WINK Engineering set out to test whether neural super‑resolution could boost a custom license‑plate recognizer in 2026. The team trained a tiny 4× upscaler, then ran it on more than 5,000 sub‑100 px crops drawn from an 18 k‑image dataset. Both the home‑grown model and a 1.2 M‑parameter pretrained Real‑ESRGAN produced identical OCR results, offering no accuracy gain in real‑world deployments.

The pipeline kept every other variable constant: the same CTC‑CRNN OCR (98.6 % baseline), identical crop resizing, and a two‑millisecond inference budget. Adding the 42 K‑parameter VGGNetCompact upscaler increased total latency to roughly 7 ms but raised exact‑match scores from 0 % to a flat 0.4 % across all size buckets. The massive 1.21 M‑parameter version behaved the same, confirming that model capacity and training data did not matter.

Because the system aggregates 15‑20 crops per vehicle, high‑resolution distant crops are outvoted by clear close‑range images, so the SR step merely injects plausible but wrong characters into the voting pool. WINK Engineering concludes that unless a commercial OCR cannot be retrained, super‑resolution adds engineering overhead without improving plate‑read rates. Deploying end‑to‑end models remains the simpler, reliable path.