Neural Super-Resolution as a Pre-Filter for License Plate OCR
Summary
The article evaluates neural super-resolution as a pre-filter for license plate OCR and reports that SR did not improve recognition accuracy across multiple pipelines, including a small SR model and a much larger pretrained model. It argues that training OCR on low-resolution data with multi-scale augmentation and leveraging multi-crop voting yields stronger performance, and that SR can introduce hallucinated characters and unnecessary latency. The piece provides a nuanced view of when SR is actually beneficial and emphasizes data quality and end-to-end OCR training over upscaling in typical production scenarios.