HAT Super-Resolution + PARSeq+CLIP4STR Ensemble Nails Extreme License Plates

Researchers Karthik Sivarama Krishnan and Koushik Sivarama Krishnan have developed a novel system for extreme in-the-wild license plate recognition, submitted to the ICIP 2026 Grand Challenge on Extreme In-the-Wild License Plate Super-Resolution (XLPSR). The system treats license plate super-resolution not as an image enhancement task but as a recognition problem gated by legibility. It first applies a Hybrid Attention Transformer (HAT) super-resolution front-end to lift characters out of sub-pixel territory, then feeds the output into an ensemble of two complementary scene-text recognizers: PARSeq-S and CLIP4STR-B.

The key innovation is a confidence-weighted character-voting scheme that explicitly exploits the challenge's asymmetric scoring rule (+2 for correct character, -1 for wrong, 0 for abstention). By abstaining on character positions where the ensemble is uncertain, the system avoids penalties and boosts its overall score. On the public validation leaderboard, the pipeline achieved a wECR (weighted Edit Cumulative Reward) of 9.73. Performance is also real-time capable, with an average inference time of 1.7s per sequence on an RTX 3090 (max 2.7s, p99 2.4s), well under the 60s/sequence Docker budget. The work was accepted as a top-8 finalist at the IEEE ICIP 2026 Grand Challenge and is described in a 2-page paper on arXiv.