Peruvian License Plate Recognition at the Edge: Compact LPRNet and CRNN Recognizers Distilled from a Rule-Corrected Generic OCR

High recognition rates reported for automatic license plate recognition (ALPR) rarely survive deployment. We analyze a fixed-camera edge ALPR system that ran without interruption for 465 days at the vehicle entrance of a university on the Peruvian altiplano and logged 579,292 crops from 29,928 unique plates over 294,900 vehicle passes. Because the camera is mounted away from the lane, the median plate measures 101 px wide by 51 px high, a resolution at which character segmentation is no longer reliable and the whole string must be read at once. The pipeline is a two-stage one. A YOLOv8n detector locates the plate on the full frame at 1.1 ms per image; the native crop is then read by PaddleOCR PP-OCRv4 reinforced with a deterministic corrector that encodes the Peruvian LNN-DDD pattern, and that pairing acts as a weak teacher producing silver labels at corpus scale. From those labels, we distill two compact recognizers on a unified 32 × 100 px input, LPRNet (0.97 M parameters, purely convolutional) and CRNN (8.73 M, two additional bidirectional recurrent layers), both trained with connectionist temporal classification (CTC) loss over 37 classes and decoding in parallel without character segmentation, and both fine-tuned on 1850 crops verified against the official vehicle registry. How the evaluation set is composed decides the conclusion. On 782 crops sampled by logged access, the teacher reached 95.4% and the student 97.4%, whereas on 1130 crops of arbitrary plates, the teacher collapsed to 54.7% while the student held 91.9% (+37.2 points; p = 7.4 × 10−93). The measured cause is that automatic label reliability grows with the number of times a vehicle passes the gate, rising from 40.4% on plates seen once or twice to 87.6% on those seen fifty times or more, a dependency the distilled model does not inherit. A rule-corrected optical character recognition (OCR) engine is, therefore, sufficient as a teacher for distilling edge-ready recognizers, and stating the sampling unit of the evaluation set is a precondition for interpreting any ALPR metric.