Annotation-Free Fitness Grading of Nigerian Naira Banknotes via Pretrained Feature Clustering
Year: 2026
Type: conference-paper
Abstract: The assessment of banknote fitness, defined as the determination of whether a note is suitable for continued circulation, remains an understudied problem for the Nigerian Naira, where no publicly available fitness-labeled dataset exists and prior work has concentrated on denomination recognition and counterfeit detection. This study proposes an unsupervised approach to Naira fitness grading that operates without fitness annotations. A working dataset of 5,446 images was consolidated from six sources and filtered to 5,349 for surface-based condition analysis. Grayscale images were encoded using a frozen EfficientNet-B0 network pretrained on ImageNet, and the resulting embeddings were grouped into two clusters with k-means, corresponding to the fit and unfit classes defined by the Central Bank of Nigeria's Clean Note Policy. The groupings were evaluated against a balanced set of 200 manually labeled images by aligning each cluster to a fitness class after clustering and reached a balanced accuracy of 0.950 that was identical across ten clustering runs and under two initialization strategies. A source-controlled evaluation confirmed that the agreement reflected physical condition rather than imaging modality. A comparison with a ResNet-50 backbone under the same pipeline produced substantially lower and less stable agreement, indicating that the separation depends on the choice of pretrained representation.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225757934