Developing a Model for the Detection of Ethiopian Fake Banknotes Using Deep Learning

Various methods for recognizing and detecting fake banknotes have recently become a major concern in finance and business. Fake detection is an increasingly important approach due to the significance and technological complexity of quickly and accurately processing large amounts of banknote image data, presenting a significant opportunity for data exploration. A proposed deep CNN technique utilized a detection method to distinguish between real and fake banknotes. This is achieved by examining the computer vision features of digital content to detect banknotes using smartphone cameras in a cross-dataset environment. The banknote detection model begins with preprocessing steps: image acquisition, size normalization, grayscale conversion, and histogram equalization. The processed grayscale image is fed into an Input Layer (zero parameters), followed by 14 Convolutional Layers and two Fully Connected Layers. This architecture, designed to reduce computational complexity and enable hardware deployment, results in a significantly smaller model than the VGG19 baseline (which has a million parameters). The customized model achieves high performance despite the parameter reduction, demonstrated by its impressive results on the dataset split into 80% training, 10% validation, and 10% testing datasets. After an iterative experimental process, the proposed model achieved 99.9% training accuracy, 99.4% validation accuracy, and 97.6% testing accuracy.