A Dynamic Adaptive Deep Coding-Classification Framework for Intelligent Pragmatic Tasks in Semantic Communication

Semantic communication is a promising paradigm for intelligent task-oriented services under dynamic and resource-constrained environments. However, current designs suffer from limited generalization, high model computation, and inefficient training. To address these issues, a universal Dynamic Adaptive Deep Coding-Classification (DADC-C) framework for image reconstruction and classification is first proposed. Within this framework, by integrating Attention Feature Blocks (AFBs) and a semantic code (SC) mask mechanism, a deep joint source-channel coding (DeepJSCC) model, i.e. DynamicJSCCR, is then developed with dual dynamic adaptation to signal-to-noise ratio (SNR) and compression rate (CR). Through selective application of AFBs and streamlined Residual Convolution Blocks, the proposed lightweight DynamicJSCC-R reduces computation of floating-point operations (FLOPs) by 85% and improves the peak SNR (PSNR) by 1.61 dB and classification accuracy by 1.4%, respectively. Furthermore, a joint semantic coding-pragmatic task (SC-PT) training strategy is proposed, where a novel CR-based binary loss function is designed with dynamic weights to balance the reconstruction and classification objectives. Experimental results show that the proposed single-model DADC-C framework outperforms traditional methods with up to 30% higher classification accuracy under high SNRs, while achieving a comparable performance to multi-model DeepJSCC schemes with order-of-magnitude advantages in parameters, FLOPs, latency, frames per second (FPS) and storage.