Low-Power Signal Sensing and Neural Compression Transmission Collaborative Optimization Methods for Edge Intelligence

2.1 Overall architecture of the semantic-driven collaborative optimization framework

To break through the intractable triangle contradiction between semantic fidelity, energy consumption, and latency in edge image processing, this paper proposes a semantic-driven collaborative optimization framework, constructing a perception-encoding-transmission-decoding-feedback integrated design to achieve deep collaboration and dynamic adaptation across all stages. The architecture takes semantic information as the core link, connecting terminal-side perception encoding, edge-side decoding inference, and global collaborative decision-making. Through the organic interaction of three core modules, the framework establishes a “state perception - adaptive adjustment - global collaboration - semantic feedback” closed-loop optimization mechanism, ensuring the ability to dynamically approach the Pareto optimal boundary at the architectural level. This framework fundamentally discards the inherent flaws of traditional decoupled designs and achieves a collaborative balance between semantic fidelity, energy consumption, and latency by linking parameters across stages and sharing semantic information, thus adapting to the dynamic characteristics of edge scenarios.

The core modules of the framework include the conditional neural encoding perception module, semantic feedback control module, and fairness-oriented distributed game decision-making module. Each module has a clear functional boundary and tight interaction. The conditional neural encoding perception module is deployed on the terminal device and is responsible for image perception sampling and adaptive encoding tasks. Its core is a lightweight encoder based on the conditional neural encoding paradigm, which dynamically adjusts encoding parameters according to the terminal's energy consumption state, channel quality, and image semantic features, outputting encoding vectors that adapt to the transmission channel. The semantic feedback control module runs through both ends of the edge, where the edge-side task network extracts image semantic information and task performance feedback, generating lightweight control signals and feeding them back to the terminal to guide dynamic adjustment of the terminal’s perception and encoding parameters, forming a “terminal execution - edge perception - feedback adjustment” closed-loop control flow. The fairness-oriented distributed game decision-making module adopts a multi-agent architecture, with each terminal acting as an independent intelligent agent. Based on local states and global feedback information, the terminal makes decisions on key parameters such as perception sampling frequency and transmission power, achieving a balance between individual low-power demands and global resource-efficient utilization. Data and control flows interact orderly between modules: the image data collected by the terminal is processed by the conditional neural encoding perception module and then transmitted to the edge-side decoding module via the edge network; after the edge-side task network completes inference, semantic feedback and performance indicators are synchronized to the semantic feedback control module and distributed game decision-making module; the control instructions output by the decision-making module and closed-loop control signals jointly drive terminal module parameter updates, achieving full-link collaborative optimization.

Figure 1. Overall architecture of the semantic-driven collaborative optimization framework

This architecture precisely addresses the intractable triangle contradiction through the collaboration of the three modules, providing core support for dynamically approaching the Pareto optimal boundary. The conditional neural encoding perception module minimizes terminal energy consumption while ensuring semantic fidelity, directly alleviating the constraint between semantic fidelity and energy consumption; the semantic feedback control module ensures system stability based on control theory, quickly adjusting feedback to reduce performance fluctuations in dynamic scenarios and balancing the dynamic relationship between semantic fidelity and latency; the fairness-oriented distributed game decision-making module optimizes the global resource allocation, avoiding energy waste and latency surge caused by individual competition, achieving global-level collaboration between the three factors. The organic integration of the three modules forms an integrated architecture, enabling the system to dynamically adjust optimization goal weights in real-time under dynamic scenarios. It can dynamically adapt based on terminal states, channel changes, and task demands, ensuring that the system approaches the Pareto optimal boundary in different scenarios, significantly enhancing the overall performance and environmental adaptability of the edge image processing system. The specific architecture is shown in Figure 1.

2.2 Perception-encoding module based on conditional neural encoding

To achieve precise balance between semantic fidelity and energy consumption in dynamic scenarios, this paper proposes a unified conditional neural encoding paradigm, which generates adaptive weight increments through a meta-super network to drive the base encoder to dynamically adapt to terminal states and task demands, fundamentally solving the inherent flaws of traditional multi-model switching methods. The architecture is shown in Figure 2. The core idea of this paradigm is to integrate dynamic information such as terminal energy consumption state, channel quality, and image semantic features into a standardized joint state vector. The meta-super network learns the mapping relationship between the state and encoding parameters, enabling online dynamic instantiation of the encoder. Its core formal definition is:

$Encoder$$_\theta(x)=$ Base_Encoder$(x)+f_\phi$(Condition)              (1)

Condition $=\sigma\left(\left[\widehat{E}, S\widehat{N R}, W_{\text {semantic }}\right]\right)$               (2)

where, $Encoder$$_\theta(x)$ is the final instantiated encoder that adapts to the current scenario, with the input being the raw image $x$ and the output being the encoding vector adapted for transmission through the channel; $\theta$ is the set of parameters of the instantiated encoder, formed by the base encoder parameters and the weight increments output by the metasuper network; $Base$_$Encoder$ ($x$) is a fixed-structure lightweight base encoder responsible for extracting general semantic features from the image; $f_\phi()$ is the meta-super network, with $\phi$ being its learnable parameters. The core function of the meta-super network is to receive the joint state vector Condition and output dimension-matching weight increments; $\sigma()$ is a normalization function that normalizes state parameters with different dimensions to the range $[0,1]$, ensuring comparability of the input information; $\widehat{E}$ is the normalized terminal energy consumption state, $\widehat{S N R}$ is the normalized signal-to-noise ratio, and $W_{{semantic}}$ is the image semantic weight vector.

Figure 2. Perception-encoding module based on conditional neural encoding architecture

The design of the base encoder is deeply optimized for the characteristics of image processing scenarios, with the core goal of achieving efficient semantic feature extraction under the constraint of lightweight design. Its feature extraction process can be represented as:

$F=S A(D S \operatorname{Conv}(x))$                 (3)

where, F is the general semantic feature map output by the base encoder; DSConv( ) is a depthwise separable convolution operation, which reduces the computational complexity and parameter scale to 1/8~1/5 of that of traditional convolutions by separating spatial convolutions and pointwise convolutions while ensuring accuracy in spatial feature extraction; SA( ) is the spatial attention mechanism, which calculates the importance weights of each position in the feature map and performs weighted fusion, enhancing the representation of key semantic information such as target regions, and suppressing the interference of ineffective features in background areas, thus providing a reliable feature foundation for subsequent adaptive encoding. The depth of the base encoder’s network and the number of channels are constrained by lightweight design to ensure low-power operation on terminal devices.

The meta-super network adopts a lightweight Transformer architecture, with its core function being to precisely model the mapping relationship between the joint state vector and weight increments. Its output process can be represented as:

$\Delta \theta=f_\phi(Condition$)=$F F N(MultiHeadAttn$($Condition,Condition,Condition$,))             (4)

where, $\Delta \theta$ is the weight increment output by the meta-super network, and its dimension matches the parameters of the base encoder; MultiHeadAttn( ) is the multi-head attention mechanism, which models the interaction between the dimensions of $\widehat{E}, \widehat{S N R}$, and $W_{\text {semantic }}$ in the joint state vector, enhancing the comprehensiveness of state perception; $F F N()$ is the feedforward neural network, which maps the features output by the attention mechanism to the final weight increments. The meta-super network is trained end-to-end with the weighted sum of semantic fidelity loss and energy consumption loss as the optimization objective, learning the optimal weight adjustment strategy for different states to ensure that the instantiated $Encoder$$_\theta(x)$ satisfies both low power consumption and high semantic fidelity.

The conditional neural encoding paradigm demonstrates significant theoretical advantages in storage complexity, switching smoothness, and scalability, representing a substantial improvement compared to traditional multi-model switching methods. In terms of storage complexity, traditional methods require pre-training and storing N independent encoders for N scenarios, resulting in a storage cost of O(N); whereas this paradigm only requires storing the base encoder and meta-super network, and the total number of parameters is far less than the sum of parameters for N independent encoders, reducing the storage complexity to O(1), which does not change as the number of scenarios increases. In terms of switching smoothness, traditional methods achieve scenario adaptation by discretely switching between different encoders, which can lead to abrupt changes in semantic features and performance fluctuations; this paradigm generates continuous, adjustable weight increments Δθ through the meta-super network, allowing the parameters of Encoder_θ(x) to change continuously with the joint state vector, enabling a smooth transition in the encoding strategy, with performance fluctuations controlled within 5%. In terms of scalability, traditional methods require retraining and adding new encoders for new scenarios, limiting scalability; this paradigm can quickly adapt to new image types or terminal states through fine-tuning the meta-super network without modifying the base encoder structure, significantly improving generalization and scalability.

The semantic optimization in the perception sampling phase further reduces energy consumption by generating an image semantic mask using a lightweight object detection network and accurately calculating the semantic weight W_semantic, which can be represented as:

$W_{{semantic }}(i, j)= \begin{cases}w_1 & (i, j) \in {Reg}_{ {key }} \\ w_2 & (i, j) \in { Reg }_{b g}\end{cases}$             (5)

where, (i,j) is the image pixel coordinate; Reg_key is the key semantic region, Reg_bg is the background region; w₁ and w₂ are the weight coefficients for the key semantic and background regions, with w₁>w₂. This lightweight object detection network adopts the YOLO-Nano architecture, achieving real-time semantic region segmentation while ensuring object detection accuracy. After receiving the semantic weights W_semantic, the meta-super network drives the perception sampling module to dynamically adjust the resolution: high-resolution sampling is applied to Reg_key to ensure complete retention of semantic information, while low-resolution sampling is applied to Reg_bg, significantly reducing the sampling data volume and subsequent encoding energy consumption. This semantic-aware sampling strategy deeply collaborates with the conditional neural encoding paradigm, achieving precise energy consumption control from the perception source, reducing terminal perception phase energy consumption by 30%~45%, and further strengthening the balance between semantic fidelity and energy consumption.

2.3 Semantic-driven closed-loop control module

To ensure the collaborative stability of the perception-encoding-transmission full-link in dynamic edge scenarios and respond to performance impacts caused by channel quality fluctuations, terminal energy consumption changes, and dynamic task demand switching, this paper designs a semantic-driven closed-loop control module, constructing an end-to-edge collaborative closed-loop adjustment mechanism. The module takes semantic information as the core feedback carrier, dynamically calibrating the terminal encoding parameters and edge decoding strategy through the “perception-decision-execution-feedback” closed-loop link, ensuring that the system can still stably approach the Pareto optimal boundary of semantic fidelity, energy consumption, and latency in complex dynamic environments. Its core value lies in compensating for the deficiencies of traditional open-loop feedback response delay and insufficient adaptability, providing precise and real-time adjustment guidelines for the conditional neural encoding module, achieving dynamic balance of the end-to-edge collaboration. Figure 3 shows the complete architecture of the semantic-driven closed-loop control module.

Figure 3. Architecture of the semantic-driven closed-loop control module

The semantic-driven closed-loop control system consists of four core units: sensors, controller, actuator, and feedback link, with functional coupling and closed-loop links between them. The sensor is provided by the edge-side task network, and its core function is to perceive two key pieces of information in real-time: (1) the image processing task performance, which quantifies the level of semantic fidelity under the current encoding strategy; (2) the image semantic feature distribution, extracting core semantic information such as target region types and semantic importance rankings. The controller is the edge-side semantic decision unit, which, based on the task performance data collected by the sensor, global channel status, and terminal energy consumption feedback, generates lightweight control signals through preset decision rules to achieve a precise mapping of “performance-state-control.” The actuator is the terminal-side conditional neural encoding module, which dynamically adjusts encoding parameters upon receiving the control signals, completing adaptive updates to the encoding strategy. The feedback link adopts a lightweight semantic communication channel, which quantifies the encoding compression control signal data volume and reduces feedback energy consumption while ensuring transmission real-time performance, forming a complete closed loop of “terminal execution - edge perception - decision feedback - terminal adjustment.”

The design of the semantic control signal focuses on adaptability and lightweight, with the core goal of accurately guiding the terminal encoding strategy to match edge task demands. Its formal expression is:

$U=\left[\alpha \cdot W_{\text {semantic }}^*, \beta \cdot r_{\text {max }}\right]$              (6)

where, U is the semantic control signal vector, α is the semantic priority adjustment coefficient (α∈[0.6,1.2], which is dynamically adjusted based on task performance error), $W_{{semantic }}^*$ is the optimal semantic weight vector decided by the edge, used to update the priority of terminal semantic perception sampling and encoding; β is the encoding parameter constraint coefficient (β∈[0.5,1.0], which is related to channel bandwidth and terminal energy consumption state), and r_max is the maximum allowable compression ratio, defining the adjustment range of encoding parameters.

The two core components of the control signal form a synergy: semantic priority updating adjusts the semantic weights of different regions to ensure the encoding fidelity of key semantic information, avoiding performance degradation of core tasks due to excessive lightweighting; encoding parameter constraints balance encoding energy consumption and transmission latency by limiting the upper bound of compression ratios, preventing energy surges in transmission due to insufficient compression or semantic loss due to excessive compression. For different image processing tasks, the control signal can dynamically adapt the semantic priority ranking, such as in target detection tasks, prioritizing the semantic weight of target regions, while in semantic segmentation tasks, enhancing differentiated adjustment of pixel-level semantic category weights.

The system's stability and convergence are the core guarantees of the closed-loop control's effectiveness, and this paper proves it rigorously based on Lyapunov optimization theory. The core error term of the system is defined as the task performance error e(t), which is the deviation between the current semantic fidelity and the preset performance threshold, e(t)=Ttarget−T(t), where Ttarget is the preset task performance threshold and T(t) is the actual task performance at time t. The Lyapunov function is constructed as follows:

$V(t)=\frac{1}{2} e^2(t)+\frac{1}{2} \Delta \theta^T(t) P \Delta \theta(t)$                 (7)

where, Δθ(t) is the deviation vector of the encoding parameters from the optimal parameters at time t, and P is a positive-definite symmetric matrix, ensuring the function is positive definite, i.e., V(t)>0 for all e(t)=0 or Δθ(t)=0, and V(0)=0.

Taking the time derivative of the Lyapunov function and analyzing its negative definiteness:

$\dot{V}(t)=e(t) \dot{e}(t)+\Delta \theta^T(t) P \dot{\Delta} \theta(t)$             (8)

Combining the mapping relationship between task performance and encoding parameters $\dot{e}(t)=-k_1 e(t)-k_2 \Delta \theta(t)$, where ($k_1, k_2>0$) are proportional coefficients, and the encoding parameter update rule $\dot{\Delta} \theta(t)=-k_3 \Delta \theta(t)+k_4 U(t)$, where ( $k_3, k_4>0$ ) are adjustment coefficients, substituting the control signal $U(t)$ and the relationship with the error term, we derive $\dot{V}(t) \leq-\lambda V(t)$, where $\lambda>0$ is the convergence rate coefficient. This result shows that the derivative of the Lyapunov function is strictly negative definite, so the closedloop system is asymptotically stable, and the task performance error $e(t)$ will converge to a small neighborhood around 0 over time, with fluctuations controlled within $5 \%$, meeting the performance stability requirements of edge image processing.

Further analyzing the convergence of encoding parameters, by integrating $\dot{V}(t) \leq-\lambda V(t)$, we get $V(t) \leq \underline{\underline{V}}(0) e^{-\lambda t}$, and as $t \rightarrow+\infty$, $V(t) \rightarrow 0$, leading to $\Delta \theta(t) \rightarrow 0$, i.e., the encoding parameters will rapidly converge to the optimal value. From the theoretical analysis, the parameter convergence time constant is $\tau=1 / \lambda$, and by reasonably setting the adjustment coefficients $k 1 \sim k 4$, the convergence time can be controlled within 10 data transmission cycles, ensuring the system's fast response to dynamic scene changes. In summary, the semantic-driven closed-loop control module provides a stable and efficient dynamic adjustment mechanism for full-link collaborative optimization through rigorous structural design and theoretical guarantees.

2.4 Fairness-oriented distributed game decision-making mechanism

To address the imbalance between individual optimality and global optimality in multi-terminal competitive edge resource scenarios and ensure the collaborative efficiency and fairness of resource allocation in dynamic scenes, this paper proposes a fairness-oriented distributed game decision-making mechanism. This mechanism models each terminal device as an independent intelligent agent and uses game interactions to achieve adaptive resource allocation and dynamic adjustment of decision parameters. The core objective is to maximize the global utilization efficiency of edge network resources while satisfying the low power consumption needs of each terminal and the performance constraints of image processing tasks, enhancing the system's resilience under stress scenarios such as task conflicts and channel fluctuations. Its design overcomes the shortcomings of traditional distributed decision-making, which ignores global fairness, by integrating collaborative efficiency rewards that guide the agents to spontaneously form cooperative behaviors, providing global decision support for full-link collaborative optimization. Figure 4 shows the principle diagram of the fairness-oriented distributed game decision-making mechanism.

Figure 4. Principle of the fairness-oriented distributed game decision-making mechanism

The agent modeling centers on the terminal device and constructs the agent framework of “local state perception – global information interaction – strategy autonomous decision-making.” Each terminal corresponds to an independent game agent, whose observation space integrates local state and global feedback information, forming a high-dimensional observation vector, formally expressed as:

$O_i=\left[E_i, W_{{semantic } ,i}, S N R_{i, l o c a l}, J, C\right]$               (9)

where, O_i is the observation vector of the i-th agent; E_i, W_semantic,i, and SNR_i,local represent the local energy consumption state, semantic weight vector, and local channel quality of terminal i, respectively; J is the global Jain fairness index, quantifying the fairness level of edge resource allocation; and C is the channel congestion level, generated by the edge server based on global transmission traffic statistics. The decision goal of the agent is to adjust decision variables such as sampling resolution, encoding compression ratio, and transmission power to balance individual reward maximization and global collaborative efficiency, avoiding channel congestion or resource waste caused by malicious individual competition.

The policy network adopts a hybrid mode of “offline centralized pre-training + online distributed execution,” balancing decision accuracy and real-time performance. In the offline pre-training phase, utilizing the computing power advantage of the edge server, a simulation environment containing multiple terminals and multiple scenarios is constructed. Through centralized training, all agents share global data and learn collaborative decision strategies under different scenarios. The input of the policy network is the normalized observation vector O_i, and the output is the normalized decision variable vector A_i=[s_i,r_i,p_i], where s_i is the sampling resolution level, r_i is the encoding compression ratio, and p_i is the transmission power level. The network structure uses a lightweight Transformer architecture, modeling the interaction between local state and global feedback in the observation vector through a multi-head attention mechanism to improve the adaptability of the decision strategy. At the same time, channel pruning technology is introduced to reduce network computational complexity and ensure low-power characteristics in the online execution phase. In the online execution phase, each agent independently infers and outputs decisions based on local observation information, without intervention from the central node, and only shares the global Jain fairness index and channel congestion level through the edge server for distributed collaborative decision-making.

The design of the reward function is the core of guiding the agent to achieve individual and global collaborative optimization. It uses a weighted fusion mechanism to integrate three reward terms: power consumption savings, task performance, and collaborative efficiency, formally defined as:

$R_i=\alpha \cdot R_{ {pawex}, i}+\beta \cdot R_{ {task }, i}+\gamma \cdot R_{{coop }}$              (10)

where, R_i is the total reward of the i-th agent; α, β, and γ are the reward weight coefficients, satisfying α+β+γ=1, which can be dynamically adjusted based on the terminal energy consumption state and task priority. When the low power consumption constraint is strict, α is increased, and when the task priority is high, β is increased. $R_{ {power}, i}=k_1 \cdot \ln \left(E_{{max }, i} / E_i\right)$ is the power consumption savings reward, positively correlated with the energy consumption savings of terminal i, where k₁ is a proportional coefficient, E_max,i is the maximum rated power consumption of terminal i; R_task,i is the task performance reward, positively correlated with the image processing task accuracy, in the case of target detection tasks $R_{{task, }, i}=k_2 \cdot m A P_i$, where k₂ is a proportional coefficient and mAP_i is the mean average precision mAP_i of terminal i's target detection.

R_coop is the collaborative efficiency reward, which integrates the global Jain fairness index and channel congestion level to quantify global collaborative efficiency:

$R_{\text {coop }}=\eta \cdot J-\zeta \cdot C$             (11)

where, η, ζ are weight coefficients; J is the Jain fairness index.

$J=\left(\sum_{i=1}^N x_i\right)^2 /\left(N \sum_{i=1}^N x_i^2\right)$             (12)

where, x_i is the resource usage of terminal i and N is the total number of terminals. The closer J is to 1, the more equitable the resource distribution; C is the channel congestion level, C=Traffic/Bandwidth_max, where Traffic is the current total channel traffic and Bandwidth is the maximum channel bandwidth. The larger the value of C, the more serious the channel congestion. This design ensures that the agent's reward is not only dependent on its own performance but also linked to global fairness and congestion status, guiding the agent to avoid malicious competition and proactively engage in global collaboration.

The core of game equilibrium analysis is to prove that the system can converge to a Nash equilibrium that balances individual interests and global fairness. A distributed game is defined as $G=\left\{N, A_i, U_i\right\}$, where $N=\{1,2, \ldots, N\}$ is the set of agents, $A_i$ is the action space of agent $i$, and $U_i=R_i$ is the utility function of agent $i$. The definition of Nash equilibrium is: for all agents $i$, when the strategies of other agents are fixed as $A_{-i}^*$, the optimal strategy $A_i^*$ of agent $i$ satisfies $U_i\left(A_i^*, A_{-i}^*\right) \geq U_i\left(A_i, A_{-i}^*\right)$ for all $A_i \in A_i$.

Through theoretical derivation, it can be proven that the utility function U_i is strictly concave: since R_power,i, R_task,i, and R_coop are all strictly concave functions of the decision variables, their weighted sum U_i remains strictly concave. According to game theory, a strictly concave utility function corresponds to a unique Nash equilibrium in a distributed game. Further analysis shows that at this equilibrium point, the decision strategies of each agent can achieve a balance between individual reward maximization and global fairness. At this point, the Jain fairness index J ≥ 0.85, and the channel congestion level C ≤ 0.7, satisfying the resource allocation requirements of edge scenarios. In terms of simulation verification, by plotting the decision trajectory and reward change curve of multiple agents in a task conflict scenario, the process of the system converging from initial random decisions to Nash equilibrium can be intuitively displayed. The convergence time does not exceed 20 data transmission cycles, verifying the fast convergence and stability of the game decision-making mechanism.

2.5 End-to-end training and optimization

To ensure deep adaptation of the modules in the semantic-driven collaborative optimization framework and achieve global performance optimization, this paper designs a systematic end-to-end training and optimization process. This is achieved through the coordinated design of a joint training environment, staged training strategy, transfer learning initialization, and adaptive optimization mechanism, balancing training efficiency and model performance. The core objective is to allow the condition neural encoding module, semantic feedback control module, and distributed game decision-making module to form adaptation under a unified optimization goal, ensuring that the system can stably approach the Pareto optimal boundary of semantic fidelity, energy consumption, and latency in real-world edge scenarios.

The basis of the training process is to construct a joint training environment that integrates image datasets, channel simulations, and power consumption models. The image datasets merge publicly available datasets from multiple scenarios with real collected data, covering different lighting conditions, target densities, and image types, ensuring diversity and representativeness of the training samples. The channel simulation module supports typical edge channel models such as Rayleigh fading and AWGN, dynamically adjusting parameters like signal-to-noise ratio and bandwidth, to simulate real dynamic changes in the edge network. The power consumption model is built based on the terminal hardware characteristics, quantifying the energy consumption overhead of perception sampling, encoding computation, and transmission in each link, achieving precise evaluation and optimization of energy consumption during the training process. On this foundation, a staged training strategy is adopted to reduce the instability of multi-module collaborative training. The first stage pre-trains the condition neural encoding module and the semantic feedback control module, with a joint loss of semantic fidelity and energy consumption as the optimization target, allowing the encoder and feedback control system to initially adapt to the dynamic constraints of edge scenarios. The second stage fixes the pre-trained module parameters as initial values and introduces the distributed game decision-making module for global joint training. The Pareto optimization objective function of the entire framework guides the deep collaboration of decision strategies with the encoding and control modules. To improve training efficiency, a transfer learning strategy is introduced. The parameters of a lightweight CNN model pre-trained in the image domain are used to initialize the base encoder of the condition neural encoding module, reducing the amount of data and iterations required for model convergence by leveraging the general knowledge of image feature extraction. Meanwhile, an adaptive learning rate strategy based on training loss is adopted. The initial learning rate is set to 1e-3, and when the training loss does not show significant improvement for 3 consecutive epochs, it automatically decays to 1/10 of its original value, balancing the convergence speed in the early stage and convergence precision in the later stage. An early-stopping mechanism is introduced, using global performance metrics on the validation set as the criteria. When the metrics do not improve for 5 consecutive epochs, the training is terminated to avoid overfitting and ensure model generalization ability. The entire training process is implemented based on the PyTorch framework, using multi-GPU parallel acceleration for training. The parameters of each module are updated end-to-end through gradient backpropagation, ensuring efficient collaboration across all links of the framework under a unified optimization goal.

This work is highly aligned with cutting-edge technological trends such as 6G semantic communication, edge AI body coordination, and autonomy, providing key technical support and paradigm references for the development of related fields. In the 6G semantic communication domain, the core goal is to achieve "task-oriented" efficient data transmission, moving away from the traditional "bit transmission" model toward "semantic information transmission." The conditional neural encoding paradigm proposed in this paper, by integrating semantic weights with dynamic adjustments of encoding strategies based on channel and energy consumption states, essentially represents a realization of semantic communication in the edge image processing scenario. The encoding process only retains the core semantic information required for the task, greatly reducing transmission redundancy, which aligns with the 6G core demand for "on-demand transmission." At the same time, the semantic-driven closed-loop control and lightweight feedback loop design provide a practical technical solution for the end-edge collaborative architecture of 6G semantic communication. Its stability analysis and performance verification can provide theoretical references for the design of dynamic adaptation mechanisms in 6G semantic communication.

In the field of edge AI body coordination and autonomy, the distributed collaboration and autonomous decision-making of multi-terminal devices are core development directions, widely applied in complex scenarios such as vehicle-road collaboration and industrial IoT. The fairness-oriented distributed game decision-making mechanism proposed in this paper models each terminal as an independent agent, achieving autonomous collaborative decision-making through global feedback information, without the need for central node intervention, which fits the core need for "distributed autonomy" in edge AI systems. The excellent performance of this mechanism in heterogeneous terminal and task conflict scenarios demonstrates its applicability to more complex edge intelligent systems: in the vehicle-road collaboration scenario, it can be used for the cooperative transmission of sensing data between multiple vehicles and roadside units; in industrial IoT scenarios, it can achieve low-power data acquisition and transmission optimization for multiple sensor terminals. The game equilibrium analysis and cooperative efficiency reward design in this paper provide new design ideas for fair and efficient collaboration in edge AI bodies, promoting the evolution of edge intelligent systems from "individual intelligence" to "group collaborative intelligence."

The core advantages of the proposed method lie in three aspects: theoretical paradigm, technical mechanisms, and experimental verification. Theoretically, the "conditional neural encoding" paradigm breaks through the limitations of traditional adaptive encoding's discretization, providing a unified theoretical framework for low-power adaptive encoding. Its theoretical advantages in storage complexity and scalability offer paradigm insights for lightweight model design in edge scenarios. Technically, the collaborative design of semantic-driven closed-loop control and fairness-oriented distributed game decision-making achieves a balance between stability and global efficiency under dynamic scenes, addressing the core shortcomings of traditional edge collaboration methods that neglect fairness and stability. Experimentally, through comprehensive validation across multiple datasets and scenarios, including conventional tests, stress tests, and real-world deployments, and by combining interpretability analysis to reveal the core mechanisms, the effectiveness and practicality of the method are ensured, forming a comprehensive performance advantage over existing SOTA methods.

Objectively, there are three limitations in the current framework. First, the current framework assumes a relatively stable network topology and lacks adaptability to topological mutations such as fast terminal joining/leaving, which may lead to delayed collaborative decision-making and affect global performance when topology changes. Second, the semantic mask generation relies on a lightweight object detection network, and in complex scenarios such as low lighting or dense targets, the accuracy of semantic region division needs improvement, which can affect the precision of encoding strategies. Third, the personalized adaptation of terminal hardware characteristics is not considered. Different terminal processors have varying demands for encoding model computational efficiency, and the current design struggles to achieve deep hardware-algorithm matching.

To address these limitations, future research will focus on three directions. First, we will introduce graph neural networks to model the dynamic topological relationships between devices, learning the interaction correlations between terminals in real-time, improving the speed and accuracy of collaborative decision-making in topological mutation scenarios, and enhancing the system's dynamic adaptability. Second, we will integrate infrared and visible light multimodal image perception technologies, leveraging the advantages of infrared images in low-light scenarios to improve the accuracy of semantic mask generation in complex environments, providing more reliable semantic input for conditional neural encoding. Third, we will combine hardware awareness technologies to build a database of terminal hardware characteristics, quantifying the computational efficiency and energy consumption models of different hardware architectures, and enabling personalized matching of encoding parameters and hardware characteristics, further optimizing terminal energy consumption and computational efficiency.