (such as data augmentation and normalization) are performed on the training data to improve the model's generalization ability. Then, the model parameters are initialized, and hyperparameters such as the number of training iterations and the learning rate are set. In each iteration, the training data is input into the model for forward propagation calculation to obtain the model output and the total loss value. The gradient of the total loss function with respect to each model parameter is calculated using the backpropagation algorithm, and the model parameters are updated using the gradient descent optimization algorithm. The above process is repeated until the number of training iterations reaches a preset value or the total loss value converges to a preset threshold, completing the model training and obtaining the pre-trained physical information neural network model.

[0045] S5. Standardize and encode the initially enhanced image to generate a standard format image stream; encapsulate the standard format image stream into network data packets according to the real-time transmission protocol, and send the network data packets to the predetermined receiving end via multicast.

[0046] Specifically, the initial enhanced image undergoes standardized encoding to generate a standard format image stream. The purpose is to convert the enhanced image data into a universal standard format, improving data compatibility and transmittability. The standardized encoding process selects a suitable standard encoding format based on the application scenario and the support of the receiving end. Static images typically use a universal static image standard format, while dynamic images use a universal video standard format. The specific implementation of the encoding process follows the encoding specifications of the corresponding standard format. Taking static image encoding as an example, the encoding process first performs color space conversion on the initial enhanced image, converting the image's color space to a color space supported by the encoding standard. Then, a discrete cosine transform is performed, converting the image's pixel values ​​from the spatial domain to the frequency domain, achieving data compression by removing correlations between pixels. Next, quantization processing is performed, quantizing the coefficients after the discrete cosine transform to further compress the data volume. The quantization step size can be dynamically adjusted according to image quality requirements. Finally, entropy encoding is performed, losslessly compressing the quantized coefficients to generate a standard format image stream.

[0047] During the encoding process, an adaptive encoding parameter adjustment mechanism optimizes the balance between encoding quality and efficiency. For example, the quantization step size and encoding mode are dynamically adjusted based on the image's texture complexity, brightness distribution, and other characteristics. For image regions with complex textures and rich details, a smaller quantization step size and a more complex encoding mode are used to preserve image details; for image regions with simple textures and uniform brightness, a larger quantization step size and a simpler encoding mode are used to improve encoding efficiency. Simultaneously, the generated image bitstream undergoes integrity verification during encoding to ensure that the bitstream is free of syntax errors and data loss, providing a reliable data foundation for subsequent transmission.

[0048] Encapsulating standard format image streams into network packets using the Real-Time Transport Protocol (RTP) involves adding appropriate header fields according to the RTP protocol specifications to ensure the identifiability and recoverability of the packets during network transmission. The RTP header fields are encapsulated strictly according to the protocol standard, including a version number field to identify the RTP protocol version, ensuring protocol compatibility between the receiver and sender; a sequence number field to identify the transmission order of each RTP packet, providing a basis for packet sorting and loss detection at the receiver; a timestamp field to synchronize playback time at the receiver, ensuring real-time playback of image data; and a payload type field to specify the type of payload data in the RTP packet, establishing a connection for decoding at the receiver.

[0049] The aforementioned image processing method based on satellite communication networks constructs a complete automated processing chain from receiving the raw byte stream from the satellite communication network to outputting an enhanced and standardized image. This chain employs a circular buffer to smooth network fluctuations, structured parsing and verification based on start identifiers and key fields to handle packet loss and out-of-order delivery while ensuring data integrity, dynamic decoding based on data format identifiers to ensure compatibility with standard and non-standard image data, and inputting image data and imaging metadata into a pre-trained physical information neural network model for forward propagation calculations. The key innovation of this network model lies in its training process incorporating physical residual loss calculated based on a differentiable atmospheric radiative transfer model as a constraint. This allows for strict adherence to the physical laws of satellite imaging while improving image visual quality. Finally, the initially enhanced image output by the model is standardized, encoded, and encapsulated for multicast forwarding. Ultimately, this achieves the technical effect of reliably receiving and orderly reassembling image data on unreliable satellite links, flexibly decoding various proprietary formats, and generating high-quality standardized image data with superior visual effects and greater physical accuracy for distribution to multiple users.

[0050] In one optional embodiment, the image data object is input into a pre-trained physical information neural network model, and a preliminary enhanced image is obtained through forward propagation calculation of the physical information neural network model, including the following steps:

[0051] S11. Based on the imaging metadata associated with the image data object, extract the parameter combination including solar zenith angle, solar azimuth angle, observation zenith angle, observation azimuth angle and atmospheric optical thickness parameters, and use the parameter combination to construct an imaging physical condition vector to describe the imaging scene.

[0052] Specifically, based on the imaging metadata associated with image data objects, specific parameter combinations are extracted and an imaging physical condition vector is constructed. The technical logic of this process lies in accurately capturing key physical influencing factors in the satellite imaging process, providing a physical constraint basis for subsequent physical information neural network models that matches the real scene, and ensuring that the model output conforms to the objective physical laws of satellite imaging. As a structured record of the entire satellite imaging process, imaging metadata covers multi-dimensional information such as satellite platform operating status, sensor operating parameters, and environmental meteorological conditions. The solar zenith angle, solar azimuth angle, observation zenith angle, observation azimuth angle, and atmospheric optical thickness are core parameters that directly dominate the radiative transfer process and imaging quality. The accuracy and completeness of their extraction directly determine the effectiveness of the physical constraints.

[0053] The calculation of the solar zenith angle is based on the principles of spherical trigonometry, and the corresponding formula is: .in Represents the solar zenith angle, which is the angle between the incident sunlight and the normal direction of the image target area; The solar declination angle is calculated from the imaging date using the laws of celestial motion, reflecting the sun's position on the celestial sphere relative to the equator. The geographic latitude of the imaging area is extracted from the target area coordinate information of the imaging metadata. The solar azimuth angle is calculated by converting the imaging timestamp to Universal Time and then combining it with the geographical longitude of the imaging area. It represents the time shift of the sun relative to local noon. The formula for calculating the solar azimuth angle is: ,in The solar azimuth angle is the angle between the projection of the sun's incident rays onto the horizontal plane and the direction of true north. The other parameters are consistent with the definition in the formula for calculating the solar zenith angle. This formula is derived from the geometric relationship between the solar altitude angle and the geographic coordinates of the target area, ensuring the physical rigor of the angle calculation.

[0054] The extraction of the observation zenith angle and observation azimuth angle depends on the relative position of the satellite's attitude parameters and the target area. The observation zenith angle is the angle between the satellite's observation direction and the normal direction of the target area, while the observation azimuth angle is the angle between the projection of the observation direction onto the horizontal plane and true north. Both are calculated through coordinate transformation between the satellite's body coordinate system and the Earth's fixed coordinate system. In practice, attitude parameters such as the satellite's roll angle, pitch angle, and yaw angle are extracted from imaging metadata. Combined with the position vector of the target area relative to the satellite, the attitude parameters are converted into observation angles through matrix transformation. The entire transformation process strictly follows the rules of spatial geometric transformation to ensure that the observation angles are consistent with the actual imaging perspective.

[0055] Atmospheric optical thickness parameters are acquired using a dual-channel complementary mechanism. When the imaging metadata contains direct measurements from the onboard atmospheric sounding module, these measurements are directly extracted; this value reflects the attenuation of electromagnetic waves propagating in the atmosphere. When no direct measurements are available in the metadata, the parameters are retrieved using a mature atmospheric radiative transfer model based on parameters such as solar zenith angle, imaging time, and target area altitude. A pixel-by-pixel adaptive adjustment strategy is employed during the retrieval process. The land cover type identifier in the metadata is used to determine the target area's surface type, and the retrieval parameters are adjusted for the radiative characteristics of different surface types to improve the accuracy of atmospheric optical thickness calculation.

[0056] After all parameters are extracted, standardization is performed to eliminate differences in the units of measurement between different parameters, thus preventing model training imbalance caused by significant differences in numerical range. Standardization uses the Z-score standardization formula. ,in These are the standardized parameter values. These are the original parameter values. To obtain the statistical mean of this parameter in the training dataset, The statistical standard deviation of this parameter in the training dataset is determined. The standardized parameter values ​​will be distributed within a reasonable range, ensuring a balanced weighting of each parameter's influence on model training. Finally, following a fixed parameter order, the five standardized parameters are combined to construct an imaging physical condition vector. This vector is stored using high-precision floating-point data type, providing reliable physical parameter input for feature fusion in the model.

[0057] S12. Perform radiometric calibration and normalization preprocessing on the pixel data in the image data object to obtain the normalized network input tensor.

[0058] Specifically, the pixel data in the image data object undergoes radiometric calibration and normalization preprocessing to obtain a normalized network input tensor. The purpose of this step is to convert the raw digital quantized values ​​output by the sensor into physically meaningful radiance data, while unifying the data value range to adapt to the input requirements of the neural network, thus laying a data foundation for efficient model training and accurate inference. The essence of radiometric calibration is to establish a quantitative mapping relationship between the original pixel values ​​and the true radiance of ground objects. Its technical principle is based on the linear response characteristics of satellite sensors, ensuring that the preprocessed data can accurately reflect the radiometric characteristics of ground objects.

[0059] The formula for calculating radiation calibration is as follows: ,in Representing radiometrically calibrated radiance, it is a physical quantity with a clear physical meaning, reflecting the intensity of radiant energy of ground objects in a specific wavelength band; The gain coefficient of the sensor characterizes the amplification factor of the sensor to incident radiation, and is related to the optical system design and the performance of electronic components of the sensor. This is the offset coefficient of the sensor, used to correct the system error of the sensor, and is determined by the sensor's factory calibration and on-orbit calibration data; The original pixel values ​​in the image data object are digital quantization results obtained by the sensor after photoelectric conversion of the incident radiation signal. Both the gain coefficient and offset coefficient are extracted from the imaging metadata associated with the image data object. The metadata stores corresponding calibration coefficients categorized by camera type and imaging mode. The system automatically matches and calls the corresponding coefficients for calculation based on the camera type identifier in the image data object.

[0060] During radiometric calibration, a data validity detection mechanism needs to be embedded to remove abnormal pixels that exceed the sensor's dynamic range. For saturated pixels below the lower limit or above the upper limit of the dynamic range, a neighborhood mean replacement strategy is used for correction. This involves selecting a neighborhood window around the pixel, calculating the mean radiance of normal pixels within the window, and replacing the radiance value of the abnormal pixel with this mean value, ensuring data continuity and integrity. The size of the neighborhood window is adaptively adjusted according to the spatial resolution of the image to avoid problems such as blurred details due to an excessively large window or ineffective correction of abnormalities due to an excessively small window.

[0061] After radiometric calibration, normalization is required. This normalization maps the radiance data to a numerical range easily processed by the neural network, reducing the oscillations caused by large numerical differences during model training and improving training stability and convergence speed. This implementation uses the min-max normalization method, and the corresponding calculation formula is as follows: ,in These are the normalized pixel values. This represents the minimum radiance value corresponding to a specific target type. This represents the maximum radiance value corresponding to a specific target type. and Statistical analysis of a large amount of similar satellite imaging data revealed that the metadata contains an image target type identifier. The system calls the corresponding statistical extreme value based on this identifier to ensure the relevance and accuracy of the normalization process and avoid feature distortion caused by the mixed normalization of data of different target types.

[0062] After normalization, the two-dimensional image pixel matrix needs to be converted into a three-dimensional network input tensor. The tensor's dimensions consist of image height, image width, and the number of channels. The number of channels is determined by the type of imaging camera; single-band cameras correspond to single-channel tensors, and multispectral cameras correspond to multi-channel tensors. The tensor is stored using high-precision floating-point data type, and the memory arrangement prioritizes channels to improve the memory access efficiency of the neural network when reading data and accelerate the model inference process. At the end of the entire preprocessing process, a data quality verification step is set up. The mean and standard deviation of the preprocessed data are calculated and compared with a preset reasonable range. If they exceed the range, the preprocessing is judged as a failure, and the preprocessing process is automatically triggered again to ensure the reliability of the tensor data input to the neural network.

[0063] S13. The encoder of the physical information neural network model is input with the network input tensor and the imaging physical condition vector as input data. Through the multi-layer convolution and downsampling operations of the encoder, a multi-scale feature map set is extracted. The multi-scale feature map set contains feature maps of multiple scales from shallow to deep.

[0064] Specifically, the encoder of the physical information neural network model takes the network input tensor and the imaging physical condition vector as input data. Through multi-layer convolution and downsampling operations, it extracts a multi-scale feature map set. This process constructs a deep feature representation that can simultaneously capture both visual image features and physical constraint information, providing a rich and complete feature foundation for subsequent decoding, ensuring that the model can fully utilize the correlation between image data and physical conditions. The encoder adopts a deep convolutional neural network architecture, consisting of alternating stacked convolutional and downsampling layers. The convolutional layers are responsible for extracting feature information, while the downsampling layers are responsible for reducing the feature map size, expanding the receptive field, and preserving key feature information.

[0065] The core calculation formula for convolution operation is: ,in Indicates the spatial location of the feature map after convolution. ,aisle The eigenvalue at a given location reflects the intensity of the characteristic response at that location on a specific channel; Indicates the spatial location of the convolution kernel Input Channel Output channels The weights at each position are learnable parameters of the convolutional layer, which are continuously optimized through model training. This represents the network input tensor at its corresponding spatial location and input channel. The value at that location; Indicates output channel The bias term is used to adjust the baseline level of feature values, improving the flexibility of feature representation. Appropriate padding is used during convolution to ensure that the spatial size of the feature map after convolution remains consistent with the input, avoiding the loss of edge feature information.

[0066] After the convolution operation is completed, a batch normalization layer is applied to standardize the feature values. The formula for batch normalization is as follows: and .in These are the standardized intermediate feature values. These are the original feature values ​​after convolution. The mean of all feature values ​​within the batch. The variance of all eigenvalues ​​within the batch. It is a very small constant used to avoid the case where the denominator is zero; These are the final eigenvalues ​​after batch normalization. For learnable scaling factors, These are learnable bias coefficients. The introduction of batch normalization layers can effectively reduce the risk of gradient vanishing, accelerate model convergence, and suppress overfitting.

[0067] After batch normalization, a nonlinear transformation is introduced through an activation function to enhance the model's ability to express complex features. This implementation uses the GELU activation function, whose calculation formula is as follows: ,in These are the eigenvalues ​​after batch normalization. The cumulative distribution function is the standard normal distribution. Compared to traditional activation functions, the GELU activation function can better preserve gradient information, especially in deep networks, effectively mitigating gradient decay and improving model training stability and feature extraction capabilities.

[0068] After the convolutional layers are processed, a downsampling layer is applied to reduce the size of the feature map. Downsampling is achieved using convolutional operations with a stride greater than 1. Compared to traditional max pooling or average pooling, this method can retain more feature information while reducing size, thus minimizing information loss. During downsampling, the spatial size of the feature map is reduced by a fixed ratio, while the number of channels increases according to a certain pattern. Through this "size reduction and channel expansion" approach, the abstraction level and semantic expressive power of the feature map are gradually improved.

[0069] The integration of imaging physical condition vectors employs a feature concatenation mechanism. In the intermediate layer of the encoder, the physical condition vectors are dimensionally expanded to match the spatial dimensions of the current layer's feature map. This is then converted into a single-channel feature map, which is concatenated with the multi-channel feature map output from the convolutional layer of that layer to form a feature map incorporating physical information. This fusion method ensures that physical condition information deeply participates in subsequent feature extraction processes, enabling the extracted multi-scale features to contain both visual details of the image and the physical constraints of the imaging scene, thus supporting the model's physically realistic output.

[0070] The multi-scale feature map set contains feature maps at multiple levels, from shallow to deep. Shallow feature maps mainly capture low-level visual features such as edges and textures, mid-level feature maps integrate local semantic and physical information, and deep feature maps reflect the global physical properties and high-level semantic information of the image. During feature extraction, a gradient clipping mechanism is used to limit the gradient magnitude, avoiding gradient explosion during training. Simultaneously, an appropriate weight initialization method is used to initialize the convolutional kernel weights, ensuring consistent variance of feature values ​​across layers and laying the foundation for stable model training.

[0071] S14. Take the deepest feature map in the multi-scale feature map set as the initial input feature map of the decoder, perform a second upsampling operation on the initial input feature map of the decoder to obtain the upsampled feature map; perform skip connections and channel concatenation between the upsampled feature map and the encoder feature map of the corresponding scale in the multi-scale feature map set to obtain the concatenated feature map.

[0072] Specifically, the deepest feature map in the multi-scale feature map set is used as the initial input feature map of the decoder. A stitched feature map is constructed through upsampling, skip connections and channel concatenation operations. This step gradually restores the spatial resolution of the feature map through upsampling. At the same time, skip connections are used to introduce feature maps of different levels of the encoder into the decoder to supplement detailed information and achieve effective fusion of shallow and deep features. This ensures that the final output image has both global semantic consistency and retains rich detailed information.

[0073] The deepest feature map, as the final result of the encoder's feature extraction, possesses the highest level of abstraction and the richest global semantic information, but its spatial resolution is the lowest. Therefore, it is used as the initial input to the decoder, and the feature map size is gradually increased through upsampling operations. Upsampling is implemented using transposed convolution, and the core calculation formula for transposed convolution is... ,in Indicates the spatial location of the upsampled feature map. ,aisle Eigenvalues ​​at; Indicates the spatial location of the transposed convolution kernel Input Channel Output channels The weight value at the location; This indicates the input feature map of the decoder at the corresponding spatial location and input channel. Eigenvalues ​​at; Indicates output channel The bias term. Transposed convolution, through inverse convolution, can learn how to reconstruct detailed information while expanding the feature map space size, and has a stronger feature learning ability compared to traditional interpolation upsampling.

[0074] After the first upsampling operation, the first upsampled feature map is obtained, whose spatial resolution is doubled compared to the initial input feature map. The number of channels is adjusted according to the decoder design. Then, skip connections and channel concatenation operations are performed. The core logic of skip connections is to introduce a feature map from the encoder with the same spatial resolution as the current upsampled feature map into the decoder. This encoder feature map contains rich detail information at that resolution, while the upsampled feature map contains deep semantic information. The fusion of the two achieves complementarity between detail and semantics.

[0075] The channel concatenation operation merges the upsampled feature map with the corresponding encoder feature map along the channel dimension to form a concatenated feature map. During concatenation, it is crucial to ensure that the spatial resolution of the two feature maps is completely identical. If slight differences exist, adaptive interpolation is used to adjust the size of the encoder feature map to ensure the accuracy of the concatenation operation. The number of channels in the concatenated feature map is the sum of the number of channels in the upsampled feature map and the encoder feature map. In this way, the detailed features extracted by the encoder and the semantic features reconstructed by the decoder are fully integrated, providing a rich feature foundation for subsequent convolutional fusion.

[0076] S15. Perform convolution processing on the stitched feature map to obtain a fused feature map; use the fused feature map as the input of the decoder, and repeatedly perform upsampling, skip connections and channel stitching in sequence until the output feature map is restored to the original image spatial resolution to obtain the final decoded feature map; perform convolution and activation processing on the final decoded feature map to output a preliminary enhanced image.

[0077] Specifically, a fused feature map is obtained by convolutional processing of the concatenated feature map. This convolutional processing performs deep fusion of the multi-channel features after concatenation, eliminating redundant information and conflicts between features from different sources, and extracting more representative fused features. The parameter settings of the convolutional layer are adjusted according to the number of channels and feature complexity of the concatenated feature map, using appropriate convolutional kernel size and number. Non-linear fusion of features is achieved through convolutional operations, so that the fused feature map retains both the detailed information of the encoder features and the semantic and physical constraint information of the decoder features.

[0078] The fused feature map is used as the next input to the decoder, and the upsampling, skip connections, and channel concatenation processes are repeated. Each repetition follows the same technical logic: the spatial resolution of the feature map is doubled through transposed convolution, and the weight parameters of the transposed convolution are adaptively optimized through model training to ensure that the detailed features at the corresponding scale can be accurately reconstructed in each upsampling process; then, the feature map of the same scale in the encoder is called through skip connections. This feature map has captured low-level visual features such as edges and textures and local physical property association information at this resolution in the encoder stage, which complements the semantic feature map after upsampling in the decoder; the channel concatenation operation fully merges the two types of feature maps in the channel dimension, so that the concatenated feature map simultaneously carries detailed information, semantic information, and physical constraint information, providing comprehensive feature support for subsequent convolutional fusion.

[0079] The number of times the above process is repeated is determined by the number of downsampling operations performed by the encoder, ensuring that the spatial resolution of the final output feature map is completely consistent with the resolution of the original image data object. This yields the final decoded feature map. The number of channels in the final decoded feature map is the total number of channels after the last stitching operation of the decoder, encompassing full-scale feature information from shallow to deep layers, as well as imaging physical constraints.

[0080] The final decoded feature map is subjected to convolution and activation processing. This convolutional layer uses a 1×1 convolutional kernel, and its core calculation formula is as follows: ,in This indicates the spatial location of the output feature map after convolution. The value at channel k; This represents the weight values ​​of the 1×1 convolution kernel at input channel c and output channel k; This represents the feature value of the final decoded feature map at the corresponding spatial location and input channel c; This represents the bias term for output channel k. The role of the 1×1 convolution kernel is to adjust the dimensionality and linearly fuse the multi-channel features of the final decoded feature map, mapping the high-dimensional features to the same output dimension as the number of channels in the input image, while preserving the key information and physical relationships of each channel feature.

[0081] After convolution, an activation function is applied. The Sigmoid activation function is used to perform a non-linear transformation on the output feature values. The calculation formula is as follows: Where x is the feature value after convolution, The output value is the result of activation. The Sigmoid activation function maps feature values ​​to the [0,1] interval, which is consistent with the range of the normalized input tensor values, ensuring that the initial enhanced image pixel values ​​have physical validity and match the quantization range of radiance.

[0082] After activation, a preliminary enhanced image is output. This image not only visually resolves the degradation issues caused by atmospheric scattering and uneven illumination in the original image, improving image clarity and contrast, but more importantly, through the full-process physical constraints of the physical information neural network model, it ensures that the radiation characteristics and texture structure of ground objects in the image conform to the physical laws of satellite imaging, without introducing artifacts that violate the atmospheric radiative transfer model, and retains the true physical attribute association information of ground objects. This provides a high-quality data foundation that combines visual effects and physical authenticity for subsequent tasks such as quantitative remote sensing inversion and scientific analysis.

[0083] refer to Figure 2 In one optional embodiment, during the training of the physical information neural network model, a pre-trained physical information neural network model is obtained by calculating the total loss function including the physical residual loss term and optimizing the network parameters through backpropagation, including the following steps:

[0084] S21. Obtain training data, which includes preprocessed satellite observation images as input, corresponding historical imaging physical condition vectors, and reference images as supervision signals; wherein, the historical imaging physical condition vectors are constructed in the same way as the imaging physical condition vectors.

[0085] Specifically, this step constructs a high-quality dataset with "input-supervision" pairings, providing reliable samples and constraints for model training. The training data contains three key types of information: preprocessed satellite observation images, corresponding historical imaging physical condition vectors, and reference images serving as supervision signals. These three elements are strictly correlated through unique identifiers for the imaging scenes, ensuring consistency between physical constraints and supervision signals during training. The preprocessed satellite observation images undergo radiometric calibration and normalization to adapt to the model's input requirements. The historical imaging physical condition vectors include solar zenith angle, solar azimuth angle, observed zenith angle, observed azimuth angle, and atmospheric optical thickness. Their construction method is consistent with imaging physical condition vectors in practical applications, accurately reflecting the imaging environment of the training samples. The reference images, serving as supervision signals, possess accurate ground reflectance information and clear details. Geometric and radiometric corrections ensure spatial and radiometric consistency with the observation images. Simultaneously, data augmentation techniques expand sample diversity and enhance the model's generalization ability.

[0086] S22. Input the preprocessed satellite observation image and imaging physical condition vector into the physical information neural network model to be trained for forward propagation to obtain the predicted image output by the model; calculate the pixel-level difference between the predicted image and the reference image to obtain the image reconstruction loss.

[0087] Specifically, the preprocessed satellite observation image and historical imaging physical condition vectors are input into the model to be trained for forward propagation, directly outputting a predicted image as an estimate of surface reflectance. During forward propagation, the model extracts multi-scale features through the encoder and fuses physical condition information, then reconstructs the predicted image through the decoder. The goal is to learn the mapping relationship from observation data to actual surface reflectance. Subsequently, the image reconstruction loss is calculated to quantify the pixel-level difference between the predicted image and the reference image, using the mean squared error loss function, with the corresponding formula being: ,in For image reconstruction loss, The total number of pixels in the image. For reference image pixel values, To predict image pixel values, this loss term ensures that the predicted image visually and structurally matches the features of real ground features.

[0088] S23. Use the predicted image as a surface reflectance estimate, input the surface reflectance estimate and the corresponding historical imaging physical condition vector into the atmospheric radiative transfer model, and calculate the resynthesized atmospheric top radiance.

[0089] Specifically, the predicted image is directly used as a surface reflectance estimate, which, along with the corresponding historical imaging physics condition vector, is input into a differentiable atmospheric radiative transfer model to calculate the resynthesized top-atmosphere radiance. Surface reflectance, as an inherent physical property of ground features, is the core input to the atmospheric radiative transfer model. Its combination with imaging physics conditions simulates the complete process of radiation propagating from the surface through the atmosphere to the satellite sensor. The final output top-atmosphere radiance has a clear physical meaning, representing the theoretical observation value of the sensor under given conditions.

[0090] S24. Calculate the difference norm between the radiance of the top atmospheric layer and the actual observed radiance in the training input, as the physical residual loss; perform a weighted summation of the image reconstruction loss and the physical residual loss to obtain the total loss function.

[0091] Specifically, the physical residual loss and the total loss function are calculated to achieve dual constraints on visual effects and physical laws. The physical residual loss is obtained by quantifying the difference between the synthesized radiance of the top atmosphere and the actual observed radiance, and is calculated using the L2 norm. The corresponding formula is as follows: ,in For physical residual loss, To synthesize the radiance of the top atmospheric layer, To ensure accurate observation of radiance, this loss term guarantees that the predicted image conforms to the physical laws of radiative transfer in satellite imaging. The total loss function is obtained by weighted summation of the image reconstruction loss and the physical residual loss, and the corresponding formula is: ,in For the total loss, , These are weighting coefficients used to balance the importance of visual fit and physical constraints.

[0092] S25. Based on the total loss function, update the parameters of the physical information neural network model to be trained using the backpropagation algorithm.

[0093] Specifically, based on the total loss function, the gradients of each model parameter are calculated using the backpropagation algorithm, and the weights and biases of the model to be trained are updated using the gradient descent optimization strategy. During backpropagation, gradient information starts from the total loss and is passed layer by layer to each layer of the model, ensuring that the optimization requirements of physical residual loss and image reconstruction loss are fed back into the parameter adjustment, so that the model continuously conforms to physical laws while improving visual effects.

[0094] S26. Repeat steps S22 to S25 until the model converges, and obtain the pre-trained physical information neural network model.

[0095] Specifically, the process from S22 to S25 is repeated to continuously iterate and train the model. Each iteration involves forward propagation, loss calculation, and parameter updates based on new parameters until the total loss value of the model converges to a preset threshold or the number of iterations reaches a set upper limit. At this point, the model has fully learned the correlation between observed data, physical conditions, and the actual surface reflectance, and outputs prediction results that are both visually plausible and physically realistic, thus obtaining a pre-trained physical information neural network model.

[0096] In an optional embodiment, the predicted image is used as a surface reflectance estimate. The surface reflectance estimate and the corresponding historical imaging physical condition vector are input into the atmospheric radiative transfer model to calculate the resynthesized top-level atmospheric radiance, including the following steps:

[0097] S31. Based on the imaging physical condition vector, obtain the atmospheric radiative transfer coefficient under the current scene by querying the lookup table.

[0098] Specifically, this step achieves rapid matching of physical parameters through a pre-built lookup table, significantly reducing real-time computational overhead while ensuring the accuracy of radiative transfer calculations. The construction of the lookup table must be based on a mature and complete atmospheric radiative transfer model (such as the 6S model or the MODTRAN model). These models can accurately simulate radiative transfer processes under different atmospheric conditions, providing a reliable physical basis for coefficient calculations. During the construction process, the coverage and resolution of the parameter grid are first determined to ensure coverage of all imaging scenarios that may occur in practical applications. For example, the solar zenith angle ranges from 0° to 80° with a step size of 5°; the solar azimuth angle ranges from 0° to 360° with a step size of 15°; the observed zenith angle ranges from 0° to 60° with a step size of 5°; the observed azimuth angle ranges from 0° to 360° with a step size of 15°; the atmospheric optical thickness (550nm band) ranges from 0.1 to 2.0 with a step size of 0.1; it also covers typical aerosol types such as continental, marine, and urban types, as well as combined scenarios with different water vapor contents (0.5~5.0g / cm²) and ozone contents (250~350DU). For each parameter combination scenario, the three-dimensional coefficient set [A(λ), B(λ), C(λ)] of the corresponding band λ is obtained by simulation calculation through a complete atmospheric radiative transfer model. All coefficient sets are stored in a multi-level index structure of "solar zenith angle - solar azimuth angle - observed zenith angle - observed azimuth angle - atmospheric optical thickness - aerosol type", forming the core data of the lookup table.

[0099] The lookup table is stored in binary file format to improve read and write efficiency, while a block storage strategy optimizes query performance: the lookup table is divided into multiple sub-blocks according to aerosol type, and each sub-block is further divided according to atmospheric optical thickness. During the query, the sub-block is first located based on the aerosol type in the imaging metadata, then the sub-block is locked by atmospheric optical thickness, and finally the coefficient groups corresponding to the remaining physical parameters are retrieved within the sub-block. In the actual query process, five core physical parameters and aerosol type information are first extracted from the imaging physical condition vector, and the parameter values ​​are matched with the parameter grid of the lookup table. Since the actual imaging parameters may not fall exactly on the grid nodes, a cubic polynomial interpolation algorithm is used to solve for the coefficients. For example, with the target parameter as the center, adjacent 4×4×4 grid nodes (corresponding to adjacent nodes of the three key parameters of solar zenith angle, observation zenith angle, and atmospheric optical thickness) are selected, and the values ​​of A(λ), B(λ), and C(λ) corresponding to the target parameter are obtained by cubic polynomial fitting. Compared to bilinear interpolation, this interpolation algorithm can better fit the nonlinear variation of coefficients with physical parameters, controlling the interpolation error within 1% and ensuring the accuracy of radiative transfer calculations. After the query is completed, the atmospheric radiative transfer coefficient corresponding to the current band λ is output, which can be directly used for subsequent radiance synthesis. This avoids the several-second delay caused by running the complete atmospheric radiative transfer model in real time, reducing the calculation time to the millisecond level and meeting the real-time requirements of model training and inference.

[0100] S32. Based on the atmospheric radiative transfer coefficient and the predicted image used as an estimate of surface reflectance, the simplified radiative transfer equation is used as the atmospheric radiative transfer model to calculate the resynthesized radiance of the top layer of the atmosphere; wherein, the simplified radiative transfer equation is expressed as:

[0101]

[0102] in, Indicates in band The calculated radiance of the top layer of the atmosphere, Indicates in band The estimated surface reflectance value of the corresponding pixel in the predicted image. , , In the band The atmospheric radiative transfer coefficient at that location is determined by the imaging physical condition vector; where, For band The product coefficient of total atmospheric transmittance and solar down-facing irradiance represents the satellite entrance pupil radiance contributed by a unit surface reflectance without considering atmospheric multiple scattering. For band The atmospheric hemispherical albedo coefficient at a given location represents the intensity of multiple scattering of upward radiation by the atmosphere and is used to characterize the surface-atmosphere coupling effect. For band The path radiance coefficient at a given location represents the radiance produced by the scattering of sunlight by atmospheric molecules and aerosols directly into the sensor, and is independent of the surface reflectance.

[0103] Specifically, a simplified radiative transfer equation is used as the atmospheric radiative transfer model. The core principle is to completely preserve the core physical mechanisms of radiative transfer while eliminating complex secondary factors, thus achieving a high-efficiency and accurate balance in radiance synthesis. In the expression of this equation, The resynthesized radiance of the top atmosphere at band λ is a theoretical physical quantity that simulates the observations of satellite sensors in this band, and its unit is consistent with the actual observed radiance. This is the estimated surface reflectance value of the pixel corresponding to the predicted image at band λ, directly taken from the predicted image output in the previous step. This value needs to be inversely normalized to restore it to the true reflectance range (0~1). The inverse normalization formula is as follows: ,in To predict the normalized pixel values ​​of an image, and These are the statistical maximum and minimum values ​​of surface reflectance for this band, ensuring that the reflectance values ​​conform to physical principles; A(λ), B(λ), and C(λ) are the band-specific atmospheric radiative transfer coefficients obtained from S31 query, all of which are wavelength-dependent parameters strongly related to imaging physical conditions.

[0104] The physical essence of A(λ) is the product of total atmospheric transmittance and downward solar irradiance. Total atmospheric transmittance includes the downward transmittance of solar radiation from the top of the atmosphere to the Earth's surface, and the upward transmittance of radiation reflected from the Earth's surface to the top of the atmosphere. The product of the two represents the transmission efficiency of radiation in the atmospheric path. Downward solar irradiance is the intensity of solar radiation reaching the Earth's surface. Therefore, the physical meaning of A(λ) is: under ideal conditions without considering atmospheric multiple scattering, the satellite entrance pupil radiance contributed by a unit of surface reflectance decreases as the solar zenith angle increases (the lower the solar altitude, the longer the atmospheric path, and the lower the transmittance), and decreases as the atmospheric optical thickness increases (the more turbid the atmosphere, the lower the transmittance). B(λ) is the atmospheric hemispherical albedo coefficient. Its physical core characterizes the multiple scattering feedback effect of atmospheric upward radiation. That is, after the radiation reflected from the Earth's surface is scattered by the atmosphere, some of the radiation returns to the Earth's surface and is reflected a second time, forming a coupled scattering process between the Earth's surface and the atmosphere. The larger the value of B(λ), the stronger this coupling effect. This coefficient mainly depends on the atmospheric optical thickness and the observation angle. The higher the aerosol content, the larger B(λ). C(λ) is the path radiative coefficient. Its sources include the direct scattering of sunlight by atmospheric molecules (Rayleigh scattering) and aerosol particles (Mie scattering). This scattered radiation enters the satellite sensor directly without being reflected by the Earth's surface, so it is independent of the Earth's surface reflectivity. The value of C(λ) increases with the increase of the solar zenith angle (the longer the atmospheric path, the more scattering opportunities) and increases significantly with the increase of atmospheric optical thickness (aerosol scattering is dominant). If there are thin clouds in the imaging scene, the contribution of cloud scattering needs to be added to C(λ). The value of C(λ) can be adjusted by the cloud amount indicator in the imaging metadata.

[0105] For multi-band satellite imagery, each band is processed separately according to the above procedure. Each band corresponds to independent A(λ), B(λ), and C(λ) coefficients (because the atmospheric transmission characteristics of different bands differ significantly; for example, the atmospheric transmittance of the shortwave infrared band is higher than that of the visible light band, while its path radiation is lower). After the composite is completed, the... Numerical verification is performed to ensure that the values ​​are within the dynamic range of the satellite sensors. For outliers that exceed the dynamic range, an adaptive truncation strategy is adopted, setting values ​​higher than the maximum value to the maximum value and values ​​lower than the minimum value to the minimum value. This ensures that the synthesized atmospheric top-layer radiance has the same physical meaning and numerical comparability as the actual observed radiance, providing a reliable basis for the subsequent calculation of physical residual loss.

[0106] In one optional embodiment, the formula for calculating the physical residual loss is:

[0107]

[0108] in, Indicates physical residual loss. This represents the total number of pixels in a training batch. For pixel index, This represents a differentiable atmospheric radiative transfer model. The parameter is The physical information neural network model to be trained Indicates the first Input data of 1 pixel, Indicates the first The imaging physical condition vector corresponding to each pixel. The model represents the first Predicted reflectance value for each pixel. Indicates the first The actual observed radiance value corresponding to each pixel.

[0109] Specifically, for each pixel in the training batch, the input data is processed by the model to be trained to obtain the corresponding surface reflectance prediction result. At the same time, the imaging physical condition vector of the pixel is combined with the input of the differentiable atmospheric radiative transfer model to complete the physical simulation from reflectance to the radiance of the top of the atmosphere, and generate synthetic radiance data with the same dimension and physical meaning as the actual observed radiance.

[0110] Subsequently, by calculating the difference between the synthesized radiance and the actual observed radiance, the degree of deviation of a single pixel at the physical level is quantified. Then, the L2 norm squared operation is performed on the difference, which can highlight the contribution weight of the larger deviation to the loss value, guide the model to prioritize correcting the prediction results that seriously violate the physical laws, and provide stable gradient information for the backpropagation process by virtue of its continuous differentiability, so as to avoid gradient oscillation affecting the parameter update efficiency.

[0111] After quantizing the individual pixel physical deviations of all pixels within a batch, the deviation results of all pixels are summed and then divided by the total number of pixels in the batch to achieve standardization. This batch averaging operation can effectively offset the interference of random deviations of a single image or a single pixel on the loss value, keeping the loss value within a reasonable range and ensuring the comparability of losses in different batches of training data, thereby improving the stability and convergence efficiency of model training.

[0112] The core design logic of the entire formula is to tightly bind the prediction results of the neural network with real physical observations through mathematical operations. Using the atmospheric radiative transfer model as a bridge, the abstract reflectance prediction value is transformed into a physically meaningful simulated radiance value. Then, through deviation quantization and summarization operations, the physical consistency requirement is transformed into a loss signal that the model can optimize. This design not only ensures the precision of loss calculation, constraining physical deviations pixel by pixel, but also achieves the stability of the training process through batch averaging. Ultimately, it forces the model to strictly follow the physical laws of radiative transfer while learning image enhancement mapping relationships, avoiding the generation of artifacts that violate common sense, and ensuring that the output results have both visual appeal and physical realism.

[0113] The aforementioned image processing method based on satellite communication networks constructs a complete automated processing chain from receiving the raw byte stream from the satellite communication network to outputting an enhanced and standardized image. This chain employs a circular buffer to smooth network fluctuations, structured parsing and verification based on start identifiers and key fields to handle packet loss and out-of-order delivery while ensuring data integrity, dynamic decoding based on data format identifiers to ensure compatibility with standard and non-standard image data, and inputting image data and imaging metadata into a pre-trained physical information neural network model for forward propagation calculations. The key innovation of this network model lies in its training process incorporating physical residual loss calculated based on a differentiable atmospheric radiative transfer model as a constraint. This allows for strict adherence to the physical laws of satellite imaging while improving image visual quality. Finally, the initially enhanced image output by the model is standardized, encoded, and encapsulated for multicast forwarding. Ultimately, this achieves the technical effect of reliably receiving and orderly reassembling image data on unreliable satellite links, flexibly decoding various proprietary formats, and generating high-quality standardized image data with superior visual effects and greater physical accuracy for distribution to multiple users.

[0114] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0115] Example 2:

[0116] Based on the same inventive concept, this application also provides a system for implementing the image processing method based on a satellite communication network as described above. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the image processing system based on a satellite communication network provided below can be found in the limitations of the image processing method based on a satellite communication network described above, and will not be repeated here.

[0117] In one exemplary embodiment, such as Figure 3 As shown, an image processing system 30 based on a satellite communication network is provided to implement the methods in the above-described method embodiments. The system includes:

[0118] The data stream preprocessing module 31 is used to write the raw byte stream from the satellite communication network into a circular buffer to obtain buffered byte data; by scanning the buffered byte data based on the start identifier, the data packet header field is extracted; wherein, the data packet header field includes the data packet sequence number, data payload length, timestamp, camera type identifier, data format identifier and check code.

[0119] The data packet verification and reassembly module 32 is used to perform integrity verification on the corresponding data payload based on the check code in the data packet header field to obtain a data packet that has passed verification; and to sort and reassemble the data packets based on the data packet sequence number and timestamp in the data packet that has passed verification to obtain an ordered sequence of valid image data packets.

[0120] The image decoding and encapsulation module 33 is used to call the corresponding decoding library to decode the data payload according to the data format identifier of each data packet in the ordered valid image data packet sequence, so as to obtain the decoded raw pixel data; and bind and encapsulate the raw pixel data with the timestamp, camera type identifier and associated imaging metadata in the corresponding data packet header field to obtain a standardized image data object.

[0121] The physical enhancement processing module 34 is used to input image data objects into a pre-trained physical information neural network model, and obtain a preliminary enhanced image through forward propagation calculation of the physical information neural network model. In the training process of the physical information neural network model, the network parameters are optimized by calculating the total loss function including the physical residual loss term and backpropagating, and the pre-trained physical information neural network model is obtained. The physical residual loss term is calculated based on a differentiable atmospheric radiative transfer model and is used to constrain the network output to conform to the physical laws of satellite imaging.

[0122] The image transmission optimization module 35 is used to standardize the encoding of the preliminarily enhanced image to generate a standard format image stream; the standard format image stream is encapsulated into network data packets according to the real-time transmission protocol, and the network data packets are sent to the predetermined receiving end via multicast.

[0123] Example 3:

[0124] Based on the same inventive concept, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the aforementioned method embodiments.

[0125] Example 4:

[0126] Based on the same inventive concept, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0127] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and 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 modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0128] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. An image processing method based on a satellite communication network, characterized in that, The method includes: S1. Write the raw byte stream from the satellite communication network into a circular buffer to obtain buffered byte data; extract the data packet header field by scanning the buffered byte data based on the start identifier; wherein, the data packet header field includes data packet sequence number, data payload length, timestamp, camera type identifier, data format identifier and checksum; S2. Perform integrity verification on the corresponding data payload based on the checksum in the data packet header field to obtain a verified data packet; sort and reassemble the data packet sequence number and timestamp in the verified data packet to obtain an ordered sequence of valid image data packets. S3. Based on the data format identifier of each data packet in the ordered valid image data packet sequence, call the corresponding decoding library to decode the data payload to obtain the decoded raw pixel data; bind and encapsulate the raw pixel data with the timestamp, camera type identifier and associated imaging metadata in the corresponding data packet header field to obtain a standardized image data object; S4. Input the image data object into a pre-trained physical information neural network model, and obtain a preliminary enhanced image through the forward propagation calculation of the physical information neural network model; S5. Standardize and encode the preliminarily enhanced image to generate a standard format image stream; encapsulate the standard format image stream into network data packets according to a real-time transmission protocol, and send the network data packets to a predetermined receiving end via multicast; The step of inputting the image data object into a pre-trained physical information neural network model and obtaining a preliminary enhanced image through forward propagation calculation of the physical information neural network model includes: During the training process of the physical information neural network model, the pre-trained physical information neural network model is obtained by calculating the total loss function including the physical residual loss term and optimizing the network parameters through backpropagation. The physical residual loss term is calculated based on a differentiable atmospheric radiative transfer model and is used to constrain the network output to conform to the physical laws of satellite imaging. The training process of the physical information neural network model involves calculating a total loss function that includes a physical residual loss term and optimizing the network parameters through backpropagation to obtain a pre-trained physical information neural network model, including: S21. Acquire training data, which includes preprocessed satellite observation images as input, corresponding historical imaging physical condition vectors, and reference images as supervision signals; wherein the historical imaging physical condition vectors are constructed in the same way as the imaging physical condition vectors. S22. Input the preprocessed satellite observation image and the imaging physical condition vector into the physical information neural network model to be trained for forward propagation to obtain the predicted image output by the model; calculate the pixel-level difference between the predicted image and the reference image to obtain the image reconstruction loss; S23. Use the predicted image as a surface reflectance estimate, input the surface reflectance estimate and the corresponding historical imaging physical condition vector into the atmospheric radiative transfer model, and calculate the resynthesized atmospheric top radiance. S24. Calculate the difference norm between the atmospheric top layer radiance and the actual observed radiance in the training input, and use it as the physical residual loss; perform a weighted summation of the image reconstruction loss and the physical residual loss to obtain the total loss function; S25. Based on the total loss function, update the parameters of the physical information neural network model to be trained using the backpropagation algorithm; S26. Repeat steps S22 to S25 until the model converges to obtain the pre-trained physical information neural network model. The step of using the predicted image as a surface reflectance estimate, inputting the surface reflectance estimate and the corresponding historical imaging physical condition vector into the atmospheric radiative transfer model, and calculating the resynthesized atmospheric top-layer radiance includes: S31. Based on the imaging physical condition vector, obtain the atmospheric radiative transfer coefficient of the current scene by querying a lookup table; S32. Based on the atmospheric radiative transfer coefficient and the predicted image used as an estimate of surface reflectance, the resynthesized atmospheric top-layer radiance is calculated using a simplified radiative transfer equation as the atmospheric radiative transfer model; wherein the simplified radiative transfer equation is expressed as: in, Indicates in band The calculated radiance of the top layer of the atmosphere, Indicates in band The estimated surface reflectance value of the corresponding pixel in the predicted image. , , In the band The atmospheric radiative transfer coefficient at that location is determined by the imaging physical condition vector; wherein, For band The product coefficient of total atmospheric transmittance and solar down-facing irradiance represents the satellite entrance pupil radiance contributed by a unit surface reflectance without considering atmospheric multiple scattering. For band The atmospheric hemispherical albedo coefficient at a given location represents the intensity of multiple scattering of upward radiation by the atmosphere and is used to characterize the surface-atmosphere coupling effect. For band The path radiance coefficient at a given location represents the radiance produced by the scattering of sunlight by atmospheric molecules and aerosols directly into the sensor, and is independent of the surface reflectance.

2. The method according to claim 1, characterized in that, The step of inputting the image data object into a pre-trained physical information neural network model and obtaining a preliminary enhanced image through forward propagation calculation of the physical information neural network model includes: S11. Based on the imaging metadata associated with the image data object, extract the parameter combination including solar zenith angle, solar azimuth angle, observation zenith angle, observation azimuth angle and atmospheric optical thickness parameters, and use the parameter combination to construct an imaging physical condition vector to describe the imaging scene. S12. Perform radiometric calibration and normalization preprocessing on the pixel data in the image data object to obtain the normalized network input tensor. S13. Input the network input tensor and the imaging physical condition vector as input data into the encoder of the physical information neural network model, and extract a multi-scale feature map set through multi-layer convolution and downsampling operations of the encoder; wherein, the multi-scale feature map set contains feature maps of multiple scales from shallow to deep. S14. The deepest feature map in the multi-scale feature map set is used as the initial input feature map of the decoder. The initial input feature map of the decoder is upsampled to obtain an upsampled feature map. The upsampled feature map is then connected to the encoder feature map of the corresponding scale in the multi-scale feature map set via skip connections and channel concatenation to obtain a concatenated feature map. S15. Perform convolution processing on the stitched feature map to obtain a fused feature map; use the fused feature map as the input of the decoder, and repeatedly perform the upsampling operation, the skip connection, and the channel stitching in sequence until the output feature map is restored to the original image spatial resolution to obtain the final decoded feature map; perform convolution and activation processing on the final decoded feature map to output the preliminary enhanced image.

3. The method according to claim 1, characterized in that, The formula for calculating the physical residual loss is as follows: in, This represents the physical residual loss. This represents the total number of pixels in a training batch. For pixel index, The atmospheric radiative transfer model is described as differentiable. The parameter is The physical information neural network model to be trained, Indicates the first Input data of 1 pixel, Indicates the first The imaging physical condition vector corresponding to each pixel. The model represents the first Predicted reflectance value for each pixel. Indicates the first The actual observed radiance value corresponding to each pixel.

4. An image processing system based on a satellite communication network, used to implement the method according to any one of claims 1 to 3, characterized in that, The system includes: The data stream preprocessing module is used to write the raw byte stream from the satellite communication network into a circular buffer to obtain buffered byte data; by scanning the buffered byte data based on the start identifier, the data packet header field is extracted; wherein, the data packet header field includes data packet sequence number, data payload length, timestamp, camera type identifier, data format identifier and check code; The data packet verification and reassembly module is used to perform integrity verification on the corresponding data payload based on the check code in the data packet header field to obtain a verified data packet; and to sort and reassemble the data packet sequence number and timestamp in the verified data packet to obtain an ordered sequence of valid image data packets. The image decoding and encapsulation module is used to call the corresponding decoding library to decode the data payload according to the data format identifier of each data packet in the ordered valid image data packet sequence, so as to obtain the decoded raw pixel data; and to bind and encapsulate the raw pixel data with the timestamp, camera type identifier and associated imaging metadata in the corresponding data packet header field to obtain a standardized image data object. The physical enhancement processing module is used to input the image data object into a pre-trained physical information neural network model, and obtain a preliminary enhanced image through forward propagation calculation of the physical information neural network model. During the training process of the physical information neural network model, the network parameters are optimized by calculating the total loss function including the physical residual loss term and backpropagating, resulting in the pre-trained physical information neural network model. The physical residual loss term is calculated based on a differentiable atmospheric radiative transfer model and is used to constrain the network output to conform to the physical laws of satellite imaging. The image transmission optimization module is used to standardize and encode the preliminarily enhanced image to generate a standard format image stream; encapsulate the standard format image stream into network data packets according to a real-time transmission protocol, and send the network data packets to a predetermined receiving end via multicast.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 3.

Images