RMSProp Optimizer and KAN Method-Based CNN on Rupiah Banknote Classification for Visually Impaired

Abstract

The visually impaired refers to individuals who experience a loss of visual function. Approximately 4 million people, or about 1.5% of Indonesia's total population, are visually impaired. They rely on their sense of touch to recognize banknote denominations in financial transactions. However, damaged banknotes often hinder identification and increase the risk of fraud. Therefore, this study aims to develop a rupiah banknote denomination classification model to assist them in conducting independent transactions. The researchers developed a CNN-KAN model with the RMSProp optimizer using a private dataset comprising 800 images of Rupiah banknotes with denominations of IDR 1,000, IDR 2,000, IDR 5,000, IDR 10,000, IDR 20,000, IDR 50,000, IDR 75,000, and IDR 100,000 from the 2016, 2020, and 2022 emission years. The dataset encompasses variations in image perspectives, lighting conditions, and the physical state of banknotes, including both intact and damaged ones, with up to 30% of the samples comprising damaged banknotes. Data augmentation techniques were implemented to improve data diversity. The dataset was then utilized for training and testing with different split ratios: 50:50, 60:40, 70:30, 80:20, and 90:10. Performance evaluation was conducted using loss, accuracy, precision, recall, and AUC-ROC metrics. Experimental results indicate that the CNN-KAN model with the RMSProp optimizer achieved optimal performance. In the 90:10 data split scenario, the model achieved 100% accuracy, precision, and recall, with an AUC-ROC of 1 and a loss of 0.008. Therefore, the CNN-KAN model with the RMSProp optimizer has been proven effective for implementing Rupiah banknote denomination detection for the visually impaired in an automated system.