Impact of image pre-processing on enhancing paddleOCR performance for automatic indonesia number plate recognition

Automatic Number Plate Recognition (ANPR) systems use Optical Character Recognition (OCR) to accurately retrieve license plate information. While models like PaddleOCR function well in ideal conditions, their accuracy drops significantly in real-world scenarios with poor lighting, low resolution, or skewed angles. This study compares the effects of five preprocessing techniques which are grayscale, bilateral filtering, Otsu thresholding, adaptive thresholding, and CLAHE along with its baseline model on OCR performance for Indonesian license plates. Using a dataset of cropped plate images taken under various challenging conditions, we assess each method based on the exact match rate percentage, average edit distance, and average character accuracy. The results demonstrate that bilateral filtering produces the smallest average edit distance, whilst Otsu thresholding produces the maximum character accuracy. However, severe angles and unclean or broken plates continue to be important causes of recognition failure across all approaches. Our findings highlight the importance of preprocessing in enhancing OCR robustness and offer practical insights for deploying ANPR in uncontrolled environments.