Banknote identification plays an increasingly important role in financial fields due to the diffusion of automatic bank systems in terms of vending machines. Nowadays, YOLOv5 has become the state-of-the-art detector of visual objects because of its relatively outperformed accuracy with a high speed of computing. In this chapter, the squeeze-excitation (SE) attention module is mingled with the terminal of the backbone in YOLOv5 to further improve visual watermark recognition of paper banknotes. The main contribution of this chapter is that the excellent precision reaches 99.99% by utilizing the novel model YOLOv5+SE.
TopIntroduction
Despite the escalating commence of electronic currency, nowadays, banknotes remain galore owing to the indispensability in circulation, which means currency issuers have still confronted the menace of forging. With the prevalence of automated systems such as vending machines, currency recognition has become increasingly significant in a number of financial sectors such as currency exchange centers, shopping malls, banking systems and ticket counters (Mittal, 2018). Meanwhile, fraud techniques have been increasingly, resulting in the light of recognizing fake currency (Zhang & Yan, 2018; Yan, 2021). Besides, numerous nations suffer from the forged currency on a large scale due to its ease of printing (Trinh et al., 2020). Hence the identification of counterfeit currency has become one of the most redhot topics.
As a genre of image classifications in the computer vision, currency recognition is defined as the process of identifying the denomination and the authenticity of currency (Singh et al., 2010). In order to effectively determine its credibility, it is necessary for banknotes to be inspected for several specialized security features involving serial number, puzzle number, the color-changing bird, raised ink and transparent window. There are a vast variety of methods to detect currency that majorly consists of digital image processing, machine learning, and deep learning algorithms.
In recent years, deep learning has boomed in image classification and detection areas. As a kind of machine learning methods, they take use of a neural network framework consisting of multiple layers that are mainly constructed to perform classification tasks directly from sounds, images, and textures. Deep learning approaches exceed conventional machine learning algorithms in precision and accuracy, though they require much data and training time. Another contributing factor of deep learning for being a popular technology in computation is that the complexity is increasingly declined with the enhancement of data and the layers of a neural network. There are various deep learning architectures in terms of VGG (Simonyan & Zisserman, 2015; Russakovsky et al., 2015), YOLOv5 (Jocher, 2020), Faster R-CNN (Ren, et al., 2015), AlexNet (Krizhevsky et al., 2017) and GoogleNet (Szegedy et al., 2015), which are utilized to find patterns from training data.
A state-of-the-art algorithm is selected to perform this task in this chapter. YOLOv5 is an appropriate model because of its excellent performance on object detection, acceptable precision, the first implementation on currency recognition with an attention mechanism.
Therefore, the focus of this research project is on visual watermark recognition of paper currencies through implementing deep learning algorithms involving YOLOv5 and its variants (YOLO-SE), which comprises of the SE attention block. Remarkably, the experiments have the huge size of the dataset to improve the precision and generalization ability. Hence, data augmentation consisting of cropping, flipping, rotation, colour modification, and noise addition was implemented.
Figure 1.
The security features of currency. The highlighted window is the target for object detection.
This book chapter aims to achieve currency identification based on the transparent window whose phases are separated into data collection, data augmentation, denomination recognition and the analysis of the outcomes. The contributions in this research are majorly summarized into three-folds: The construction of comprehensive samples, the proposal of YOLOv5-SE, and the result analysis.
First of all, regarding the requirement of dataset in deep learning, the samples in this experiment involve the front and back sides, the changes of location, size, and others. As a result, we create a relatively full-scale dataset, which is beneficial for the experiments.