WINK Engineering tested two SR models as OCR pre-filters on 2,000 sub-100px production plate crops and found zero accuracy gain at any model size.
Key Takeaways
A custom 42K-parameter SRVGGNetCompact and a 1.21M-parameter Real-ESRGAN pretrained model both produced identical results: 0.0% exact match, 0.4% character accuracy on low-res crops.
The root cause is hallucination: SR generates plausible but wrong characters (8/B, 0/D, 7/T confusion), polluting the multi-crop voting pool with confident garbage.
Multi-crop voting already solves the problem. 15-20 crops per vehicle let large clear crops (100-200px) outvote unreadable small crops; SR on small crops changes nothing.
The fix is free: train the OCR model with multi-scale augmentation on actual low-res crops. The model reads blurry inputs directly without a second model or hallucination risk.
SR as a pre-filter only makes sense when you cannot retrain the OCR model, such as with a vendor-locked commercial API or legacy black-box system.
