Adaptive Multi-Variant Enhancement and OCR Selection for Robust License Plate Recognition in Hazy Conditions

Abstract

Automatic License Plate Recognition (ALPR) systems often experience performance degradation under adverse visual conditions such as haze, low contrast, and limited image resolution. This study proposes an adaptive ALPR framework that integrates Dark Channel Prior (DCP)-based dehazing, YOLOv11-based license plate detection, lightweight multi-variant image enhancement, and plate-aware OCR selection to achieve robust recognition under hazy conditions. After dehazing and plate detection, each cropped license plate image is processed with three enhancement variants: RAW, CLAHE-light, and SR2x + CLAHE-light. Each variant is recognized with PaddleOCR to generate multiple OCR candidates, which are then evaluated with a plate-aware scoring mechanism that combines OCR confidence, structural validity, and ambiguity penalties. Experimental evaluation on 945 hazy license plate images achieved a readability rate of 100.00%, an exact plate accuracy of 93.12%, and a character-level accuracy of 98.63%. Ablation analysis showed that the proposed framework improved exact plate accuracy by 3.70 percentage points over the strongest single-variant baseline, and statistical significance testing confirmed that the improvement was not due to random variation. Parameter sensitivity analysis further demonstrated that the proposed scoring mechanism remained stable across different weighting configurations. Although the complete pipeline requires an average processing time of 471.42 ms per image (2.12 FPS), the results indicate that adaptive multi-variant enhancement combined with structure-aware OCR selection provides a robust and accurate solution for ALPR under degraded hazy conditions without requiring retraining of either the detector or the OCR model.