Ethiopian Original, Different Side and Counterfeiting Bank Notes of Both Side Verification Using Deep Learning Approach

Abstract

Counterfeit currency is a growing concern in Ethiopia, affecting economic stability and public trust. This study presents a deep-learning system that uses Convolutional Neural Networks (CNNs) to detect counterfeit Ethiopian banknotes of 100-birr and 200-birr denominations. Both front and back sides of genuine and counterfeit samples are analyzed. A dataset of 1,500 images was augmented to 7,800 and split into eight classes. Three pre-trained CNNs—MobileNet, DenseNet121 and InceptionV3—were fine-tuned and compared with a custom CNN. MobileNet achieved the highest accuracy (96.58%), followed by DenseNet121 (95.17%) and InceptionV3 (93.99%). The custom CNN reached 94.20%. The results demonstrate the suitability of transfer learning for scalable and accurate currency authentication in banking and commercial applications.