Classification of Rupiah Banknote for the Visually Impaired using Convolutional Neural Network with Custom Architecture and RMSProp Optimizer

Visual impairment is a condition that affects the sense of sight. In Indonesia, around 1.5% of the 4 million people are blind. In their daily activities, the visually impaired rely on their sense of touch to recognize the value of money by feeling the double lines on the right and left sides of the rupiah banknote. However, if the money is in bad condition, they are unable to recognize it, so they have to involve the help of others with the risk of fraud. Research is underway to develop a paper currency detection tool to assist the visually impaired in everyday transactions and reduce the risk of fraud. The research in this paper focuses on creating a classification model for detecting Indonesian rupiah banknotes. The method approach used in this research involves using a custom architecture Convolutional Neural Network model that combines elements from the AlexNet and VGGNet architectures, along with the RMSProp Optimizer. Before training the model, the private dataset that has been collected is augmented to increase the dataset and then split data. Next, the model is trained and tested using a private dataset of 800 images of rupiah banknotes for 2016, 2020, and 2022 emissions years, divided into eight classes of banknotes with denominations of IDR 1000, IDR 2000, IDR 5000, IDR 10,000, IDR 20,000, IDR 50,000, IDR 75,000, and IDR 100,000. Finally, the model's performance is evaluated using loss function, accuracy, precision, recall, and AUC-ROC evaluation metrics. The model evaluation results show a value loss function of 1.28, an accuracy of 95.31 %, a precision of 95.24%, a recall of 93.75%, and an AUC-ROC value of 0.99. This evaluation result indicates that the model is highly effective for use in the rupiah currency nominal detection system for the visually impaired.