Knowledge Distillation Enhanced DeepJSCC for Edge-Optimized Image Transmission Systems

The deployment of deep joint source-channel coding (DeepJSCC) on resource-limited devices faces significant challenges due to high model complexity, despite its proven efficacy in reducing latency and bandwidth requirements for next-generation semantic and task-oriented communications. To support heterogeneous architectures essential for diverse edge devices, we propose a cross-architecture knowledge distillation framework for heterogeneous DeepJSCC models. Central to our approach is a teaching assistant network integrating adaptive feature fusion modules (FFMs), which dynamically reconcile architectural gaps between teacher and student models. These modules dynamically resolve architectural mismatches between teacher and student models by generating student-adaptive representations, thereby mitigating feature space misalignment and enhancing distillation efficacy. Comprehensive experiments demonstrate that our approach substantially enhances reconstruction fidelity and perceptual quality in student models without increasing inference latency overhead, while significantly alleviating performance degradation induced by model compression.