Machine Learning Framework for Extraction and Recognition of Nigerian Banknote Serial Numbers

Abstract

The traceability of physical currency remains an essential component of financial security systems, particularly in economies where cash transactions are prevalent. Serial numbers printed on banknotes provide unique identifiers that can support forensic auditing, fraud detection, and circulation monitoring. However, automated extraction and recognition of these identifiers are technically challenging due to background interference, substrate variability, typographic inconsistencies, and physical wear. This study proposes a tailored machine learning framework for the extraction and recognition of serial numbers on Nigerian Naira banknotes. The framework integrates adaptive preprocessing, connected component-based region localization, structured character segmentation, engineered feature representations, and supervised classification. Five handcrafted feature strategies were comparatively evaluated alongside a convolutional neural network (CNN) trained directly on raw image patches. Using a dataset comprising 2,199 segmented characters across eight denominations, experimental analysis demonstrated that serialized pixel representation combined with an Artificial Neural Network achieved 99% classification accuracy. A CNN architecture attained equivalent performance without explicit feature engineering. The findings indicate that deep learning approaches provide strong generalization capability for financial document recognition and offer practical utility for automated currency monitoring systems.

Keywords:    Serial Number Recognition; Currency Authentication; Deep Learning;

                       Feature Engineering; Nigerian Naira; Computer Vision.