Implementasi CLAHE dan YOLOv11 untuk Pembacaan Pelat Nomor Kendaraan Beresolusi Rendah pada Raspberry Pi 5

Abstract

Vehicle license plate recognition in bus fleets often faces obstacles due to small plate size, low image resolution, and poor contrast, making characters difficult to recognize. This study proposes a low-resolution bus license plate character reading system using the You Only Look Once version 11 (YOLOv11) model combined with the Contrast Limited Adaptive Histogram Equalization (CLAHE) method as a preprocessing stage in the inference process. The training dataset was developed from a combination of synthetic license plate images and low-resolution real vehicle license plate images. The model was trained using an NVIDIA A100 GPU for 500 epochs and implemented on a Raspberry Pi 5 as an edge device. Validation results show the model achieved a precision of 88.9%, a recall of 76.2%, an mAP@50 of 87.2%, and an mAP@50–95 of 53.4%. Testing on real vehicle license plate images shows that the application of CLAHE increases the success rate of character reading from 82.13% to 91.04% with a relatively small additional inference time. Furthermore, implementation on a Raspberry Pi 5 demonstrates that the model can be run efficiently with low resource usage, making it worthy of consideration as an Artificial Intelligence of Things (AIoT)-based Automatic Number Plate Recognition (ANPR) solution.

GS Cited Analysis