A low-light robust semantic communication method for embodied intelligence scenarios

Abstract

The application discloses a low-light robust semantic communication method for embodied intelligent scenarios, which is named semantic flow matching SemFlow. First, at the transmitting end, the low-light image is separated into structure prior and texture detail subbands through logarithmic domain wavelet decomposition, and each subband is mapped into a compact latent representation by a frequency-differentiated semantic encoder, and spliced into a unified power-normalized symbol stream transmission. Then, at the receiving end, the coarse reconstruction is restored through a symmetric decoder and an inverse wavelet transform, and the reconstruction quality is gradually improved through a hierarchical generative refinement module. Finally, the inconsistency between the subbands is eliminated through the conditional high-frequency refinement and pixel domain correction stages, and the final normal-light reconstructed image is output. The application can complete generative reconstruction within 24 ms delay with only 8 solver steps, and obtain a gain of 5-7 decibels on the PSNR index relative to the representative baseline, meeting the needs of real-time visual communication of embodied intelligent agents.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically relating to a low-light robust semantic communication method for embodied intelligence scenarios. Background Technology

[0002] With the convergence of 6G wireless communication networks and embodied intelligence technology, real-time visual perception of intelligent agents is becoming a core service in next-generation wireless networks. Embodied intelligence refers to the ability of physically formed intelligent agents (such as autonomous robots, drones, and intelligent vehicles) to achieve task objectives through perception, decision-making, and action in real physical environments. In typical embodied intelligence application scenarios such as underground facility inspection, nighttime corridor navigation, post-disaster rubble search and rescue, and mine exploration, intelligent agents often face severe light-deficient environments. Under these photon-scarce working conditions, images captured by image sensors mounted on embodied intelligent agents often exhibit severe problems such as brightness attenuation, color degradation, and sensor noise amplification.

[0003] This degradation of visual information presents a fundamental contradiction: on the one hand, degraded image content needs to be restored to support reliable downstream decisions by remote servers or edge nodes; on the other hand, limited wireless channel capacity requires aggressive image compression to meet real-time transmission needs. Traditional split coding methods have approached the Shannon limit through complex channel coding, but their modular architecture precludes semantic-level joint optimization, thus failing to provide a unified mechanism for coordinating content restoration and compression.

[0004] Semantic communication mitigates the aforementioned contradictions by unifying source coding and channel coding within an end-to-end neural network architecture. The Deep Joint Source-Channel Coding (DeepJSCC) paradigm achieves elegant quality degradation without the clip effect and has been extended to areas such as rate control, channel adaptation, and physical layer integration. However, these methods primarily aim to minimize pixel-level distortion metrics such as mean square error (MSE) during training, tending to produce reconstructions that preserve global structure but lack fine-grained texture, especially when the source is degraded by low illumination.

[0005] Generative methods compensate for these shortcomings by synthesizing reasonable textures at the receiver. Diffusion-based methods absorb channel noise into the forward process at the noise model level, or guide the inverse trajectory using transmitter semantic descriptors at the conditional level; conditional decoders based on Generative Adversarial Networks (GANs) guide the output onto a natural image manifold through a discriminator. However, diffusion models typically require hundreds of steps of sequential denoising inference, resulting in high latency and making them unsuitable for the real-time response requirements of embodied agents; while GANs suffer from pattern instability, affecting the reproducibility of decoding. More fundamentally, existing methods neither initialize the generation process with prior knowledge of the channel transmission structure nor adaptively adjust the generated trajectory according to the brightness distribution of low-light sources.

[0006] In the field of low-light image enhancement, starting from Retinex-based illuminance-reflectance decomposition, various methods have been developed, including unpaired adversarial training, zero-reference self-supervised learning, Transformer-based global illuminance modeling, and generative methods such as score diffusion and rectified flow. However, these methods treat enhancement as an independent visual task, completely excluding channel-induced distortion from their optimization loop. When deployed downstream of semantic codecs, they fail to distinguish between source-level noise and channel transmission structure, leading to suboptimal reconstruction; conversely, existing semantic codecs assume the input image is well-exposed and lack mechanisms to handle brightness degradation.

[0007] In summary, in embodied intelligence scenarios, the joint optimization of augmented perception coding and channel-aware generative decoding remains an open problem at the intersection of the two fields. There is an urgent need for an end-to-end semantic communication framework that can adapt to the degradation characteristics of low-light sources and efficiently transmit and quickly recover high-quality visual information under limited bandwidth, in order to support reliable decision-making and autonomous action of embodied agents in low-light environments. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this invention provides a robust semantic communication method for low-light conditions in embodied intelligence scenarios, named Semantic Flow Matching (SemFlow). First, at the transmitting end, the low-light image is separated into structural prior and texture detail subbands through logarithmic wavelet decomposition. A frequency-differentiated semantic encoder maps each subband to a compact latent representation and concatenates them into a unified power-normalized symbol stream for transmission. Then, at the receiving end, a coarse reconstruction is recovered through a symmetric decoder and inverse wavelet transform, and the reconstruction quality is progressively improved through a hierarchical generative refinement module. The prior-conditionalized flow matching decoder synthesizes texture details along the learned ordinary differential equation (ODE) trajectory. Finally, conditional high-frequency refinement and pixel-domain correction stages eliminate inconsistencies between subbands, outputting the final normal-light reconstructed image. This invention can complete generative reconstruction with only 8 solver steps and a 24-millisecond delay, achieving a 5-7 dB gain in PSNR relative to a representative baseline, meeting the requirements of real-time visual communication for embodied intelligent agents.

[0009] The technical solution adopted by this invention to solve its technical problem is as follows:

[0010] Step 1: At the transmitting end, the low-light image is decomposed into a logarithmic wavelet decomposition to separate the image into a low-frequency approximate sub-band and a high-frequency detail sub-band. Each sub-band is encoded into a compact latent representation by a frequency-differentiated semantic encoder, spliced ​​into a unified symbol stream, and then transmitted after power normalization.

[0011] Step 2: At the receiving end, each subband is recovered by a symmetric semantic decoder, and a rough logarithmic domain estimate is generated by inverse amplitude scaling and inverse discrete wavelet transform. A hierarchical generative refining module is constructed to progressively optimize the rough estimate.

[0012] Step 3: In the hierarchical generative refining module, the prior conditional flow matching stage enhances the low-frequency approximate subband, the conditional high-frequency residual module calibrates the high-frequency detail subband, and the pixel domain refining stage resolves cross-subband inconsistencies, outputting the final reconstructed normal illumination image.

[0013] Furthermore, step 1 specifically includes:

[0014] Step 1-1: Perform logarithmic transform and second-order discrete wavelet transform on the low-light image;

[0015] Given a low-light image First, a logarithmic transformation is performed to expand the dynamic range of dark pixels. These represent the image's height, width, and number of channels, respectively; pixel values ​​are first mapped to unit intervals for representation.

[0016]

[0017] In the formula, This indicates cropping the pixel values ​​to... Interval; logarithmic transformation from each pixel Mapped to And expand contrast in dark pixel areas; then linearly rescale to This allows the distribution to be centered at zero to match the tanh output range of the subsequent encoder; this two-step mapping is denoted as... Its inverse transform is denoted as ;

[0018] right Recursively applying the second-order two-dimensional Haar discrete wavelet transform (DWT) to decompose the data into multi-scale subbands:

[0019]

[0020] In the formula, , For approximate LL subband, , , These are detailed HF subbands that capture information in the vertical, horizontal, and diagonal directions, respectively; among them, For the current decomposition level index, To determine the maximum decomposition level, after the wavelet decomposition process, the final level (i.e., the [number]th level) is retained. The approximate subbands of each level are used for subsequent processing; the three directional detail subbands of each level are stitched together along the batch dimension into a single tensor. ; This represents a two-dimensional discrete wavelet transform operation;

[0021] Applying amplitude normalization limits all sub-bands to the encoder's operating range:

[0022]

[0023] In the formula, and These are the amplitude normalization coefficients for the approximate subband and the detail subband, respectively, used to counteract the energy scaling effect during wavelet decomposition.

[0024] Steps 1-2: Encode each subband using a frequency-differentiated semantic encoder;

[0025] LL Semantic Encoder The final stage generated by the wavelet decomposition is the 1st wavelet. Level approximation subbands, and use multi-scale convolution heads from Extract parallel feature maps, and then... Convolutional stitching projection proceeds through two sequential stages. Each stage contains two residual dense blocks (RRDBs) and a convolutional block attention module (CBAM). Spatial resolution is halved between stages via straddle convolutions. The final projection head outputs the latent representation.

[0026]

[0027] In the formula, For LL encoder parameters, This indicates the number of channels in the output feature map of the LL encoder;

[0028] HF subbands of each group Lightweight single-branch HF encoder The encoding consists of a convolutional backbone, an RDB-CBAM block, two consecutive strided convolutions for spatial compression, and... projection:

[0029]

[0030] In the formula, For the first HF encoder parameters This indicates the number of channels in the output feature map of the HF encoder;

[0031] Steps 1-3: Concatenate all subband latent representations into a unified symbol vector, normalize the power, and then transmit.

[0032] All subband latent representations are concatenated into a unified symbol vector. and normalized to unit average power ;

[0033] The channel transmission model is as follows:

[0034]

[0035] In the formula, The fading coefficient is... For element-wise multiplication, ;

[0036] The actual physical transmission environment corresponding to the channel transmission model includes: additive white Gaussian noise (AWGN) channel, Rayleigh fading, and Rice fading.

[0037] Furthermore, step 2 specifically includes:

[0038] Step 2-1: Recover each wavelet subband using a symmetric semantic decoder;

[0039] set up and These are the power denormalized and demultiplexed received latent vectors, respectively; LL decoder It consists of a convolutional expansion layer, two sequential RRDB-CBAM stages, and a final projection head, with the two sequential RRDB-CBAM stages separated by a pixel shuffling upsampling layer.

[0040]

[0041] In the formula, For LL decoder parameters;

[0042] Each HF decoder Include Extended, two RDB-CBAM blocks and convolutional projection, each of the two RDB-CBAM blocks is followed by pixel shuffling and upsampling:

[0043]

[0044] In the formula, For the first HF decoder parameters;

[0045] Step 2-2: Generate a rough logarithmic domain estimate through inverse amplitude scaling and inverse wavelet transform;

[0046] Inverse amplitude scaling restores the original dynamic range: , Then through The inverse DWT generates a rough logarithmic field estimate:

[0047]

[0048] Steps 2-3: Construct a hierarchical feature extractor to provide multi-scale conditional information for the subsequent refining stage;

[0049] Hierarchical feature extractor For rough estimation Apply an RDB-CBAM layer with a convolution stride of 2 to perform stepwise spatial downsampling:

[0050]

[0051] In the formula, , To create spatial feature maps with progressively halved resolution, From A global scene descriptor that is pooled and compressed using a two-layer multilayer perceptron (MLP).

[0052] Furthermore, step 3 specifically includes:

[0053] Step 3-1: Enhance approximate subbands in the prior conditional flow matching stage;

[0054] The low-frequency enhancement is formalized as a stream matching problem, and the decoded approximate subbands are... Transmitted along the learning trajectory to the truth target ;

[0055] set up , The interpolation path is defined as:

[0056]

[0057] Nonlinear sinusoidal transmission scheduling is employed.

[0058]

[0059] Its bell-shaped derivative satisfies And in The peak value is reached at this point, concentrating the transmission budget in the area with the greatest brightness correction;

[0060] To introduce state feedback to resist channel disturbances, a state-dependent velocity field is used, from equation... Rewritten as:

[0061]

[0062] For nonlinear sinusoidal transmission scheduling , The closed-form ratio is obtained as follows:

[0063]

[0064] because ( In practice, truncated variants are used. , ;

[0065] Constructing a state-dependent velocity field It has the following two key properties:

[0066] Property 1, or path equivalence: For any... , equal ;

[0067] Property two, namely local asymptotic stability: ODE Local asymptotic stability along the ideal trajectory: All eigenvalues ​​in The above is The disturbance decays according to the following rules:

[0068]

[0069] The ordinary differential equation (ODE) evolution trajectory used to enhance the approximate subband, i.e. the generating flow, is determined by the parameterized velocity field. The driver is implemented as a conditional U-Net with two-level spatial downsampling; three conditional mechanisms are used to inject prior and contextual cues: first, and The input is concatenated along the channel dimension to form a 6-channel input, enabling the network to directly estimate the residuals. Second, time step and global descriptor Third, quarter-resolution spatial features are jointly embedded using two MLP layers and added as a time-conditional signal to each residual block; U-Net bottleneck layer modulated by feature-level linear modulation FiLM:

[0070]

[0071] In the formula, and for convolution;

[0072] Enhanced subband The velocity field is obtained through uniform Euler step integration:

[0073]

[0074] Final iteration As an enhanced approximate subband ;

[0075] Step 3-2: Conditional high-frequency residual module calibration of detail sub-band;

[0076] Introducing a conditional high-frequency residual module , utilizing from Guided artifact removal is performed based on structural priors; detail subbands are received through independent branch coding. The conditional features matching the resolution are spliced ​​and fused through channels, and the projector generates residual correction via the RDB-CBAM backbone network; the corrected sub-band is:

[0077]

[0078] In the formula, For channel distortion input, To correct the output, A learnable residual scaling factor initialized to 0.3; each With spatial resolution and the first Feature maps of subband matching at the level of detail As a condition: Using full-resolution features Refining , Using half-resolution features Refining ;

[0079] Step 3-3: Wavelet synthesis and pixel-domain refinement;

[0080] Enhanced wavelet components Inverse DWT assembly:

[0081]

[0082] In the formula, and The outputs of steps 3-1 and 3-2 are respectively; Results After inverse logarithmic transformation Mapped to the linear pixel domain;

[0083] Pixel Domain Refinement Module This problem is solved by applying a final bounded residual correction in the pixel domain; shared A dual-branch coding-fusion architecture for encoding pixel domain estimation. and full-resolution conditional features After channel splicing and fusion and the RDB-CBAM backbone network, the final output is generated:

[0084]

[0085] In the formula, For pixel-domain input, For the final reconstructed normally lit image, a residual scaling factor can be learned. Initialized to 0.3.

[0086] Furthermore, the joint training objective of the low-light robust semantic communication method reflects the hierarchical structure of semantic flow matching SemFlow:

[0087]

[0088] In the formula, For the current training epoch; weights depend on the epoch. and Implementing course learning strategies; weighting and Fixed during training;

[0089] Codec loss With low light input The reference is a supervised semantic codec, while the brightness enhancement is delegated to a hierarchical generative refining module;

[0090] Stream matching loss train Predicting self-correction targets Add perturbations to the network input according to the random interpolation framework:

[0091]

[0092] Velocity target by clean interpolation Calculated using the deterministic conditional flow matching framework:

[0093]

[0094] High-frequency loss With true detail coefficients Supervision Output:

[0095]

[0096] End-to-end loss Optimization through complementary pixel, structure, perception, and chromaticity parameters with truth value The difference between them:

[0097]

[0098] In the formula, Measuring CIE Fidelity.

[0099] Preferably, the .

[0100] An electronic device includes: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to enable the electronic device to perform the aforementioned low-light robust semantic communication method.

[0101] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned low-light robust semantic communication method.

[0102] A chip includes a processor for retrieving and running a computer program from a memory, causing a device on which the chip is mounted to perform the aforementioned low-light robust semantic communication method.

[0103] A computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the aforementioned low-light robust semantic communication method.

[0104] The beneficial effects of this invention are as follows:

[0105] 1) This invention proposes the SemFlow end-to-end framework, which jointly optimizes low-light enhancement and semantic compression in embodied intelligence scenarios. Logarithmic wavelet decomposition separates structural priors from texture details, and a frequency-differentiated encoder compresses each subband with asymmetric capacity. The latent representations of all subbands are concatenated into a single power-normalized symbol stream, and implicit power allocation is learned through end-to-end training, prioritizing the protection of structural priors without explicit rate control.

[0106] 2) To achieve efficient generative decoding robust to channel impairments, this invention develops a priori conditionalized flow-matched decoder, whose state-dependent velocity field provides provably shrinking dynamics to channel perturbations. Nonlinear sinusoidal transmission scheduling concentrates generation capabilities in the brightness-critical mid-trajectory region, and initialization from the structured subband of channel decoding restricts the generation process to the residual-enhanced manifold.

[0107] 3) To recover high-frequency textures beyond those conveyed by structural priors, this invention introduces a hierarchical generative reconstruction module with conditional detail refinement and pixel domain correction stages, each stage being guided by multi-scale features reconstructed from coarse wavelet.

[0108] 4) System experiments under AWGN, Rayleigh, and Rice fading channels with three bandwidth ratios show that SemFlow surpasses the representative baseline by 5 to 7 dB in PSNR, while completing generative reconstruction in only 8 solver steps with a total latency of only 102 milliseconds, meeting the real-time visual transmission requirements of embodied agents. Ablation experiments confirm the complementary contributions of each stage, with the flow matching module alone contributing more than 10 dB of total quality gain. Attached Figure Description

[0109] Figure 1 This is the overall architecture diagram of the SemFlow framework proposed in this invention, which shows the logarithmic domain wavelet decomposition and frequency-differentiated semantic coding and channel transmission at the transmitter end, as well as the semantic decoding, inverse wavelet transform and hierarchical generative refining module (including three stages: stream matching, conditional high-frequency refining and pixel domain correction) at the receiver end.

[0110] Figure 2The graph shows the training convergence curves, where (a) is the curve of the weighted composite training loss changing with the number of training rounds under the AWGN channel and CBR = 1 / 6 conditions, and (b) is the curve of the global brightness difference between the codec output and input in the logarithmic domain changing with the number of training rounds.

[0111] Figure 3 Four channel models (AWGN, Rayleigh fading, Rice fading) are used under the condition of CBR = 1 / 6. Rice's decline The reconstructed PSNR follows The change curve.

[0112] Figure 4 The three channel bandwidth ratios under AWGN channel are CBR SSIM of {1 / 3, 1 / 6, 1 / 12} The change curve.

[0113] Figure 5 This is a comparison chart of visual quality for three test sets (LOL-v1, LOL-v2-real, LOL-v2-syn) under the conditions of CBR = 1 / 6 and SNR = 10 dB.

[0114] Figure 6 The PSNR and stream matching module delay vary with the number of Euler steps. A graph showing the changes.

[0115] Figure 7 This is a bar chart showing the distribution ratio of transmit power between the LL and HF subbands under three CBR configurations. Detailed Implementation

[0116] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0117] This invention provides a robust semantic communication method for low-light conditions based on flow matching, named Semantic Flow Matching (SemFlow), for embodied intelligent scenarios. First, at the transmitting end, the low-light image is separated into structural prior and texture detail subbands through logarithmic wavelet decomposition. A frequency-differentiated semantic encoder maps each subband to a compact latent representation and concatenates them into a unified power-normalized symbol stream for transmission. Then, at the receiving end, a coarse reconstruction is recovered through a symmetric decoder and inverse wavelet transform, and the reconstruction quality is progressively improved through a hierarchical generative refinement module. The prior-conditionalized flow matching decoder synthesizes texture details along the learned ordinary differential equation (ODE) trajectory. Finally, inconsistencies between subbands are eliminated through conditional high-frequency refinement and pixel-domain correction stages, outputting the final normal-light reconstructed image. This invention can complete generative reconstruction within a 24-millisecond delay with only 8 solver steps, achieving a 5-7 dB gain in PSNR relative to a representative baseline, meeting the requirements of real-time visual communication for embodied intelligent agents.

[0118] The technical solution adopted by this invention to solve its technical problem is as follows:

[0119] Step 1: At the transmitting end, the low-light image is decomposed into a logarithmic wavelet decomposition to separate the image into a low-frequency approximate sub-band and a high-frequency detail sub-band. Each sub-band is encoded into a compact latent representation by a frequency-differentiated semantic encoder, spliced ​​into a unified symbol stream, and then transmitted after power normalization.

[0120] Step 1-1: Perform logarithmic transform and second-order discrete wavelet transform on the low-light image;

[0121] Given a low-light image First, a logarithmic transformation is performed to expand the dynamic range of dark pixels. Pixel values ​​are first mapped to a unit interval representation:

[0122]

[0123] In the formula, Crop pixel values ​​to The range; this transformation will change each pixel from Mapped to And expand contrast in dark pixel areas; then linearly rescale to This makes the distribution centered at zero to match the tanh output range of the subsequent encoder; this two-step mapping is denoted as... Its inverse transform is denoted as ;

[0124] right Recursively applying a second-order two-dimensional Haar discrete wavelet transform (DWT) decomposes the data into multi-scale subbands:

[0125]

[0126] In the formula, , For approximate (LL) subbands, , , These are detail (HF) subbands that capture information in the vertical, horizontal, and diagonal directions, respectively. The decomposition is of the series; the three directional subbands of each series are spliced ​​along the batch dimension to form a single tensor. ;

[0127] Since the Haar coefficients scale with the decomposition depth, applying amplitude normalization limits all subbands to the encoder's operating range:

[0128]

[0129] In the formula, , ( hour);

[0130] Steps 1-2: Encode each subband using a frequency-differentiated semantic encoder;

[0131] LL Semantic Encoder Using multi-scale convolution heads from Extract parallel feature maps, and then... Convolutional stitching projection, through two sequential stages (each stage containing two residual dense blocks RRDB and convolutional block attention modules CBAM), with spatial resolution halved between stages via straddle convolution, finally outputs the latent representation from the projection head:

[0132]

[0133] In the formula, For LL encoder parameters;

[0134] HF subbands of each group Lightweight single-branch HF encoder The encoding consists of a convolutional backbone, an RDB-CBAM block, two consecutive strided convolutions for spatial compression, and... projection:

[0135]

[0136] In the formula, For the first HF encoder parameters;

[0137] Steps 1-3: Concatenate all subband latent representations into a unified symbol vector, normalize the power, and then transmit.

[0138] All subband latent representations are concatenated into a unified symbol vector. and normalized to unit average power Because the LL encoder uses a deep multi-scale RRDB stage while the HF encoder uses a shallow RDB branch, the capacity asymmetry leads to implicit power allocation: the high-capacity LL encoder generates a larger potential coefficient, which automatically occupies a larger share of the total transmit power after normalization; this allocation emerges naturally from the encoder design and end-to-end training without the need for an explicit power splitting module.

[0139] The channel transmission model is as follows:

[0140]

[0141] In the formula, The fading coefficient is... For element-wise multiplication, The evaluation was conducted under three standard channel models: Additive White Gaussian Noise (AWGN) Rayleigh decline ( ) and Rice's decline (factor) ).

[0142] Step 2: At the receiving end, each subband is recovered by a symmetric semantic decoder, and a rough logarithmic domain estimate is generated by inverse amplitude scaling and inverse discrete wavelet transform. A hierarchical generative refining module is constructed to progressively optimize the rough estimate.

[0143] Step 2-1: Recover each wavelet subband using a symmetric semantic decoder;

[0144] set up and The received potential vector after power denormalization and demultiplexing; LL decoder It consists of a convolutional expansion layer, two sequential RRDB-CBAM stages (separated by a pixel shuffling upsampling layer), and a final projection head:

[0145]

[0146] In the formula, For LL decoder parameters;

[0147] Each HF decoder Include Extended, two RDB-CBAM blocks (each followed by shuffled upsampling of pixels) and convolutional projection:

[0148]

[0149] In the formula, For the first HF decoder parameters;

[0150] Step 2-2: Generate a rough logarithmic domain estimate through inverse amplitude scaling and inverse wavelet transform;

[0151] Inverse amplitude scaling restores the original dynamic range: , Then through The inverse DWT generates a rough logarithmic field estimate:

[0152]

[0153] The estimate The global scene structure is preserved, but residual distortion still exists in the detail sub-bands;

[0154] Steps 2-3: Construct a hierarchical feature extractor to provide multi-scale conditional information for the subsequent refining stage;

[0155] Hierarchical feature extractor right Applying RDB-CBAM layers with a step-2 increment:

[0156]

[0157] In the formula, ( (This refers to a spatial feature map whose resolution is gradually halved.) From A global scene descriptor that is pooled and compressed using two layers of multilayer perceptron (MLP);

[0158] Step 3: In the hierarchical generative refining module, the prior conditional flow matching stage enhances the low-frequency approximate subband, the conditional high-frequency residual module calibrates the high-frequency detail subband, and the pixel domain refining stage resolves cross-subband inconsistencies, outputting the final reconstructed normal illumination image.

[0159] Step 3-1: Enhance approximate subbands in the prior conditional flow matching stage;

[0160] Decoded approximate subband It provides structurally faithful but hue-degraded priors; it formalizes low-frequency enhancement as a flow matching problem. Transmitted along the learning trajectory to the truth target ;set up , The interpolation path is defined as:

[0161]

[0162] Nonlinear sinusoidal transmission scheduling is employed.

[0163]

[0164] Its bell-shaped derivative satisfies And in The peak value is reached at the point where the transmission budget is concentrated in the area with the greatest brightness correction; this allocation is consistent with the nonlinearity of human visual perception, where changes in dark areas produce disproportionately large perceptual brightness changes, thereby directing the generation capability to the position where it is most needed perceptually.

[0165] To introduce state feedback to resist channel disturbances, a state-dependent velocity field is employed. From equation... Rewritten as:

[0166]

[0167] Regarding the aforementioned sinusoidal scheduling The closed-form ratio is obtained as follows:

[0168]

[0169] because ( In practice, truncated variants are used. , ;

[0170] This velocity field has the following two key properties:

[0171] Property 1 (Path Equivalence): For any... , equal ;

[0172] Proof: On the path, Substituting into ;

[0173] Property 2 (Local asymptotic stability): ODE Local asymptotic stability along the ideal trajectory: All eigenvalues ​​in The above is The disturbance decays according to the following rules:

[0174]

[0175] Proof: Let ,but ,because Integrating, we obtain the above expression;

[0176] because Due to channel noise disturbances, the ODE deviates from the ideal training trajectory; Property 2 indicates that in the ideal velocity field Under these conditions, such perturbations decay exponentially; when the learned... Approximating with sufficient accuracy At this time, this contraction behavior is approximately inherited, providing robustness to channel impairments that standard position-independent training targets do not possess;

[0177] The flow is learned from the velocity field The driver is implemented as a conditional U-Net with two-level spatial downsampling; three conditional mechanisms are used to inject prior and contextual cues: first, and The input is concatenated along the channel dimension to form a 6-channel input, enabling the network to directly estimate the residuals. Second, time step and global descriptor Third, quarter-resolution spatial features are jointly embedded using two MLP layers and added as a time-conditional signal to each residual block; Modulating the U-Net bottleneck layer using feature-level linear modulation (FiLM):

[0178]

[0179] In the formula, and for convolution;

[0180] Enhanced subband The velocity field is obtained through uniform Euler step integration:

[0181]

[0182] because It is already a meaningful reconstruction, trajectory length Channel distortion The scaling is much shorter than that of the unconditional diffusion model, which starts from Gaussian noise. Path; this prior conditional initialization combined with the contraction dynamics of property two, together make it possible to achieve only... High-fidelity reconstruction can be achieved in just one Euler step; final iteration As an enhanced approximate subband ;

[0183] Step 3-2: Conditional high-frequency residual module calibration of detail sub-band;

[0184] Introducing Conditional High Frequency Residual Module (HFRM) ), utilizing from The module performs guided artifact removal based on structural priors; the detail subbands are received through independent branch encoding. The conditional features matching the resolution are spliced ​​and fused through channels, and the projector generates residual correction via the RDB-CBAM backbone network; the corrected sub-band is:

[0185]

[0186] In the formula, For channel distortion input, To correct the output, The learnable residual scaling factor is initialized to 0.3; for each HFRM With spatial resolution and the first Feature maps of subband matching at the level of detail Condition: HFRM Using full-resolution features Refining HFRM Using half-resolution features Refining ;

[0187] Step 3-3: Wavelet synthesis and pixel-domain refinement;

[0188] Enhanced wavelet components Inverse DWT assembly:

[0189]

[0190] In the formula, and The outputs of steps 3-1 and 3-2 are respectively; Results After inverse logarithmic transformation Mapped to the linear pixel domain;

[0191] Because each subband is operated independently in the wavelet domain, slight cross-subband phase inconsistencies may remain in the synthesized image; pixel domain refining module This problem is solved by applying a final bounded residual correction in the pixel domain; Shared HFRM A dual-branch coding-fusion architecture for encoding pixel domain estimation. and full-resolution conditional features After channel splicing and fusion and the RDB-CBAM backbone network, the final output is generated:

[0192]

[0193] In the formula, For pixel-domain input, For the final reconstructed normally lit image, a residual scaling factor can be learned. Initialize to 0.3;

[0194] The joint training objectives of steps 1 to 3 reflect the hierarchical structure of SemFlow:

[0195]

[0196] In the formula, For the current training epoch; weights depend on the epoch. and The curriculum learning strategy is implemented as follows: codec supervision is emphasized in the early stages of training to establish a stable transmission baseline, followed by gradual enhancement of perceptual supervision to optimize output quality; the remaining weights... and Fixed during training;

[0197] Codec loss With low light input To reference the supervised semantic codec and ensure faithful channel transmission, while delegating luminance enhancement to a hierarchical generative refining module; it combines subband, logarithmic, and linear pixel domains. Structural similarity (SSIM) and spatial mean matching terms;

[0198] Stream matching loss train Predicting self-correction targets To improve generalization performance, perturbations are added to the network input according to a random interpolation framework:

[0199]

[0200] Velocity target by clean interpolation Calculated using the deterministic conditional flow matching framework:

[0201]

[0202] High-frequency loss With true detail coefficients HFRM supervision Output:

[0203]

[0204] End-to-end loss Optimization through complementary pixel, structure, perception, and chromaticity parameters with truth value The difference between them:

[0205]

[0206] In the formula, Measuring CIE fidelity;

[0207] Example:

[0208] This invention proposes SemFlow, a low-light robust semantic communication method based on flow matching for embodied intelligent scenarios. This method unifies low-light enhancement and semantic compression, and provides an end-to-end solution for reliable visual communication of embodied intelligent agents in low-light environments through wavelet domain structure-texture separation and receiver flow matching reconstruction.

[0209] Step S1: At the transmitting end, perform logarithmic domain wavelet decomposition and frequency differential semantic coding on the low-light image.

[0210] Step S101: Perform logarithmic transformation and second-order Haar discrete wavelet transform on the low-light image.

[0211] In embodied intelligence scenarios, image sensors mounted on intelligent agents such as robots and drones capture low-light images in low-light environments. The pixel values ​​are concentrated near zero, which leads to two problems: unstable gradients during training and underutilization of channel bandwidth. Logarithmic transformation addresses these two problems by expanding the dynamic range of dark pixels.

[0212] First, pixel values ​​are mapped to unit interval representations:

[0213]

[0214] This transformation changes each pixel from Mapped to And expand the contrast in dark pixel areas. Then linearly rescale to This makes the distribution centered at zero to match the tanh output range of the subsequent encoder. This two-step mapping is denoted as... Its inverse transform is denoted as .

[0215] Then, to Recursively applying a second-order two-dimensional Haar discrete wavelet transform (DWT) decomposes the data into multi-scale subbands:

[0216]

[0217] In the formula, , To capture the global brightness structure, approximate (Low-Low, LL) subbands are used. , , These are high-frequency (HF) subbands that capture information in the vertical, horizontal, and diagonal directions, respectively. This represents the total decomposition level. The three directional subbands of each level are concatenated along the batch dimension to form a single tensor. This allows subsequent operations to be uniformly applied to all three directions.

[0218] Since the Haar coefficients scale with the decomposition depth, applying amplitude normalization limits all subbands to the encoder's operating range:

[0219]

[0220] In the formula, , ( (Time). Low-frequency coefficients concentrate most of the signal energy and are less sensitive to channel noise, while high-frequency coefficients are highly sensitive to channel noise. Therefore, a differentiated coding strategy is adopted.

[0221] Step S102: Encode each subband using a frequency-differentiated semantic encoder.

[0222] LL Semantic Encoder It is a deep, multi-scale network. Multi-scale convolutional heads ( , , (Convolution kernel) from the normalized approximate subband Extract parallel feature maps, and then... Convolutional concatenation and projection fusion of features. The fused features are processed through two sequential stages, each containing two residual-in-residual dense blocks (RRDBs) and three residual dense blocks (RDBs) in the residuals, with residual scaling applied. ,growth rate A convolutional block attention module (CBAM) and a multi-stage convolution (stride = 2) are used to halve the spatial resolution. The final projection head outputs the latent representation:

[0223]

[0224] In the formula, For LL encoder parameters, For potential dimensions.

[0225] HF subbands of each group Lightweight single-branch HF encoder Encoding. The encoder contains a convolutional backbone, an RDB-CBAM block (4 layers of dense connections), ,growth rate Two consecutive stride convolutions (stride=2) are used for spatial compression and projection:

[0226]

[0227] In the formula, For the first HF encoder parameters This represents a potential dimension for HF. The asymmetric encoder capacity reflects the dominant role of the LL subband in maintaining structural fidelity under channel distortion. A deep multi-scale RRDB stage is adopted, while Shallow RDB branches are used to reduce the overhead allocated to components that are not perceived as critical.

[0228] All subband latent representations are concatenated into a unified symbol vector. and normalized to unit average power .because and Due to the capacity asymmetry, high-capacity LL encoders naturally generate larger potential coefficients, which, after normalization, account for a larger proportion of the total transmit power. This implicit power allocation emerges from the encoder design and end-to-end training, eliminating the need for explicit power partitioning modules. Experiments under three CBR configurations show that the power share of the LL subband monotonically increases from 75.5% at CBR = 1 / 3 to 91.8% at CBR = 1 / 12, indicating that the system automatically concentrates more power on structural priors as bandwidth becomes scarcer.

[0229] Step S103: Transmit via channel.

[0230] Symbolic vector Transmitted over a wireless channel. The received signal is modeled as follows:

[0231]

[0232] In the formula, The fading coefficient is... For element-wise multiplication, The noise is additive white Gaussian noise. Evaluations were conducted under three standard channel models: AWGN (...). Rayleigh decline ( ) and Rice's decline (factor) In embodied intelligence scenarios, agents face multipath fading and rapid changes in channel conditions when moving in complex environments (underground facilities, building interiors, etc.). The aforementioned channel model can fully characterize these typical transmission environments.

[0233] Step S2: At the receiving end, each subband is recovered by a symmetric semantic decoder, and a coarse logarithmic domain estimate is generated by inverse amplitude scaling and inverse discrete wavelet transform, and a hierarchical generative refinement module is constructed.

[0234] Step S201: Recover each wavelet subband using a symmetric semantic decoder.

[0235] set up and This is the received latent vector after power denormalization and demultiplexing. The decoder reverses the coding operation by replacing the strided convolution with a pixel shuffle upsampling layer, restoring the original spatial resolution.

[0236] LL decoder It consists of a convolutional expansion layer, two sequential RRDB-CBAM stages (separated by a pixel shuffling upsampling layer), and a final projection head:

[0237]

[0238] In the formula, These are the parameters for the LL decoder.

[0239] Each HF decoder Include Extended, two RDB-CBAM blocks (each followed by shuffled upsampling of pixels) and convolutional projection:

[0240]

[0241] In the formula, For the first HF decoder parameters.

[0242] Step S202: Generate a rough logarithmic domain estimate by inverse amplitude scaling and inverse wavelet transform.

[0243] Inverse amplitude scaling restores the original dynamic range: , Then through The inverse DWT generates a rough logarithmic field estimate:

[0244]

[0245] The estimate The global scene structure is preserved, but residual distortion still exists in the detailed subbands, which needs to be addressed by the hierarchical generative refining module.

[0246] Step S203: Construct a hierarchical feature extractor to provide multi-scale conditional information for the subsequent refining stage.

[0247] Hierarchical feature extractor For rough estimation Applying progressively stepped-2 RDB-CBAM layers produces a shared multi-scale representation:

[0248]

[0249] In the formula, ( (This refers to a spatial feature map whose resolution is gradually halved.) From A global scene descriptor is pooled and compressed using two layers of MLP. These conditional features are shared across the subsequent three refining stages.

[0250] Step S3: Perform three-stage refining in the hierarchical generative refining module: prior conditional flow matching enhancement, conditional high-frequency residual calibration, and pixel domain refining.

[0251] Step S301: Enhance the low-frequency approximate subband in the prior conditional flow matching stage.

[0252] Decoded approximate subband It provides structurally faithful but tone-degraded priors. Unlike applying a fixed post-processing network, this invention formalizes low-frequency enhancement as a flow matching problem. Transmitted along the learning trajectory to the truth target .set up , The interpolation path is defined as:

[0253]

[0254] Logarithmic domain brightness enhancement follows a nonlinear correction profile: in Nearby, estimated Rough scene structure has been captured; in Nearby, to The residual gaps disappear, therefore the main brightness correction is concentrated in the middle of the trajectory. Standard linear scheduling Evenly distributing the transmission budget results in capacity mismatch at both ends. Therefore, a nonlinear sinusoidal transmission scheduling method is employed.

[0255]

[0256] Its bell-shaped derivative satisfies And in At its peak, the transmission budget is concentrated in the area of ​​greatest brightness correction. This allocation aligns with the perceptual nonlinearity of human vision—changes in dark areas produce disproportionately large perceptual brightness changes, thus directing generation capabilities to where they are most perceptually needed. (In the family) In the middle, the index It is a guarantee The smallest integer that is regular and allows closed derivatives without fractional powers.

[0257] Standard conditional flow matching target It is position-independent and does not carry information about the current ODE state. The information, therefore the discretization error and Channel disturbances will propagate uncorrected. Due to prior knowledge... In wireless transmission, disturbances from channel noise are unavoidable, and this uncorrected drift is particularly detrimental. To introduce state feedback, the interpolation path is rewritten... This yields the state dependency equivalent form:

[0258]

[0259] Regarding the aforementioned sinusoidal scheduling The closed-form ratio is obtained as follows:

[0260]

[0261] The choice of [aspect / factor] affects the difficulty of velocity field learning: smaller [aspect / factor] A larger ratio is generated at the mid-track. This forces the network to regress to a higher-amplitude target, thus making training unstable. In integer exponents, minimize Simultaneously satisfy Regularity. Due to ( In practice, truncated variants are used. , ; Cut-off only Activation time, within the effective training range The internal effects are negligible.

[0262] The following two properties establish the consistency and robustness of the proposed velocity field:

[0263] Property 1 (Path Equivalence): For any... , equal .

[0264] Proof: On the path, Substituting... .

[0265] Property 2 (Local asymptotic stability): ODE Local asymptotic stability along the ideal trajectory: All eigenvalues ​​in The above is The disturbance decays according to the following rules:

[0266]

[0267] Proof: Let .but ,because Integrating, we obtain the above expression.

[0268] because Due to channel noise disturbances, the ODE deviates from the ideal training trajectory. Property 2 indicates that in an ideal velocity field... Under these conditions, such perturbations decay exponentially. When the learning speed of the network... Approximating with sufficient accuracy In this case, this contraction behavior is approximately inherited, providing robustness to channel impairments not found in standard position-independent training objectives. This property is particularly important for embodied intelligence scenarios, as agents face constantly changing channel conditions when moving through complex environments.

[0269] The flow is learned from the velocity field The driver is implemented as a conditional U-Net with two levels of spatial downsampling. Three conditional mechanisms are used to inject prior and contextual cues:

[0270] First, and The input is concatenated along the channel dimension to form a 6-channel input, enabling the network to directly estimate the residuals. .

[0271] Second, time step and global descriptor It is jointly embedded by two MLP layers and added to each residual block as a time-conditional signal.

[0272] Third, quarter-resolution spatial features Modulating the U-Net bottleneck layer using feature-wise linear modulation (FiLM):

[0273]

[0274] In the formula, and for Convolution. The Dropout2d layer after the first downsampling stage is trained with regularization on a small LOL dataset.

[0275] The enhanced subband is obtained through a velocity field learned by uniform Euler step integration:

[0276]

[0277] because It is already a meaningful reconstruction, trajectory length Channel distortion The scaling is much shorter than that of the unconditional diffusion model, which starts from Gaussian noise. Path. This prior conditional initialization combined with the contraction dynamics of property two, together make it possible to achieve only... High-fidelity reconstruction can be achieved in a few Euler steps. Training and inference use the same step size to avoid distribution mismatch. Final iteration. As an enhanced approximate subband .

[0278] Step S302: Conditional high-frequency residual module calibration of high-frequency detail subband.

[0279] Although step S301 recovers the approximate subband, the decoding detail coefficients... It still carries channel-induced artifacts. A Conditional High-Frequency Residual Module (HFRM) is introduced. ) using from Guided artifact removal is performed based on structural priors.

[0280] This module receives detail subbands via independent branch encoding. The conditional features matching the resolution are then combined and fused through the RDB-CBAM backbone network, and the projector generates residual correction. The corrected sub-band is as follows:

[0281]

[0282] In the formula, For channel distortion input, To correct the output, This is a learnable residual scaling factor initialized to 0.3. (For each HFRM) With spatial resolution and the first Feature maps of subband matching at the level of detail Condition: HFRM Using full-resolution features Refining HFRM Using half-resolution features Refining .

[0283] Step S303: Wavelet synthesis and pixel domain refinement stage.

[0284] Enhanced wavelet components Inverse DWT assembly:

[0285]

[0286] In the formula, and The outputs for steps S301 and S302 are respectively. Results After inverse logarithmic transformation Mapped to the linear pixel domain.

[0287] Because each subband is operated independently in the wavelet domain, slight cross-subband phase inconsistencies may remain in the synthesized image. Pixel Domain Refinement Module This problem is addressed by applying a final bounded residual correction in the pixel domain. Shared HFRM A dual-branch coding-fusion architecture for encoding pixel domain estimation. and full-resolution conditional features After channel splicing and fusion through the RDB-CBAM backbone network, the final output is generated:

[0288]

[0289] In the formula, For pixel-domain input, For the final reconstructed normally lit image, a residual scaling factor can be learned. Initialized to 0.3. In embodied intelligence scenarios, the reconstructed image is sent to a remote server or edge node for downstream decision-making (such as obstacle detection, path planning, target recognition, etc.).

[0290] Step S304: End-to-end joint training.

[0291] The training objective reflects the hierarchical structure of SemFlow:

[0292]

[0293] In the formula, This refers to the current training epoch. The weights depend on the epoch. (linearly decays from 1.4 to 0.8) and (Increase linearly from 0.8 to 1.5) Implement course learning strategies.

[0294] Codec loss Combining subband, logarithmic, and linear pixel domains SSIM and spatial mean matching term, where the mean matching term explicitly prevents brightness drift caused by channel noise.

[0295] Stream matching loss (weight) The training speed network predicts self-correcting targets. To improve generalization performance, perturbations are added to the network input according to a random interpolation framework:

[0296]

[0297] In the formula, The velocity target is calculated using clean interpolation.

[0298]

[0299] High-frequency loss (weight) ):

[0300]

[0301] End-to-end loss :

[0302]

[0303] In the formula, , , , Measuring CIE Color fidelity.

[0304] Training uses the AdamW optimizer ( , Weight decay ), initial learning rate After the first 10 rounds of linear preheating, cosine-controlled annealing is performed until... Gradient clipping to the unit norm, with EMA decay of 0.999 applied during inference. Channel SNR changes from [previous value] during each mini-batch training period. dB uniform sampling.

[0305] To verify the effectiveness of the method of the present invention, a system experiment was conducted under the following settings.

[0306] Dataset and Evaluation Settings: The network was trained on the LOL-v1 dataset, which provides 485 pairs of low-light and normal-light images for training and 15 pairs for testing. Generalization ability was evaluated on LOL-v2-real (100 test pairs) and LOL-v2-syn (100 test pairs) without fine-tuning. Reconstruction fidelity was measured by Peak Signal-to-Noise Ratio (PSNR) and SSIM, and perceptual quality was measured by learning perceptual patch similarity (LPIPS). All methods were tested under AWGN, Rayleigh fading, and Ricean fading channels. Evaluation under dB conditions.

[0307] Implementation details: All models are implemented in PyTorch, running on an NVIDIA RTX 4080 GPU with a batch size of 16 and a block size of [missing information]. Training is performed. Wavelet decomposition series. Basic channel width Three Channel Bandwidth Ratio (CBR) configurations: CBR = 1 / 3, 1 / 6, and 1 / 12, corresponding to potential dimensions. = (24, 8), (17, 4), (5, 3). Stream matching inference uses... Each ODE step, the training time step is sampled from a logit-normal distribution ( ), Disturbance amplitude Speed ​​ratio cutoff The training can be done in a maximum of 1000 rounds, with an early stop time of 80 rounds.

[0308] Baseline Comparison: Comparison with two baseline types. Separate Scheme: Source codec (JPEG or JPEG 2000) paired with downstream enhancer (Retinex-Net or EnlightenGAN), reporting practical and capacity-reaching variants of LDPC with 16-QAM for each codec. End-to-End Scheme: DeepJSCC retrained under normal illumination supervision. All baselines are run under the same channel conditions and matched CBR.

[0309] Analysis of experimental results:

[0310] like Figure 2 As shown, regarding training convergence, the composite loss decreases sharply within the first 200 epochs and then stabilizes at approximately 1.7, confirming that the joint optimization with multiple heterogeneous loss terms does not diverge. The brightness difference reaches a low plateau around 50 epochs, indicating that the encoder / decoder first addresses global illumination alignment and then optimizes towards higher-order objectives such as structural details and flow velocity estimation.

[0311] like Figure 3 As shown, in terms of performance under different channel conditions, SemFlow performs better in AWGN channels. Achieving approximately 23.5 dB PSNR at 10 dB and exceeding 25 dB at 10 dB, with a marginal improvement of approximately 5 to 7 dB relative to the best baseline. This advantage persists across all four channel models. DeepJSCC avoids the clamp effect, but its PSNR remains limited to approximately 17 to 18 dB, attributed to its MSE training objective, which causes the encoder to learn to faithfully reproduce dark images rather than extract features beneficial for illumination recovery.

[0312] like Figure 4 As shown, in terms of bandwidth efficiency, even under the most aggressive compression (CBR = 1 / 12), The SSIM exceeds 0.79 at 1 dB and reaches 0.849 at 15 dB. The three CBR curves are approximately parallel across the SNR range, with a difference of about 0.04 between adjacent configurations, indicating that bandwidth compression and channel noise degrade quality in a roughly separable manner within the proposed framework. An SSIM of 0.870 at 15 dB is achieved at CBR = 1 / 6, within 0.022 at CBR = 1 / 3, while consuming only half the channel bandwidth.

[0313] like Figure 5 The visual quality comparison shown and the ablation experiment results shown in the table demonstrate that removing the flow matching module resulted in a 10.5 dB decrease in PSNR, disabling the HFRM module reduced PSNR by 1.6 dB, and removing the pixel refiner resulted in the largest numerical degradation (PSNR dropped to 5.19 dB), confirming the complementary contributions of each stage.

[0314] like Figure 6 As shown, in terms of inference efficiency, the latency of the streaming module scales approximately linearly with the number of solver steps, approximately 3.0 milliseconds per step. The latency is 24 milliseconds, with a total system delay of 102 milliseconds. Quality saturation can be achieved in 8 steps, stemming from two properties: prior initialization confines the ODE trajectory within the residual enhancement gap; and discretization errors on the contractile decay trajectory of the self-correcting velocity field. This latency level is perfectly feasible for the real-time visual communication requirements of embodied agents.

[0315] like Figure 7 As shown, regarding implicit power allocation, the LL subband receives a significantly larger share of power than its symbol share at each operating point: at CBR = 1 / 12, it accounts for 30.8% of the transmitted symbols but 91.8% of the total power. The LL power share monotonically increases from 75.5% at CBR = 1 / 3 to 91.8% at CBR = 1 / 12, indicating that the system automatically concentrates more power on structural a priori as bandwidth becomes scarcer.

Claims

1. A low-light robust semantic communication method for embodied intelligence scenarios, characterized in that, Includes the following steps: Step 1: At the transmitting end, the low-light image is decomposed into a logarithmic wavelet decomposition to separate the image into a low-frequency approximate sub-band and a high-frequency detail sub-band. Each sub-band is encoded into a compact latent representation by a frequency-differentiated semantic encoder, spliced ​​into a unified symbol stream, and then transmitted after power normalization. Step 2: At the receiving end, each subband is recovered by a symmetric semantic decoder, and a rough logarithmic domain estimate is generated by inverse amplitude scaling and inverse discrete wavelet transform. A hierarchical generative refining module is constructed to progressively optimize the rough estimate. Step 3: In the hierarchical generative refining module, the prior conditional flow matching stage enhances the low-frequency approximate subband, the conditional high-frequency residual module calibrates the high-frequency detail subband, and the pixel domain refining stage resolves cross-subband inconsistencies, outputting the final reconstructed normal illumination image.

2. The low-light robust semantic communication method for embodied intelligence scenarios according to claim 1, characterized in that, Step 1 specifically includes: Step 1-1: Perform logarithmic transform and second-order discrete wavelet transform on the low-light image; Given a low-light image First, a logarithmic transformation is performed to expand the dynamic range of dark pixels. These represent the image's height, width, and number of channels, respectively; pixel values ​​are first mapped to unit intervals for representation. ； In the formula, This indicates cropping the pixel values ​​to... Interval; logarithmic transformation from each pixel Mapped to And expand contrast in dark pixel areas; then linearly rescale to This allows the distribution to be centered at zero to match the tanh output range of the subsequent encoder; this two-step mapping is denoted as... Its inverse transform is denoted as ; right Recursively applying the second-order two-dimensional Haar discrete wavelet transform (DWT) to decompose the data into multi-scale subbands: ； In the formula, , For approximate LL subband, , , These are detailed HF subbands that capture information in the vertical, horizontal, and diagonal directions, respectively; among them, For the current decomposition level index, To determine the maximum decomposition level, after the wavelet decomposition process, the final level (i.e., the [number]th level) is retained. The approximate subbands of each level are used for subsequent processing; the three directional detail subbands of each level are stitched together along the batch dimension into a single tensor. ; This represents a two-dimensional discrete wavelet transform operation; Applying amplitude normalization limits all sub-bands to the encoder's operating range: ； In the formula, and These are the amplitude normalization coefficients for the approximate subband and the detail subband, respectively, used to counteract the energy scaling effect during wavelet decomposition. Steps 1-2: Encode each subband using a frequency-differentiated semantic encoder; LL Semantic Encoder The final stage generated by the wavelet decomposition is the 1st wavelet. Level approximation subbands, and use multi-scale convolution heads from Extract parallel feature maps, and then... Convolutional stitching projection proceeds through two sequential stages. Each stage contains two residual dense blocks (RRDBs) and a convolutional block attention module (CBAM). Spatial resolution is halved between stages via straddle convolutions. The final projection head outputs the latent representation. ； In the formula, For LL encoder parameters, This indicates the number of channels in the output feature map of the LL encoder; HF subbands of each group Lightweight single-branch HF encoder The encoding consists of a convolutional backbone, an RDB-CBAM block, two consecutive strided convolutions for spatial compression, and... projection: ； In the formula, For the first HF encoder parameters This indicates the number of channels in the output feature map of the HF encoder; Steps 1-3: Concatenate all subband latent representations into a unified symbol vector, normalize the power, and then transmit. All subband latent representations are concatenated into a unified symbol vector. and normalized to unit average power ; The channel transmission model is as follows: ； In the formula, The fading coefficient is... For element-wise multiplication, ; The actual physical transmission environment corresponding to the channel transmission model includes: additive white Gaussian noise (AWGN) channel, Rayleigh fading, and Rice fading.

3. The low-light robust semantic communication method for embodied intelligence scenarios according to claim 2, characterized in that, Step 2 specifically includes: Step 2-1: Recover each wavelet subband using a symmetric semantic decoder; set up and These are the power denormalized and demultiplexed received latent vectors, respectively; LL decoder It consists of a convolutional expansion layer, two sequential RRDB-CBAM stages, and a final projection head, with the two sequential RRDB-CBAM stages separated by a pixel shuffling upsampling layer. ； In the formula, For LL decoder parameters; Each HF decoder Include Extended, two RDB-CBAM blocks and convolutional projection, each of the two RDB-CBAM blocks is followed by pixel shuffling and upsampling: ； In the formula, For the first HF decoder parameters; Step 2-2: Generate a rough logarithmic domain estimate through inverse amplitude scaling and inverse wavelet transform; Inverse amplitude scaling restores the original dynamic range: , Then through The inverse DWT generates a rough logarithmic field estimate: ； Steps 2-3: Construct a hierarchical feature extractor to provide multi-scale conditional information for the subsequent refining stage; Hierarchical feature extractor For rough estimation Apply an RDB-CBAM layer with a convolution stride of 2 to perform stepwise spatial downsampling: ； In the formula, , To create spatial feature maps with progressively halved resolution, From A global scene descriptor that is pooled and compressed using a two-layer multilayer perceptron (MLP).

4. A low-light robust semantic communication method for embodied intelligence scenarios according to claim 3, characterized in that, Step 3 specifically includes: Step 3-1: Enhance approximate subbands in the prior conditional flow matching stage; The low-frequency enhancement is formalized as a stream matching problem, and the decoded approximate subbands are... Transmitted along the learning trajectory to the truth target ; set up , The interpolation path is defined as: ； Nonlinear sinusoidal transmission scheduling is employed. ； Its bell-shaped derivative satisfies And in The peak value is reached at this point, concentrating the transmission budget in the area with the greatest brightness correction; To introduce state feedback to resist channel disturbances, a state-dependent velocity field is used, from equation... Rewritten as: ； For nonlinear sinusoidal transmission scheduling , The closed-form ratio is obtained as follows: ； because ( In practice, truncated variants are used. , ; Constructing a state-dependent velocity field It has the following two key properties: Property 1, or path equivalence: For any... , equal ; Property two, namely local asymptotic stability: ODE Local asymptotic stability along the ideal trajectory: All eigenvalues ​​in The above is The disturbance decays according to the following rules: ； The ordinary differential equation (ODE) evolution trajectory used to enhance the approximate subband, i.e. the generating flow, is determined by the parameterized velocity field. The driver is implemented as a conditional U-Net with two-level spatial downsampling; three conditional mechanisms are used to inject prior and contextual cues: first, and The input is concatenated along the channel dimension to form a 6-channel input, enabling the network to directly estimate the residuals. Second, time step and global descriptor Third, quarter-resolution spatial features are jointly embedded using two MLP layers and added as a time-conditional signal to each residual block; U-Net bottleneck layer modulated by feature-level linear modulation FiLM: ； In the formula, and for convolution; Enhanced subband The velocity field is obtained through uniform Euler step integration: ； Final iteration As an enhanced approximate subband ; Step 3-2: Conditional high-frequency residual module calibration of detail sub-band; Introducing a conditional high-frequency residual module , utilizing from Guided artifact removal is performed based on structural priors; detail subbands are received through independent branch coding. The conditional features matching the resolution are spliced ​​and fused through channels, and the projector generates residual correction via the RDB-CBAM backbone network; the corrected sub-band is: ； In the formula, For channel distortion input, To correct the output, A learnable residual scaling factor initialized to 0.3; each With spatial resolution and the first Feature maps of subband matching at the level of detail As a condition: Using full-resolution features Refining , Using half-resolution features Refining ; Step 3-3: Wavelet synthesis and pixel-domain refinement; Enhanced wavelet components Inverse DWT assembly: ； In the formula, and The outputs of steps 3-1 and 3-2 are respectively; Results After inverse logarithmic transformation Mapped to the linear pixel domain; Pixel Domain Refinement Module This problem is solved by applying a final bounded residual correction in the pixel domain; shared A dual-branch coding-fusion architecture for encoding pixel domain estimation. and full-resolution conditional features After channel splicing and fusion and the RDB-CBAM backbone network, the final output is generated: ； In the formula, For pixel-domain input, For the final reconstructed normally lit image, a residual scaling factor can be learned. Initialized to 0.

3.

5. A low-light robust semantic communication method for embodied intelligence scenarios according to claim 4, characterized in that, The joint training objective of the low-light robust semantic communication method reflects the hierarchical structure of semantic flow matching SemFlow: ； In the formula, This is the current training round; Round-dependent weights and Implement course learning strategies; Weight and Fixed during training; Codec loss With low light input The reference is a supervised semantic codec, while the brightness enhancement is delegated to a hierarchical generative refining module; Stream matching loss train Predicting self-correction targets Add perturbations to the network input according to the random interpolation framework: ； Velocity target by clean interpolation Calculated using the deterministic conditional flow matching framework: ； High-frequency loss With true detail coefficients Supervision Output: ； End-to-end loss Optimization through complementary pixel, structure, perception, and chromaticity parameters with truth value The difference between them: ； In the formula, Measuring CIE Fidelity.

6. A low-light robust semantic communication method for embodied intelligence scenarios according to claim 5, characterized in that, The .

7. An electronic device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 6.

8. 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 as described in any one of claims 1 to 6.

9. A chip, characterized in that, include: A processor for retrieving and running a computer program from memory, causing a device on which the chip is mounted to perform the method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the method as described in any one of claims 1 to 6.

Images