Publication

Abstract : Semantic communication is an emerging paradigm aimed at reducing bandwidth by transmitting task-relevant features instead of raw data. This work presents a semantic segmentation-based communication framework using the DeepLabv3+ architecture with a MobileNetV2 backbone for efficient feature extraction. The encoder generates latent representations, including low-level and bottleneck features, which are transmitted across the channel instead of full-resolution images. At the receiver, the features are upsampled, fused, and refined through convolutional layers to reconstruct semantic maps. To assess communication performance, a simulated noisy channel with bit flips and random packet drops is introduced. A baseline without error protection is compared to Reed–Solomon (RS) coding. Results show that RS coding enhances robustness, preserves semantic accuracy, and ensures efficient, resilient transmission.