Presentation Information
[B-17-06]A Study on Task-Aware JSCC Image Transmission Using VLM-Based Recognition
〇Manabu Takagi1, Kenji Ishii1 (1. Mitsubishi Electric Corporation Information Technology R&D Center)
Keywords:
Semantic Communication,Task-Aware JSCC,Vision-Language Model,Image Transmission
This paper proposes a task-aware image transmission scheme for Joint Source-Channel Coding (JSCC), where transmission performance is evaluated based on downstream recognition accuracy of a Vision-Language Model (VLM) rather than conventional reconstruction quality metrics. The proposed scheme consists of a coarse reconstruction using a base latent representation and a progressive refinement using enhancement latents generated from intermediate encoder features. Simulation-based evaluation on the CIFAR-10 dataset shows that the proposed latent+enhancement scheme outperforms the latent-only scheme in terms of Top-1 accuracy under the same transmission budget, demonstrating the effectiveness of the proposed approach for task-aware image transmission.