Task-Generalizable Deep Joint Source-Channel Coding via Hierarchical Masked Autoencoders

Deep learning-based joint source-channel coding (DeepJSCC) has demonstrated significant potential for task-oriented semantic communication. However, existing DeepJSCC frameworks exhibit limited generalization across multi-tasks. This limitation stems from bandwidth compression, which results a trade-off between preserving details and capturing semantics. To overcome this challenge, we propose a novel Task-Generalizable DeepJSCC framework, called TGDJSCC. In the pre-training phase, we devise a hierarchical masked autoencoder (HMAE) to learn hierarchical representations within bandwidth-constrained latent spaces. Specifically, we introduce a novel content-aware tree for establishing hierarchy, along with a dual self-supervised mask image modeling scheme to learn transferable hierarchical representations. During fine-tuning, a task-adaptive feature selection module dynamically routes hierarchical features according to specific task requirements. This enables efficient adaptation to new tasks with only a few fine-tuning epochs for both pixel-level and semantic-level tasks, eliminating the need for full retraining. Comprehensive experiments demonstrate that TGDJSCC significantly surpasses existing DeepJSCC methods in task generalization, while maintaining communication efficiency under stringent bandwidth constraints.