Aging and soiling simulators for iproved banknote classification and validation

다국어 초록 (Multilingual Abstract)

In this dissertation, efficient banknote aging and soiling simulator for generating aged and soiled banknotes from new banknotes are proposed. In artificial banknote aging and soiling simulation, it is important to generate detailed aging and soiling patterns and brightness changes which are similar to real-world circulated banknotes, since the recognition accuracy of circulated banknotes is becoming decreased in automated teller machines (ATM) trained with new banknotes. In order to generate aging and soiling details efficiently, deep convolutional neural network with residual learning is proposed for banknote soiling. Also, Gaussian brightness model and optimal Perlin noise map are proposed for banknote aging.
Although financial automatic machines were trained with new banknotes, they cause false acceptance and rejection problems to classify or validate banknotes after circulating several months due to some aging and soiling effects in circulated banknotes. Also, it is essential to analyze the characteristics of aged and soiled banknotes for efficient banknote recycling of circulated banknotes. However, additional works to train automatic machines with aged and soiled banknotes are time-consuming. Although previous studies provide several circulation simulators using chemical and mechanical operations, it is rather expensive and memory-inefficient. Therefore, low-cost and memory-efficient simulators for artificial banknote aging and soiling from new banknotes are needed, but there is no common solution that simulates artificially aged and soiled banknotes based on machine learning techniques and image processing. Efficient banknote circulation simulators must consistently produce the characteristics of circulated banknote image such as creasing, ink wear, edge blurness, stain, irregular dirt, etc.
An efficient banknote simulation method that consists of two main parts is proposed: 1) banknote soiling simulator based on deep convolutional neural network with residual learning and batch skipping, and 2) banknote aging simulator based on adaptive Perlin noise map and Gaussian brightness model for artificial aging effects such as stain, irregular dirt, etc. In the residual learning based neural networks, an output banknote image is composed of an input banknote image (corresponding to relatively new banknote image) and residual image (detailed information of soiled banknotes), which can efficiently simulate detailed soiling patterns in rarely soiled banknotes. For naturalness and performance improvement, the optimal Perlin noise parameter values are determined according to the characteristics of each denomination.
The proposed banknote simulators lead to perform better than existing methods in banknote recognition and counterfeit banknote detection for aged and soiled banknotes in multi-currency environment. The experimental results show that the performance for aged and soiled banknotes has been noticeably improved in terms of evaluation scores for banknote classification and validation. When banknote classification and validation algorithms are trained by using the proposed banknote simulators, it becomes more robust to recognize highly aged and soiled banknotes than conventional algorithms.
Additionally, according to the inputted banknotes, progressive learning methods are applied to the parameters used for banknote classification and validation, which are automatically updated from pre-trained parameters. In banknote classification algorithm, when a bundle of banknotes is inputted, the template images used to calculate correlation coefficient of target banknote are re-generated. In banknote validation, the reference parameters for calculating evaluation score are updated by using Bayesian learning method. The progressive learning method provides additional performance improvement comparing with only using pre-trained parameters from the proposed banknote simulators.