Ethiopian Banknotes RecognitionUsing Convolutional Neural Network

Abstract

Money transactions can be performed by automated self-service machines like ATM for moneydeposits and withdrawals, banknote counters and coin counters, automatic vending machines,and automatic smart card charging machines. These devices must be equipped with four essentialfunctions: banknote recognition, counterfeit banknote detection, serial number recognition, andfitness classification. Therefore, we need robust system that can recognize banknotes and classifyinto denominations that can be used in these automated machines. However, most widelyavailable banknote detectors are hardware systems that uses optical and magnetic sensors todetect and validate banknotes. These banknote detectors are usually designed for specific countrybanknotes. Reprogramming such system to detect our country, Ethiopia, banknotes is verydifficult. In addition researchers have developed Ethiopian banknotes recognition system usingDeep Learning Artificial Intelligence technology like CNN, and R-CNN. However, in thesesystems dataset used for training is relatively small, and accuracy of banknotes recognition isfound smaller. The existing systems also not included implementation using embedded system.In this research work, we collected various Ethiopian currencies with different ages andconditions and applied various optimization techniques for CNN architects to identify the bestoptimization technique and CNN model with best accuracy. Experimental analysis demonstrated,MobileNet with RMSProp optimization technique in batch size 32 is a robust and reliableEthiopian banknote detector with best accuracy. Selected model is implemented on Embeddedsystem (Raspberry Pi 3 B+) and Web based User Interface (UI) is developed and verified.