Developing Ethiopian Paper Currency Recognition Model Using Convolutional Neural Network (Cnn)

Abstract

In Ethiopia, Financial Institutions Have Adopted Various Transaction Facilities Such As Banknote Counting Machines, Multi-Currency Detecting Machines, Automatic Teller Machines (Atms), Banknotes Reader And Sorters, Etc. Using Currency Recognition As Their Main Activity, Which Makes Automated Currency Recognition Of Significant Interest. For These Devices To Work Properly Banknote Recognition Is A Mandatory Feature. In Addition To These, There Are 5.3% Of Visually Impaired People In Ethiopia According To A Survey Report Made By The Ethiopian Ministry Of Health. These People Have Faced Problems In Currency Recognition.In This Study, We Have Reviewed Literature Related To Ethiopian And Other Country's Banknotes Recognition. In The Case Of Ethiopian Banknote Recognition And Classification, There Is Little Research Work On Old Ethiopian Banknotes. There Have Not Been Found Any Methods Implemented Or Proposed For The Recognition Of Newly Released Ethiopian Banknotes. Therefore, This Thesis Presents A Deep Learning-Based Method For Newly Released Ethiopian Paper Currency Denominations From Their Color Images. A Classification Framework Has Been Implemented Using The Concept Of Transfer Learning Where A Large Convolutional Neural Network Pre-Trained On Millions Of Natural Images Is Employed For The Classification Of Images From New Classes. An Image Dataset Of Four Banknote Denominations Is Prepared By Preprocessing And Augmentation Of Real-Banknote Images Acquired In Different Viewpoints And Lighting Conditions Via A Scanner And Smartphone Camera. A New Top Layer Upon The Convolutional Base Of A Pre-Trained Mobile Net Model Is Trained For A Few Epochs Upon A Portion Of The Dataset To Achieve An Agreeable Accuracy.All The Codes Are Implemented Using Python Programming Language. In The Case Of The 80/10/10 Dataset Proportion An Average Training Accuracy And Average Validation Accuracy Of 99.7%, 99.5% Are Obtained Respectively, And In The Case Of 70/15/15 Dataset Proportion An Average Training Accuracy And Average Validation Accuracy Of 97.8%, 99.0% Are Obtained Respectively. Therefore, As The First Research Work, The Model Has Shown The Best Performance In Both Dataset Categories But The 80/10/10 Dataset Ratio Provides A Promising Result With An Average Training Accuracy Of 99.7%.