Semantic Communication System for Vehicle Detection Based on CNN Encoder and Supernet Architecture

Abstract

This paper proposes a semantic communication system for vehicle detection tasks over noisy wireless channels. A CNN based semantic encoder is designed to compress image features at multiple compression ratios (CH4, CH8, CH16), transmitting only task-relevant information rather than raw pixel data. To reduce model redundancy, a Supernet architecture with shared encoder weights is introduced, supporting all three compression ratios within a single model trained using the Sandwich Rule strategy. Experiments are conducted under AWGN channel conditions across five SNR levels (0-30 dB), evaluated using six detection metrics via the YOLOv12 framework on a multi-class vehicle dataset of 5,171 training images. Results demonstrate that at 0 dB SNR, the traditional transmission method suffers catastrophic performance collapse, while the semantic CH16 encoder retains mAP@0,5 = 0,414 (a reduction of only 8,2% from its 30 dB performance). Furthermore, Supernet_K16 achieves mAP@0,5 = 0,618 at 0 dB, compared to 0,627 for the independently trained CH16, a gap of only 1,4%, demonstrating that a single Supernet can effectively replace three independent models with negligible performance loss.