Semantic signal recovery method, electronic device, and readable storage medium

Abstract

This application provides a semantic signal recovery method, an electronic device, and a readable storage medium, belonging to the fields of wireless communication and artificial intelligence technology. The method includes: performing preliminary noise prediction on a distorted semantic signal based on the current noise propagation time step to obtain a preliminary noise component; performing noise reduction estimation on the distorted semantic signal based on the preliminary noise component to obtain an estimated clean semantic signal; performing semantic fuzzing processing on the estimated clean semantic signal to obtain a fuzzy clean semantic signal; adding noise to the fuzzy clean semantic signal based on the preliminary noise component to generate fuzzy adversarial samples; performing sample noise prediction on the fuzzy adversarial samples based on the current noise propagation time step to obtain adversarial sample noise components; and performing skip-denoising on the distorted semantic signal based on the preliminary noise component and the adversarial sample noise components to obtain a target recovery signal. This application embodiment can effectively improve semantic communication efficiency in scenarios lacking prior channel information.

Description

Technical Field

[0001] This application relates to the fields of wireless communication and artificial intelligence technologies, and in particular to a semantic signal recovery method, an electronic device, and a readable storage medium. Background Technology

[0002] Currently, with the rapid growth of IoT devices, communication systems are evolving from "bit-level reliable transmission" to "semantic-level efficient transmission." Traditional semantic signal recovery methods are typically based on generative adversarial network (GAN) architectures. Specifically, they acquire the distorted semantic signal received at the receiver of the semantic communication system after transmission through the channel, and use a generator in the GAN to denoise the distorted signal to generate clean semantic features. Further, a discriminator network in the GAN distinguishes between true and false semantic features, and updates the generator parameters based on the discrimination results. Through adversarial training between the generator and the discriminator, the generator gradually learns the mapping relationship from the distorted semantic signal to the clean semantic signal, ultimately outputting the recovered semantic signal. However, due to inherent defects such as poor training stability and susceptibility to pattern collapse, the quality of the recovered semantic signal is unstable, making it difficult to guarantee the reliability of semantic communication. Therefore, obtaining a communication semantic signal with a stable training process and reliable generation quality has become a challenging research problem.

[0003] To address this issue, existing technologies employ semantic signal recovery methods based on diffusion models. Specifically, these methods construct external conditional inputs based on communication channel state information, embedding the distorted semantic signal and the external conditions together into a noise prediction network within the diffusion model to estimate noise components. Then, through sequential iteration following the Markov chain assumption and using the noise components, the distorted semantic signal is progressively denoising to recover the original semantic signal. However, these existing methods require prior information from the external environment, such as channel state information, as guiding conditions for noise sample generation. In complex and dynamic wireless communication environments, this not only makes it difficult to obtain accurate channel state information in real time but also significantly increases communication overhead. Furthermore, the denoising generation process strictly follows the Markov chain assumption, requiring numerous sequential iterative steps to complete semantic signal recovery, resulting in high semantic communication latency. Therefore, how to effectively improve semantic communication efficiency in scenarios lacking prior channel information has become an urgent problem to solve. Summary of the Invention

[0004] The main objective of this application is to propose a semantic signal recovery method, an electronic device, and a readable storage medium, which aims to effectively improve semantic communication efficiency in scenarios where prior channel information is lacking.

[0005] To achieve the above objectives, a first aspect of this application proposes a semantic signal recovery method, applied to the receiving end of a semantic communication system, the method comprising: Obtain the current noise diffusion time step and the corresponding distortion semantic signal; wherein, the distortion semantic signal is a noisy semantic signal; Based on the current noise propagation time step, a preliminary noise prediction is performed on the distorted semantic signal to obtain a preliminary noise component; Based on the preliminary noise components, the distorted semantic signal is denoised and estimated to obtain the estimated clean semantic signal. The estimated pure semantic signal is subjected to semantic fuzzing to obtain a fuzzy pure semantic signal; wherein, the fuzzy pure semantic signal is used to provide internal guidance for the generation of adversarial examples of the distorted semantic signal; The fuzzy pure semantic signal is noise-added based on the initial noise component to generate a fuzzy adversarial sample corresponding to the distorted semantic signal. Based on the current noise diffusion time step, the noise component of the fuzzy adversarial sample is predicted. The distorted semantic signal is subjected to skip-denoising processing based on the initial noise component and the adversarial sample noise component to obtain the target recovery signal.

[0006] In some embodiments, performing semantic fuzzing on the estimated pure semantic signal to obtain a fuzzy pure semantic signal includes: Attention scoring is performed on the distorted semantic signal to obtain a signal attention score; The distorted semantic signal is semantically masked based on the signal attention score and a preset score threshold to obtain a semantic signal mask matrix; The estimated pure semantic signal is semantically blurred based on the semantic signal mask matrix to obtain a blurred pure semantic signal.

[0007] In some embodiments, the semantic signal mask matrix includes semantic mask values ​​where the signal attention score is greater than the preset score threshold; the step of performing semantic fuzzing processing on the estimated clean semantic signal based on the semantic signal mask matrix to obtain a fuzzy clean semantic signal includes: Based on the semantic mask value, the estimated pure semantic signal is filtered to obtain a pure semantic signal that is retained. According to a preset Gaussian filter, the remaining signals in the estimated pure semantic signal other than the retained pure semantic signal are subjected to Gaussian blurring to obtain the blurred remaining pure semantic signal. The fuzzy residual pure semantic signal and the retained pure semantic signal are fused to obtain the fuzzy pure semantic signal.

[0008] In some embodiments, the step of performing skip denoising processing on the distorted semantic signal based on the preliminary noise component and the adversarial example noise component to obtain the target recovery signal includes: The target noise is determined based on the preliminary noise component and the adversarial sample noise component; The distorted semantic signal is subjected to skip-denoising processing based on the target noise to obtain the target recovery signal.

[0009] In some embodiments, determining the target noise based on the preliminary noise component and the adversarial example noise component includes: Obtain the guiding scale parameter of the fuzzy pure semantic signal; wherein, the guiding scale parameter is used to characterize the credibility of the semantic structure of the fuzzy pure semantic signal; Based on the adversarial sample noise component and the preliminary noise component, component calculation is performed to obtain the noise difference value; The target noise is obtained by performing noise correction calculations based on the noise difference value and the guiding scale parameter.

[0010] In some embodiments, the step of performing skip denoising processing on the distorted semantic signal based on the target noise to obtain the target recovered signal includes: Determine the initial jump denoising time step; wherein the initial jump denoising time step is earlier than the current noise diffusion time step, and the time interval between the initial jump denoising time step and the current noise diffusion time step is greater than 1. Based on the initial jump denoising time step and the target noise, the distorted semantic signal is subjected to first jump denoising processing to obtain the first denoised signal corresponding to the initial jump denoising time step. The initial jump denoising time step is updated to obtain the target jump denoising time step; wherein, the time of the target jump denoising time step is earlier than the initial jump denoising time step. The first denoised signal is subjected to second jump denoising processing based on the target jump denoising time step and the target noise corresponding to the target jump denoising time step to obtain the target recovered signal corresponding to the target jump denoising time step.

[0011] In some embodiments, the step of performing a first jump denoising process on the distorted semantic signal based on the initial jump denoising time step and the target noise to obtain a first denoised signal corresponding to the initial jump denoising time step includes: Obtain the current cumulative attenuation coefficient corresponding to the current noise diffusion time step, and obtain the initial cumulative attenuation coefficient corresponding to the initial jump denoising time step; The denoised signal quantity is calculated based on the current cumulative attenuation coefficient, the distorted semantic signal, and the target noise. Based on the initial cumulative attenuation coefficient and the target noise, a noise compensation term is calculated. The first denoised signal is obtained by calculation based on the noise compensation term and the denoised signal quantity.

[0012] In some embodiments, the step of performing preliminary noise prediction on the distorted semantic signal based on the current noise diffusion time step to obtain preliminary noise components includes: Sine embedding is performed on the current noise diffusion time step to obtain the current time step features; Position encoding is performed on the distorted semantic signal to obtain distorted semantic signal features; The distorted semantic signal features and the current time step features are jointly encoded to obtain the fused signal time sequence features; Attention noise prediction is performed based on the temporal characteristics of the fused signal to obtain the preliminary noise component.

[0013] To achieve the above objectives, a second aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0014] To achieve the above objectives, a third aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method of the first aspect described above.

[0015] The semantic signal recovery method, electronic device, and readable storage medium proposed in this application acquire the current noise diffusion time step received by the receiver and its corresponding distorted semantic signal, and perform preliminary noise prediction on the distorted semantic signal based on the current noise diffusion time step to obtain a preliminary noise component; perform noise reduction estimation on the distorted semantic signal based on the preliminary noise component to obtain an estimated clean semantic signal; perform semantic fuzzing processing on the estimated clean semantic signal to obtain a fuzzy clean semantic signal; perform noise addition processing on the fuzzy clean semantic signal based on the preliminary noise component to generate a fuzzy adversarial sample corresponding to the distorted semantic signal; perform sample noise prediction on the fuzzy adversarial sample based on the current noise diffusion time step to obtain an adversarial sample noise component; and perform jump denoising processing on the distorted semantic signal based on the preliminary noise component and the adversarial sample noise component to obtain the target recovered signal. In this application, for complex dynamic wireless communication scenarios lacking prior channel information, an internal guiding condition is constructed by estimating a clean semantic signal derived from the distorted semantic signal itself and performing semantic fuzzing processing. This replaces the reliance on external channel state information, avoiding the increased communication overhead caused by the difficulty in obtaining accurate and real-time prior channel information in existing technologies. Simultaneously, by performing skip denoising processing based on preliminary noise components and adversarial sample noise components, the sequential iteration limitation of traditional diffusion models strictly following the Markov chain assumption is effectively solved, significantly reducing the number of iteration steps required for denoising. This reduces the inference latency of semantic communication and effectively improves the efficiency of semantic communication in scenarios lacking prior channel information. Attached Figure Description

[0016] Figure 1 This is a flowchart of the semantic signal recovery method provided in the embodiments of this application; Figure 2 yes Figure 1 The flowchart of step S102 in the document; Figure 3 yes Figure 1 The flowchart of step S104 in the process; Figure 4 yes Figure 3 The flowchart of step S303 in the process; Figure 5 yes Figure 1 The flowchart of step S107 in the process; Figure 6 yes Figure 5 The flowchart of step S501 in the text; Figure 7 yes Figure 5 The flowchart of step S502 in the document; Figure 8 yes Figure 7 The flowchart of step S702 in the process; Figure 9This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0017] 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.

[0018] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit the scope of this application.

[0020] First, let's analyze some of the terms used in this application: Artificial intelligence (AI) is a new branch of computer science that studies, develops, and applies theories, methods, technologies, and systems to simulate, extend, and expand human intelligence. It aims to understand the essence of intelligence and produce intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thought. Furthermore, AI utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to achieve optimal results.

[0021] This application provides a semantic signal recovery method, an electronic device, and a readable storage medium, aiming to effectively improve semantic communication efficiency in scenarios lacking prior channel information.

[0022] The semantic signal recovery method, electronic device, and readable storage medium provided in this application are specifically described through the following embodiments. First, the semantic signal recovery method in this application is described.

[0023] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0024] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0025] The semantic signal recovery method provided in this application relates to the fields of wireless communication and artificial intelligence technologies. This semantic signal recovery method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the semantic signal recovery method, but is not limited to the above forms.

[0026] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0027] Figure 1This is an optional flowchart of the semantic signal recovery method provided in this application embodiment, applied to the receiving end of a semantic communication system. Figure 1 The method may include, but is not limited to, steps S101 to S107.

[0028] Step S101: Obtain the current noise diffusion time step and the corresponding distortion semantic signal; wherein, the distortion semantic signal is a noisy semantic signal.

[0029] Step S102: Perform preliminary noise prediction on the distorted semantic signal based on the current noise diffusion time step to obtain preliminary noise components.

[0030] Step S103: Based on the preliminary noise components, perform noise reduction estimation on the distorted semantic signal to obtain the estimated clean semantic signal.

[0031] Step S104: Perform semantic fuzzing on the estimated pure semantic signal to obtain a fuzzy pure semantic signal; wherein, the fuzzy pure semantic signal is used to provide internal guidance for the generation of adversarial examples for the distorted semantic signal.

[0032] Step S105: Add noise to the fuzzy pure semantic signal based on the initial noise components to generate fuzzy adversarial samples corresponding to the distorted semantic signal.

[0033] Step S106: Based on the current noise diffusion time step, perform sample noise prediction on the fuzzy adversarial sample to obtain the adversarial sample noise component.

[0034] Step S107: Perform skip denoising on the distorted semantic signal based on the initial noise component and the adversarial sample noise component to obtain the target recovery signal.

[0035] Steps S101 to S107 as shown in the embodiments of this application involve acquiring the current noise propagation time step received by the receiving end and its corresponding distorted semantic signal, performing preliminary noise prediction on the distorted semantic signal based on the current noise propagation time step to obtain a preliminary noise component; performing noise reduction estimation on the distorted semantic signal based on the preliminary noise component to obtain an estimated clean semantic signal; performing semantic fuzzing processing on the estimated clean semantic signal to obtain a fuzzy clean semantic signal; performing noise addition processing on the fuzzy clean semantic signal based on the preliminary noise component to generate a fuzzy adversarial sample corresponding to the distorted semantic signal; performing sample noise prediction on the fuzzy adversarial sample based on the current noise propagation time step to obtain an adversarial sample noise component; and performing jump denoising processing on the distorted semantic signal based on the preliminary noise component and the adversarial sample noise component to obtain the target recovery signal. In this application, for complex dynamic wireless communication scenarios lacking prior channel information, an internal guiding condition is constructed by estimating a clean semantic signal derived from the distorted semantic signal itself and performing semantic fuzzing processing. This replaces the reliance on external channel state information, avoiding the increased communication overhead caused by the difficulty in obtaining accurate and real-time prior channel information in existing technologies. Simultaneously, by performing skip denoising processing based on preliminary noise components and adversarial sample noise components, the sequential iteration limitation of traditional diffusion models strictly following the Markov chain assumption is effectively solved, significantly reducing the number of iteration steps required for denoising. This reduces the inference latency of semantic communication and effectively improves the efficiency of semantic communication in scenarios lacking prior channel information.

[0036] In step S101 of some embodiments, specifically, the semantic communication system is a communication architecture that takes semantic information as the main transmission object and aims to achieve efficient semantic-level information exchange. The system includes a transmitter and a receiver. The transmitter is used to extract semantic features and compress and encode the original source data to generate a semantic signal and send it to the channel. The receiver is used to receive the distorted semantic signal transmitted through the channel and perform semantic signal recovery processing to reconstruct the original semantic information.

[0037] Specifically, the current noise diffusion time step refers to the discrete-time index parameter in the noise addition process of the original semantic information. This time step is used to characterize the intensity level of the noise contained in the distorted semantic signal.

[0038] Specifically, distorted semantic signals refer to noisy semantic signals that have degraded due to interference from factors such as channel noise and fading after the original semantic information has been transmitted through a wireless channel.

[0039] Specifically, the receiver can obtain the current noise diffusion time step and the corresponding distorted semantic signal after the original semantic information has undergone forward noise addition processing using a diffusion model, which is transmitted by the transmitter via a wireless channel.

[0040] For example, in a semantic communication scenario, if the sender intends to transmit the semantic information "There is a traffic accident ahead, please slow down", this information is first encoded into a semantic feature vector. After being transmitted through a wireless channel, the distorted semantic signal received by the receiver may contain interference components introduced by channel noise, making the originally clear semantic features blurry, and some semantic details (such as slowing down) may be masked by noise.

[0041] In this embodiment, by acquiring the current noise propagation time step and the corresponding distortion semantic signal received by the receiver, a quantitative characterization of the received semantic signal state is achieved, providing data support for subsequent noise prediction and signal recovery.

[0042] Please see Figure 2 In some embodiments, step S102 includes, but is not limited to, steps S201 to S204: Step S201: Perform position encoding on the current noise diffusion time step to obtain the current time step features.

[0043] Step S202: Position encoding is performed on the distorted semantic signal to obtain the distorted semantic signal features.

[0044] Step S203: Jointly encode the distorted semantic signal features and the current time step features to obtain the fused signal time sequence features.

[0045] Step S204: Based on the temporal characteristics of the fused signal, perform attention noise prediction to obtain preliminary noise components.

[0046] In step S201 of some embodiments, specifically, the current time step feature refers to the vector representation of the current noise diffusion time step.

[0047] Furthermore, the current noise diffusion can be encoded using sine or cosine position coding. Specifically, for the dimension index of the current noise diffusion time step, if the dimension index is even, sine position coding is performed using a sine function; if the dimension index is odd, cosine position coding is performed using a cosine function. This converts the absolute position information of the current noise diffusion time step into a continuous and periodic vector representation.

[0048] In step S202 of some embodiments, specifically, the distorted semantic signal feature refers to the semantic location information vector representation of the distorted semantic signal.

[0049] Specifically, word embedding representation can be performed on the distorted semantic signal to obtain the distorted signal vector. The dimension D of this vector can be obtained, and a position vector of dimension D can be generated. The position vector is then added element by element to the distorted signal vector at the corresponding position to obtain the distorted semantic signal feature with position information.

[0050] In step S203 of some embodiments, specifically, the fused signal timing features refer to the vector representation that integrates the distorted semantic signal features and the current time step features.

[0051] Specifically, the current time step features can be aligned based on the distorted semantic signal features to obtain aligned time step features with the same feature dimensions as the distorted semantic signal features. The aligned time step features and the distorted semantic signal features are then concatenated to obtain the fused signal temporal features.

[0052] In step S204 of some embodiments, specifically, the preliminary noise component refers to the noise vector representation predicted for the distorted semantic signal, the noise dimension being the same as the characteristics of the distorted semantic signal, used to represent the noise component predicted from the distorted semantic signal.

[0053] Specifically, noise prediction can be achieved by using a pre-trained semantic signal recovery model to analyze the temporal features of the fused signal. This model is a neural network composed of a diffusion model and a Transformer model. The diffusion model is used to iteratively denoise the distorted semantic signal, while the Transformer model is used to predict the noise components of the distorted semantic signal at different diffusion noise time steps and to construct the semantic mask matrix of the distorted semantic signal.

[0054] The training loss function of the preset semantic signal recovery model is the binary cross-entropy loss function, which can be expressed as:

[0055] in, Indicates the loss value. This represents the total number of distorted semantic signal samples within a batch. Indicates the first The true label of each distorted semantic signal sample takes a value of 0 or 1. Indicates the first The probability that a distorted semantic signal sample is predicted to be the true label.

[0056] Furthermore, if the loss value meets the preset training conditions, the preset semantic signal recovery model is determined as the pre-trained semantic signal recovery model. If the loss value does not meet the preset training conditions, the parameters of the preset semantic signal recovery model are adjusted according to the loss value through the backpropagation algorithm until the loss value meets the preset training conditions, so as to obtain the pre-trained semantic signal recovery model.

[0057] Specifically, the preset training condition is the convergence criterion of the semantic signal recovery model. If the loss value is less than the preset threshold (such as 0.2), the semantic signal recovery model will stop training.

[0058] Furthermore, the temporal features of the fused signal can be input into the noise prediction network in the Transformer model. This network generates the query matrix, key matrix, and value matrix of the fused signal temporal features through three linear transformations via a multi-head self-attention layer. The query matrix and the transpose of the key matrix are multiplied to obtain the attention score matrix, which is then normalized using the Softmax function to obtain the attention weight matrix. The attention weight matrix is ​​multiplied by the value matrix to obtain the attention aggregation feature. The attention aggregation feature is then normalized to obtain the standardized aggregation feature, and a linear function (such as the Softmax function) is applied to activate the standardized aggregation feature to predict the initial noise component with the same dimension as the distorted semantic signal features.

[0059] In this embodiment, attention noise prediction is performed based on the temporal features of the fused signal. The temporal features of the fused signal can effectively compensate for the lack of sequence awareness in the Transformer model itself, and inject necessary temporal and structural priors into the semantic feature modeling of the Transformer model, effectively enhancing the Transformer model's ability to model noise interference and semantic structural distortion.

[0060] Through steps S201 to S204, sinusoidal embedding ensures the consistency and stability of time step features, positional encoding preserves the spatial structure of the distorted semantic signal, joint encoding realizes the organic fusion of temporal and spatial information, and the attention mechanism effectively focuses on the key semantic information of the distorted semantic signal, enabling the noise prediction network to more comprehensively understand the content, structure, and noise stage of the distorted semantic signal, thereby obtaining more accurate preliminary noise components and laying a solid data foundation for subsequent noise reduction estimation, fuzzy adversarial sample generation, and skip denoising.

[0061] In step S103 of some embodiments, specifically, estimating the clean semantic signal refers to an approximate estimate of the clean semantic signal obtained after removing noise from the distorted semantic signal at the current time step based on the preliminary noise components. Ideally, this signal should be close to the clean original semantic information transmitted by the sending end, and is used for subsequent semantic ambiguity processing and adversarial sample generation.

[0062] Specifically, the estimated pure semantic signal can be determined using the following formula:

[0063] in, This represents the estimated clean semantic signal corresponding to the current noise propagation time step t. This represents the distorted semantic signal corresponding to the current noise propagation time step t. This represents the current cumulative attenuation coefficient at the current noise propagation time step t, used to characterize the cumulative retention rate of the original semantic information from the noise propagation start time step to the current time step. This represents the initial noise component corresponding to the current noise diffusion time step t.

[0064] Furthermore, the current cumulative attenuation coefficient is ,in, This represents the instantaneous attenuation coefficient corresponding to the current noise propagation time step t, used to characterize the degree of signal attenuation within a single noise propagation time step. This represents the variance scheduling parameter corresponding to the current noise diffusion time step t, used to determine the degree of Gaussian noise added to the original semantic information by the diffusion model at each time step. The instantaneous decay coefficient is represented by time step index l, and T represents the total number of sampling steps in the diffusion model.

[0065] In this embodiment, preliminary noise prediction is performed on the distorted semantic signal based on the current noise diffusion time step. This generates a data foundation that retains the overall structural information of the original semantics while removing the main noise components at the current noise diffusion time step. This provides a semantically consistent and less noisy data foundation for subsequent semantic fuzziness processing and adversarial sample generation.

[0066] Please see Figure 3 In some embodiments, step S104 includes, but is not limited to, steps S301 to S303: Step S301: Perform attention scoring on the distorted semantic signal to obtain the signal attention score.

[0067] Step S302: Perform semantic masking on the distorted semantic signal based on the signal attention score and a preset score threshold to obtain a semantic signal mask matrix.

[0068] Step S303: Perform semantic fuzzing on the estimated pure semantic signal based on the semantic signal mask matrix to obtain the fuzzy pure semantic signal.

[0069] In step S301 of some embodiments, specifically, the signal attention score refers to the quantification value of the semantic importance of each semantic unit in the distorted semantic signal in the subsequent Gaussian blurring process. The higher the score, the more important the semantic unit is.

[0070] For example, if the distorted semantic signal is text information, the semantic unit can be the text semantic vector representation corresponding to the text segmentation; if the distorted semantic signal is image information, the semantic unit can be the image semantic vector representation corresponding to the segmented image block; if the distorted semantic signal is audio information, the semantic unit can be the audio semantic vector representation corresponding to the audio frame.

[0071] Specifically, the distorted semantic signal features corresponding to the distorted semantic signal can be input into the Transformer to perform self-attention calculation on the semantic units at each position in the distorted semantic signal features, thereby obtaining the semantic unit attention score matrix at each position, i.e., the signal attention score.

[0072] In step S302 of some embodiments, specifically, the preset score threshold refers to a pre-set score boundary value used to distinguish between key semantic signals and non-key semantic signals.

[0073] Specifically, the semantic signal mask matrix is ​​a binary matrix generated according to the importance of semantic units. Each element in the matrix is ​​used to indicate whether the semantic unit at the corresponding position should be masked. A mask value of 0 indicates that the position is a critical semantic unit that needs to be selectively masked, and a mask value of 1 indicates that the position is a non-critical semantic unit that needs to be retained.

[0074] For example, the preset score threshold can be set as the statistical mean of multiple attention scores to ensure that the masking strategy has stability and adaptability at different denoising time steps, such as 0.5.

[0075] Specifically, the attention score of each semantic unit can be compared with a preset score threshold. If the attention score is greater than the threshold, the mask value corresponding to the semantic unit at that position is set to 0, indicating that the semantic unit is a key semantic unit and will be selectively masked. If the attention score is less than or equal to the threshold, the mask value corresponding to the semantic unit at that position is set to 1, indicating that the semantic unit is a non-key semantic unit and will be retained. The mask values ​​of all semantic units at all positions constitute a semantic signal mask matrix.

[0076] For example, if the distorted semantic signal is a feature sequence of the semantic content "There is a traffic accident ahead, please slow down", after attention scoring, key semantic units such as "traffic accident" and "slow down" will receive higher attention scores, while semantic units such as "ahead" and "road" will receive lower attention scores. Then, the mask value corresponding to "traffic accident" and "slow down" with attention scores higher than the preset score threshold can be set to 0, and the mask value corresponding to "ahead" and "road" with attention scores lower than the preset score threshold can be set to 1, so as to generate a semantic signal mask matrix.

[0077] Please see Figure 4 In some embodiments, the semantic signal mask matrix includes semantic mask values ​​where the signal attention score is greater than a preset score threshold. Step S303 includes, but is not limited to, steps S401 to S403: Step S401: Based on the semantic mask value, the estimated pure semantic signal is filtered to obtain the retained pure semantic signal.

[0078] Step S402: According to the preset Gaussian filter, the remaining signals in the estimated pure semantic signal except for the pure semantic signal are subjected to Gaussian blurring to obtain the blurred remaining pure semantic signal.

[0079] Step S403: The fuzzy residual pure semantic signal and the retained pure semantic signal are fused to obtain the fuzzy pure semantic signal.

[0080] In step S401 of some embodiments, specifically, the semantic mask value refers to the binarized value of the semantic unit mask after the signal attention score in the semantic signal mask matrix is ​​greater than a preset score threshold, i.e., 0.

[0081] Specifically, preserving the pure semantic signal means that the semantic unit corresponding to the position with a semantic mask value of 0, i.e. the key semantic unit, is retained in its original state in subsequent processing without being blurred.

[0082] Specifically, semantic units with a mask value of 0 can be selected from the estimated pure semantic signal based on the semantic mask value to determine the retained pure semantic signal.

[0083] In step S402 of some embodiments, specifically, the Gaussian filter refers to a linear smoothing filter based on the Gaussian function, which smooths the distorted semantic signal through convolution operation, so as to reduce the local detail semantics in the distorted semantic signal while preserving the overall structural information in the distorted semantic signal.

[0084] Specifically, the remaining signal refers to the non-critical semantic units corresponding to the positions in the estimated pure semantic signal where the mask value is 1.

[0085] Specifically, the fuzzy residual pure semantic signal refers to the semantic signal obtained by Gaussian fuzzing the remaining signals in the estimated pure semantic signal after retaining the pure semantic signal.

[0086] Specifically, the residual signal corresponding to the position with a mask value of 1 can be extracted from the estimated pure semantic signal. The residual signal is then convolved with a Gaussian filter to obtain the fuzzy residual pure semantic signal.

[0087] In step S403 of some embodiments, specifically, the fuzzy pure semantic signal refers to the semantic signal obtained after selectively fuzzing non-key semantic units in the estimated pure semantic signal, which is used to provide internal guidance for the generation of adversarial examples for the distorted semantic signal.

[0088] Specifically, the fuzzy pure semantic signal can be determined using the following formula:

[0089] in, This represents the fuzzy, pure semantic signal corresponding to the current noise propagation time step t. This represents the semantic signal mask matrix corresponding to the current noise diffusion time step t. This indicates element-wise multiplication. This represents the estimated clean semantic signal corresponding to the current noise propagation time step t. denoted as Gaussian filter, where v represents the standard deviation parameter of the Gaussian filter. The larger the value of this parameter, the lower the degree of preservation of detailed information in the distorted semantic signal.

[0090] Through steps S401 to S403, for complex dynamic wireless communication scenarios lacking prior channel information, the semantic importance distinction mechanism of the semantic signal mask matrix is ​​used to identify and retain distorted semantic signals that have a key impact on model decision-making. The remaining distorted semantic signals are then subjected to Gaussian blurring, which weakens local details while preserving the overall structural information of the distorted semantic signals, forming an internal reference that is "semantically missing but structurally consistent". This means that the internal key information contained in the distorted semantic signals can be used as a guiding condition for the generation of adversarial signal sample diffusion. This avoids dependence on external channel state information and solves the semantic structural ambiguity problem caused by global semantic blurring of distorted semantic signals under strong constraints. This provides high-quality internal guiding conditions for subsequent fuzzy adversarial sample generation and skip denoising processing.

[0091] Through steps S301 to S303, a semantic signal mask matrix is ​​constructed by comparing attention scores and preset score thresholds. This achieves quantitative assessment and binarization of the importance of distorted semantic signals, which helps avoid semantic structure uncertainty problems that may be caused by global blurring of distorted semantic signals by Gaussian blurring. It can utilize the internal key information contained in the distorted semantic signals themselves as guiding conditions for the spread generation process of adversarial signal samples. This avoids dependence on external channel state information and ensures the semantic consistency and reconstruction accuracy of the subsequently generated adversarial samples and distorted semantic signals, thereby improving the accuracy and reliability of subsequent semantic signal recovery.

[0092] In step S105 of some embodiments, specifically, a fuzzy adversarial sample refers to a sample constructed based on a fuzzy pure semantic signal and a preliminary noise component, according to the noise addition rule of the forward process of the diffusion model. The sample maintains consistency with the distorted semantic signal in terms of semantic structure, but fuzzification is introduced at the detail level.

[0093] Specifically, fuzzy adversarial examples can be determined using the following formula:

[0094] in, This represents the fuzzy adversarial sample corresponding to the current noise propagation time step t. This represents the cumulative attenuation coefficient corresponding to the current noise propagation time step t. This represents the fuzzy, pure semantic signal corresponding to the current noise propagation time step t. This represents the initial noise component corresponding to the current noise diffusion time step t, and .

[0095] In step S106 of some embodiments, specifically, the adversarial example noise component refers to the noise vector representation for the prediction of the fuzzy adversarial example, and the noise dimension is also the same as the distorted semantic signal feature.

[0096] Specifically, the method for predicting sample noise of fuzzy adversarial samples based on the current noise diffusion time step is the same as the method for preliminary noise prediction of distorted semantic signals based on the current noise diffusion time step, and will not be elaborated here.

[0097] Specifically, based on the current noise diffusion time step, the noise of the fuzzy adversarial sample is predicted. The resulting adversarial sample noise component, together with the preliminary noise component obtained in the previous steps, constitutes the input basis for the subsequent skip denoising process. Through the difference information between the two, the degree of influence on semantic details can be quantitatively evaluated, thereby providing data support for the subsequent skip denoising.

[0098] Please see Figure 5 In some embodiments, step S107 includes, but is not limited to, steps S501 to S502: Step S501: Determine the target noise based on the preliminary noise component and the adversarial sample noise component.

[0099] Step S502: Perform skip denoising processing on the distorted semantic signal based on the target noise to obtain the target recovery signal.

[0100] Please see Figure 6 In some embodiments, step S501 includes, but is not limited to, steps S601 to S603: Step S601: Obtain the guiding scale parameter of the fuzzy pure semantic signal; wherein, the guiding scale parameter is used to characterize the credibility of the semantic structure of the fuzzy pure semantic signal.

[0101] Step S602: Component calculation is performed based on the adversarial sample noise component and the preliminary noise component to obtain the noise difference value.

[0102] Step S603: Based on the noise difference value and the guiding scale parameter, noise correction calculation is performed to obtain the target noise.

[0103] In step S601 of some embodiments, specifically, the guidance scale parameter is an adjustable parameter for the fuzzy pure semantic signal. This parameter is used to characterize the credibility of the semantic structure of the fuzzy pure semantic signal, so as to control the weight of the internal semantic guidance information in the noise correction process. The larger the value of this parameter, the higher the credibility of the semantic structure of the fuzzy pure semantic signal and the stronger the internal semantic guidance; conversely, the lower the credibility of the semantic structure of the fuzzy pure semantic signal, the weaker the internal semantic guidance.

[0104] Specifically, the values ​​of the guiding scale parameters can be determined according to the specific application scenario, and are not limited here.

[0105] For example, if the channel conditions are poor and the noise in the distorted semantic signal is strong, a larger guiding scale parameter (such as 7.5) can be set, making the noise correction process more dependent on the semantic structure information carried by the fuzzy pure semantic signal. This can make a stronger use of the adversarial example noise component to correct the initial noise estimate, while avoiding semantic structure ambiguity caused by an excessively large guiding scale. If the channel conditions are good and the distorted semantic signal itself is of high quality, a smaller guiding scale parameter (such as 2.5) can be set, making the noise correction process more dependent on the original noise estimation result, avoiding the introduction of bias due to over-correction.

[0106] In step S602 of some embodiments, specifically, the noise difference value refers to the difference between the adversarial sample noise component and the initial noise component, which is used to quantify the adjustment direction and adjustment magnitude of the internal guidance conditions on the noise estimation. This difference reflects the degree of deviation of the noise estimation result with or without internal structure guidance.

[0107] Specifically, the noise difference value can be obtained by subtracting the adversarial sample noise component from the initial noise component.

[0108] In step S603 of some embodiments, specifically, the target noise refers to the final noise estimate that has been weighted and corrected by the guiding scale parameter, which fuses the adversarial sample noise component and the initial noise component.

[0109] Specifically, the target noise can be determined using the following formula:

[0110] in, This represents the target noise corresponding to the current noise diffusion time step t. This represents the adversarial sample noise component corresponding to the current noise propagation time step t. This represents the initial noise component corresponding to the current noise diffusion time step t.

[0111] Through steps S601 to S603, for complex dynamic wireless communication scenarios lacking prior channel information, the correction strength of the internal guidance conditions is adjusted by the guidance scale parameter. While utilizing the structural information provided by the fuzzy pure semantic signal to guide noise estimation, the semantic structural ambiguity caused by over-reliance on the internal guidance is avoided. This ensures that the generated target noise retains the noise characteristics of the distorted semantic signal while incorporating the semantic constraints of the internal structural guidance, providing an accurate noise estimation basis for subsequent skip denoising processing and effectively improving the accuracy and reliability of semantic signal recovery.

[0112] Please see Figure 7 In some embodiments, step S502 includes, but is not limited to, steps S701 to S702: Step S701: Determine the starting jump denoising time step; wherein the starting jump denoising time step is earlier than the current noise diffusion time step, and the time interval between the starting jump denoising time step and the current noise diffusion time step is greater than 1.

[0113] Step S702: Perform first jump denoising processing on the distorted semantic signal based on the initial jump denoising time step and the target noise to obtain the first denoised signal corresponding to the initial jump denoising time step.

[0114] Step S703: Update the initial jump denoising time step to obtain the target jump denoising time step; wherein, the time of the target jump denoising time step is earlier than the initial jump denoising time step.

[0115] Step S704: Perform second jump denoising processing on the first denoised signal based on the target jump denoising time step and the target noise corresponding to the target jump denoising time step to obtain the target recovered signal corresponding to the target jump denoising time step.

[0116] In step S701 of some embodiments, specifically, the initial jump denoising time step is the first target time step in the jump sampling sequence pre-set in the jump sampling strategy. This time step is earlier than the current noise diffusion time step and the time interval between the two is greater than 1, which is used to identify the target denoising stage that jumps directly from the current noise state.

[0117] Furthermore, the skip sampling sequence can be represented as ,in T represents the total number of sampling steps in the diffusion model. This scheme follows the principle of exponential sparse sampling to select the skip time step sequence.

[0118] For example, if the total number of diffusion time steps in the diffusion model is Then the selected skip sampling sequence can be: Furthermore, the sequence uses a large step size for fast noise reduction in high-noise regions (1000 to 400), and gradually reduces the step size in low-noise detail recovery regions (400 to 10) to ensure the accuracy of semantic signal recovery. If the current noise diffusion time step is 1000 steps, the starting jump denoising time step is determined to be 800 steps.

[0119] By selecting a skip sampling sequence, this application can minimize the number of sampling steps while ensuring the quality of semantic signal recovery, thereby effectively reducing inference latency and computational overhead.

[0120] Please see Figure 8 In some embodiments, step S702 includes, but is not limited to, steps S801 to S804: Step S801: Obtain the current cumulative attenuation coefficient corresponding to the current noise diffusion time step, and obtain the initial cumulative attenuation coefficient corresponding to the initial jump denoising time step.

[0121] Step S802: Calculate the denoised signal quantity based on the current cumulative attenuation coefficient, the distorted semantic signal, and the target noise.

[0122] Step S803: Calculate the noise compensation term based on the initial cumulative attenuation coefficient and the target noise.

[0123] Step S804: Calculate based on the noise compensation term and the denoised signal quantity to obtain the first denoised signal.

[0124] In step S801 of some embodiments, specifically, the current cumulative attenuation coefficient refers to the cumulative noise intensity parameter corresponding to the current noise propagation time step, which is used to characterize the cumulative retention ratio of the original semantic information from the noise propagation start time step to the current time step.

[0125] Specifically, the initial cumulative attenuation coefficient refers to the cumulative noise intensity parameter corresponding to the initial jump denoising time step, which is used to characterize the cumulative retention ratio of the original semantic information from the noise diffusion start time step to the initial jump denoising time step.

[0126] In step S802 of some embodiments, specifically, the denoised signal quantity refers to an intermediate quantity calculated based on the current cumulative attenuation coefficient, the distorted semantic signal, and the target noise, used to reflect the signal components obtained after removing noise components from the distorted semantic signal at the current noise diffusion time step.

[0127] In step S803 of some embodiments, specifically, the noise compensation term refers to the noise correction term calculated based on the initial cumulative attenuation coefficient and the target noise, which is used to compensate for the noise components that need to be adjusted due to the change in noise intensity when jumping from the current time step to the initial jump denoising time step during the jump denoising process.

[0128] In step S804 of some embodiments, specifically, the first denoising signal refers to a denoised semantic signal obtained by removing noise from a part of the distorted semantic signal corresponding to the starting jump denoising time step.

[0129] Specifically, for any time step pair satisfying m < n, the first denoising signal can be determined by the following jump denoising formula:

[0130] where represents the denoising signal corresponding to the starting jump denoising time step m, represents the distorted semantic signal corresponding to the noise diffusion time step n, represents the cumulative attenuation coefficient corresponding to the noise diffusion time step n, represents the target noise corresponding to the noise diffusion time step n, represents the starting cumulative attenuation coefficient corresponding to the starting noise diffusion time step m, represents the variance parameter introduced during the denoising process at the current noise diffusion time step t, which is used to control the randomness of the denoising process; if = 0, the denoising process degenerates into a deterministic mapping.

[0131] Furthermore, represents the denoised signal amount, represents the noise compensation term. If n is t, the noise diffusion time step n can be represented as the current noise diffusion time step t.

[0132] Through steps S801 to S804, discontinuous jump sampling from the current noise diffusion time step to the starting jump denoising time step is achieved, and jump denoising of the distorted semantic signal is realized while ensuring the quality of semantic signal recovery, significantly improving the efficiency of semantic signal recovery.

[0133] In one available scenario of this embodiment, the jump denoising formula can be deduced in the following way: Specifically, it is to weaken the explicit dependence on adjacent time steps in the reverse diffusion process. If the reverse denoising process, under the condition of the given distorted semantic signal and the clean sample obeys the following Gaussian distribution:

[0134] where represents a custom conditional probability distribution symbol, represents the distorted semantic signal corresponding to the previous time step t - 1 of the current noise diffusion time step, represents the distorted semantic signal corresponding to the current noise diffusion time step t, and N represents the Gaussian distribution, Indicates a pure sample The weighting coefficients, This represents the distortion semantic signal corresponding to the current noise propagation time step t. In prediction Weight of time, I represents the variance parameter introduced during the denoising process at the current noise propagation time step t, and I represents the identity matrix.

[0135] To further ensure consistency between the reverse denoising process and the first-stage noise addition process of the diffusion model, the reverse denoising process is reparameterized as follows:

[0136] in, This represents the distorted semantic signal corresponding to the previous time step t-1 of the current noise propagation time step. This represents the distorted semantic signal corresponding to the current noise propagation time step t. Indicates a pure sample The weighting coefficients, This represents the distortion semantic signal corresponding to the current noise propagation time step t. In prediction Weight of time, This represents the cumulative attenuation coefficient corresponding to the current noise propagation time step t. Indicates a pure sample. This represents the variance parameter introduced during the denoising process at the current noise propagation time step t.

[0137] Furthermore, combining the statistical characteristics of the forward process in the diffusion model, the following constraints can be obtained:

[0138]

[0139] in, Indicates a pure sample The weighting coefficients, This represents the distortion semantic signal corresponding to the current noise propagation time step t. In prediction Weight of time, This represents the cumulative attenuation coefficient corresponding to the current noise propagation time step t. This represents the variance parameter introduced during the denoising process at the current noise propagation time step t.

[0140] Furthermore, by solving the above system of equations using the method of undetermined coefficients, the parameters can be obtained. The parsing expression:

[0141]

[0142] in, Indicates a pure sample The weighting coefficients, This represents the distortion semantic signal corresponding to the current noise propagation time step t. In prediction Weight of time, This represents the cumulative attenuation coefficient corresponding to the current noise propagation time step t. This represents the variance parameter introduced during the denoising process at the current noise propagation time step t.

[0143] Furthermore, since the aforementioned reverse denoising process is no longer strictly bound by the Markov assumption, the conditional distribution of the above reasoning process can be further generalized into a skip-sampling form across time steps, that is, for any condition satisfying... The time step can be directly determined by... Sampling The jump denoising formula.

[0144] In step S703 of some embodiments, specifically, the target jump denoising time step refers to the target time step reached from the second jump to the last jump in the jump sampling sequence during subsequent jump denoising processing. This time step is earlier than the initial jump denoising time step and is used to further remove residual noise in the distorted semantic signal in order to finally recover the target recovered signal.

[0145] Specifically, the initial jump denoising time step can be replaced step by step according to the time step from the second jump to the last jump in the jump sampling sequence.

[0146] For example, if the skip sampling sequence is After completing the first jump denoising process, the current time step is updated to 800 steps, and the target jump denoising time step can be gradually updated to... step.

[0147] In step S704 of some embodiments, specifically, the target recovered signal refers to the clean semantic signal that is finally recovered after multiple skip denoising processes.

[0148] Specifically, the first denoised signal is subjected to second-jump denoising processing based on the target jump denoising time step and the target noise corresponding to the target jump denoising time step to obtain the second denoised signal. The target jump denoising time step is updated step by step according to the jump sampling sequence to repeat the above-mentioned jump denoising processing operation until the target jump denoising time step in the jump sampling sequence has been traversed to obtain the final target recovery signal. The method of performing second-jump denoising processing on the first denoised signal based on the target jump denoising time step and the target noise corresponding to the target jump denoising time step is the same as the method of performing first-jump denoising processing on the distorted semantic signal based on the initial jump denoising time step and the target noise, and will not be described in detail here.

[0149] For example, if the skip sampling sequence is The first jump denoising is the first denoised signal that jumps from time step 1000 to 800, the second jump denoising is the second denoised signal that jumps from time step 800 to 600, and the jump sampling sequence is traversed until the last jump denoising is the target recovered signal that jumps from time step 20 to 10. The entire semantic signal recovery process only requires 9 jump denoising operations, while the traditional diffusion model requires 1000 sequential iterations.

[0150] Through steps S701 to S704, the sequential iteration limitation of traditional diffusion models that strictly follow the Markov chain assumption is effectively overcome, reducing the number of iterations from thousands to tens or even a few steps, significantly reducing the inference latency of semantic communication. At the same time, through the fusion of semantic guidance information and the setting of deterministic mapping, the semantic consistency and coherence of the recovered signal are guaranteed, realizing efficient and high-quality semantic signal recovery in scenarios lacking prior channel information.

[0151] Through steps S501 to S502, while ensuring the quality of semantic recovery, the inference efficiency is significantly improved, and the semantic communication system is able to operate efficiently and stably in scenarios lacking prior channel information. This demonstrates the completeness and innovation of this technical solution in terms of noise fusion and skip sampling.

[0152] This application embodiment obtains the current noise propagation time step received by the receiving end and its corresponding distorted semantic signal, and performs preliminary noise prediction on the distorted semantic signal based on the current noise propagation time step to obtain a preliminary noise component; performs noise reduction estimation on the distorted semantic signal based on the preliminary noise component to obtain an estimated clean semantic signal; performs semantic fuzzing processing on the estimated clean semantic signal to obtain a fuzzy clean semantic signal; performs noise addition processing on the fuzzy clean semantic signal based on the preliminary noise component to generate a fuzzy adversarial sample corresponding to the distorted semantic signal; performs sample noise prediction on the fuzzy adversarial sample based on the current noise propagation time step to obtain an adversarial sample noise component; and performs jump denoising processing on the distorted semantic signal based on the preliminary noise component and the adversarial sample noise component to obtain the target recovery signal. In this application, for complex dynamic wireless communication scenarios lacking prior channel information, an internal guiding condition is constructed by estimating a clean semantic signal derived from the distorted semantic signal itself and performing semantic fuzzing processing. This replaces the reliance on external channel state information, avoiding the increased communication overhead caused by the difficulty in obtaining accurate and real-time prior channel information in existing technologies. Simultaneously, by performing skip denoising processing based on preliminary noise components and adversarial sample noise components, the sequential iteration limitation of traditional diffusion models strictly following the Markov chain assumption is effectively solved, significantly reducing the number of iteration steps required for denoising. This reduces the inference latency of semantic communication and effectively improves the efficiency of semantic communication in scenarios lacking prior channel information.

[0153] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described semantic signal recovery method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0154] Please see Figure 9 , Figure 9 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 901 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 902 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 902 can store the processing system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 902 and is called and executed by the processor 901 using the semantic signal recovery method of the embodiments of this application. The input / output interface 903 is used to implement information input and output; The communication interface 904 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 905 transmits information between various components of the device (e.g., processor 901, memory 902, input / output interface 903, and communication interface 904); The processor 901, memory 902, input / output interface 903, and communication interface 904 are connected to each other within the device via bus 905.

[0155] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described semantic signal recovery method.

[0156] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0157] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0158] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0159] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; 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 embodiment according to actual needs.

[0160] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0161] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0162] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0163] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. The coupling or direct coupling or communication connection between the shown or discussed units may be through some interfaces, or indirect coupling or communication connection between the apparatus or units, and may be electrical, mechanical, or other forms.

[0164] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0165] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0166] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0167] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A semantic signal recovery method, characterized in that, The method, applied to the receiving end of a semantic communication system, includes: Obtain the current noise diffusion time step and the corresponding distortion semantic signal; wherein, the distortion semantic signal is a noisy semantic signal; Based on the current noise propagation time step, a preliminary noise prediction is performed on the distorted semantic signal to obtain a preliminary noise component; Based on the preliminary noise components, the distorted semantic signal is denoised and estimated to obtain the estimated clean semantic signal. The estimated pure semantic signal is subjected to semantic fuzzing to obtain a fuzzy pure semantic signal; wherein, the fuzzy pure semantic signal is used to provide internal guidance for the generation of adversarial examples of the distorted semantic signal; The fuzzy pure semantic signal is noise-added based on the initial noise component to generate a fuzzy adversarial sample corresponding to the distorted semantic signal. Based on the current noise diffusion time step, the noise component of the fuzzy adversarial sample is predicted. The distorted semantic signal is subjected to skip-denoising processing based on the initial noise component and the adversarial sample noise component to obtain the target recovery signal.

2. The method according to claim 1, characterized in that, The step of performing semantic fuzzy processing on the estimated pure semantic signal to obtain a fuzzy pure semantic signal includes: Attention scoring is performed on the distorted semantic signal to obtain a signal attention score; The distorted semantic signal is semantically masked based on the signal attention score and a preset score threshold to obtain a semantic signal mask matrix; The estimated pure semantic signal is semantically blurred based on the semantic signal mask matrix to obtain a blurred pure semantic signal.

3. The method according to claim 2, characterized in that, The semantic signal mask matrix includes semantic mask values ​​for which the signal attention score is greater than the preset score threshold; the step of performing semantic fuzzing processing on the estimated pure semantic signal based on the semantic signal mask matrix to obtain a fuzzy pure semantic signal includes: Based on the semantic mask value, the estimated pure semantic signal is filtered to obtain a pure semantic signal that is retained. The remaining signals in the estimated pure semantic signal, excluding the retained pure semantic signal, are subjected to Gaussian blurring processing according to a preset Gaussian filter to obtain the blurred remaining pure semantic signal. The fuzzy residual pure semantic signal and the retained pure semantic signal are fused to obtain the fuzzy pure semantic signal.

4. The method according to claim 1, characterized in that, The step of performing skip denoising processing on the distorted semantic signal based on the preliminary noise component and the adversarial example noise component to obtain the target recovery signal includes: The target noise is determined based on the preliminary noise component and the adversarial sample noise component; The distorted semantic signal is subjected to skip-denoising processing based on the target noise to obtain the target recovery signal.

5. The method according to claim 4, characterized in that, Determining the target noise based on the preliminary noise component and the adversarial example noise component includes: Obtain the guiding scale parameter of the fuzzy pure semantic signal; wherein, the guiding scale parameter is used to characterize the credibility of the semantic structure of the fuzzy pure semantic signal; Based on the adversarial sample noise component and the preliminary noise component, component calculation is performed to obtain the noise difference value; The target noise is obtained by performing noise correction calculations based on the noise difference value and the guiding scale parameter.

6. The method according to claim 4, characterized in that, The step of performing skip-denoising processing on the distorted semantic signal based on the target noise to obtain the target recovered signal includes: Determine the initial jump denoising time step; wherein the initial jump denoising time step is earlier than the current noise diffusion time step, and the time interval between the initial jump denoising time step and the current noise diffusion time step is greater than 1. Based on the initial jump denoising time step and the target noise, the distorted semantic signal is subjected to first jump denoising processing to obtain the first denoised signal corresponding to the initial jump denoising time step. The initial jump denoising time step is updated to obtain the target jump denoising time step; wherein, the time of the target jump denoising time step is earlier than the initial jump denoising time step. The first denoised signal is subjected to second jump denoising processing based on the target jump denoising time step and the target noise corresponding to the target jump denoising time step to obtain the target recovered signal corresponding to the target jump denoising time step.

7. The method according to claim 6, characterized in that, The first jump denoising process, based on the initial jump denoising time step and the target noise, is applied to the distorted semantic signal to obtain the first denoised signal corresponding to the initial jump denoising time step, including: Obtain the current cumulative attenuation coefficient corresponding to the current noise diffusion time step, and obtain the initial cumulative attenuation coefficient corresponding to the initial jump denoising time step; The denoised signal quantity is calculated based on the current cumulative attenuation coefficient, the distorted semantic signal, and the target noise. Based on the initial cumulative attenuation coefficient and the target noise, a noise compensation term is calculated. The first denoised signal is obtained by calculation based on the noise compensation term and the denoised signal quantity.

8. The method according to any one of claims 1 to 7, characterized in that, The step of performing preliminary noise prediction on the distorted semantic signal based on the current noise diffusion time step to obtain preliminary noise components includes: Sine embedding is performed on the current noise diffusion time step to obtain the current time step features; Position encoding is performed on the distorted semantic signal to obtain distorted semantic signal features; The distorted semantic signal features and the current time step features are jointly encoded to obtain the fused signal time sequence features; Attention noise prediction is performed based on the temporal characteristics of the fused signal to obtain the preliminary noise component.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the semantic signal recovery method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the semantic signal recovery method according to any one of claims 1 to 8.

Images