Vector Quantization-Based Opportunistic Semantic Communication for Object Detection

This paper proposes a communication-efficient generative AI framework for object detection through semantic communication. Specifically, the network consists of a transmitter (Tx) and a receiver (Rx), each equipped with the encoder and decoder of a VQ-VAE-2 model. The raw image observed at the Tx is compressed into discrete codebook indices, which are then transmitted to the Rx, where they are decoded and used to determine the presence or absence of an object. While transmitting discrete representations instead of raw images improves communication efficiency, this alone is insufficient to meet the stringent latency and bandwidth requirements of time-critical object detection. To further enhance communication efficiency, we propose an opportunistic semantic communication scheme that employs thresholding based on the distributional distance between the codebook index distributions of samples with and without target objects. Furthermore, we formulate an optimization problem to minimize the object detection error under a given communication budget by tuning this threshold. Numerical evaluations corroborate that our method achieves only 45.5% of the transmission cost compared to its full-transmission counterpart, while preserving 99.98% SSIM and improving PSNR by 1.28% compared to the baseline, thereby validating the effectiveness of the proposed opportunistic transmission strategy.