Publications

Noisy, low-res license plates reduce OCR accuracy in real-world traffic.

Built a GAN pipeline combining segmentation, recognition, and super-resolution with TV-regularization.

Generated sharper plates with higher OCR accuracy, PSNR, and SSIM vs. baselines.

Improves ALPR reliability under challenging lighting, motion, and weather conditions.

ALPR must run in real time on constrained edge hardware.

Benchmarked classical vs. DNNs; measured accuracy, latency, memory, and FPS.

DNNs improved accuracy; lightweight models + pre/post-processing met real-time.

Actionable guidance to choose models/datasets for deployable edge ALPR.

Block-based motion models struggle at ultra-high resolutions.

Applied HASM® mesh motion; Python simulations on 1080p/4K/8K sequences.

Stable motion fields; cases where mesh outperformed block motion.

Guides codec design and encoder settings for UHD content.

Static power allocation wastes energy under variable traffic.

Learned traffic signatures; clustered sites; adjusted per-hour power/links.

Lower energy and resource use while maintaining QoS.

Enables greener, adaptive RAN operations and cost savings.