Semantic-Preserving Image Compression and Restoration Pipeline for Bandwidth-Constrained UAV Applications: A Task-Aware Evaluation Framework

作者：F. Hossain, Noman Saffat Sajid, Rashedur M. Rahman · 发表于：Artificial Intelligence and Applications · 年份：2026 · DOI：10.47852/bonviewaia62028610

Unmanned aerial vehicles (UAVs) face critical challenges in transmitting high-resolution imagery over bandwidth-constrained communication channels, particularly in time-sensitive applications such as search and rescue and surveillance. This paper presents a data-efficient image transmission and enhancement pipeline addressing the trade-off between transmission efficiency and visual quality preservation. Our three-stage framework consists of onboard k-means color quantization (16-color palette), efficient transmission over bandwidth-limited channels, and ground-station deep learning-based restoration using CCDNet with iterative residual learning. The pipeline achieves approximately 68% file size reduction and 3× faster transmission times under realistic network conditions, validated through probabilistic network simulations. Beyond perceptual quality metrics (PSNR: 35.01 dB, SSIM: 0.9430), we demonstrate real-world applicability through downstream task evaluation. Object detection using YOLOv8 shows restored images achieve 63.0% mAP@0.5, significantly outperforming JPEG-compressed images at 49.0% mAP while maintaining the same compression ratio—a 14.0 percentage point improvement with 77.8% retention of original performance. Zero-shot cross-dataset evaluation demonstrates strong generalization: the pipeline trained on Semantic Riverscapes achieves 89.94% PSNR retention and 85.06% detection performance retention on VisDrone2019-DET without model retraining. Computational analys...