Ijraset Journal For Research in Applied Science and Engineering Technology

Abstract

For visually challenged people, distinguishing between different denominations of cash is a difficult process. Even though unique symbols are engraved on various currencies in India, the work is still difficult for blind individuals. The inadequacy of identifying devices prompted the development of a portable gadget for denominational segregation. This project aims to create an Android application that will assist visually and hearing-impaired people in detecting Indian cash denominations by putting a banknote in front of the camera. The work uses machine learning and Android programming approaches and is based on real-world applications. The android application uses text to speech concept to read the value of note to the user and then it converts the text value into speech. To harness the power of them all, we are leveraging the Keras, TensorFlow, Fastai, and PyTorch libraries, as well as different machine learning techniques like ResNet and MobileNet. Various technologies like machine learning models, python and many more libraries are used for the backend part of application. And for front end, java concepts and android development techniques are employed. Altogether they are integrated into a single platform which is highly user-friendly and makes it easy to use and implement in our daily life.

Introduction

I. INTRODUCTION

According to the most recent estimate from the World Organization, approximately 2.3 billion people across the world suffer from visual impairment or blindness, with at least 1 billion of them suffering from impaired vision that might have been prevented or that has yet to be addressed. Visually challenged people experience several challenges in conducting regular tasks. They have number of issues with money transactions as well. They are unable to distinguish the paper currencies due to the similarities in paper structure and size between the various types.

One of the most fundamental and crucial systems that helps a person with an impairment operate around his or her problems is assistive technology. This study shows how forward-thinking attempts are being made to build assistive technologies for the visually impaired so that they can live a socially and financially independent life. The currency denomination identifier application is an artificially intelligent currency detection app that acts as an assistive tool for the visually impaired to check whether they have been handed the correct amount of money and, thus, ensure that they have not been cheated on. It outputs computer-generated audio and has a simple user interface for a better user experience.

II. LITERATURE SURVEY

1. Currency Detection for the Visually Impaired Shweta Yadav, Zulfiakr Ali Ansari, and Kaushiki Gautam Singh have studied various identification techniques for detecting currency denominations. The suggested system covers six different forms of currency papers. In their study, the YOLO-V3 approach was applied, and the input picture was pre-processed and converted to greyscale. After each feature is clustered one by one, the Sobel technique is used to extract the image's edges.
2. In 2019, Prakhar Chaturvedi et al. studied various digital processing techniques and proposed a system that leverages the use of both radial basis functions and image processing techniques. The scanned input image is transformed to a digital image, then image analysis, pre-processing, segmentation, edge detection, and pattern matching are all conducted using Radial Basis Functions.
3. In 2019, Prashengit Dhar et al. created a paper currency detection system that can recognise paper currency from an image. They took the SURF and LBP features of currencies, merged them, and used SVM classifiers to train them. The proposed system employed a large number of classifiers and was also capable of detecting paper currency in rotational positions.
4. In 2018, Shaimaa H. Shaker et al. proposed a system based on image processing techniques and algorithms. The used programming software would extract the features from the test photos, and the retrieved features would be compared to the ones saved in the MAT file. The training image features are saved in a MAT file, and the test image features are compared to them, yielding the money denomination.
5. Baharaini Paper Curreny Recognition Ebtesam Althafir, Muhammad Sarfraz, and Muhannad Alfarras designed and proposed a system to recognise paper currency. They employed the Weighted Euclidean Distance Neural Network technique, which aids in recognising currency from both sides with the help of two classifiers. The suggested method is based on extracting some unique characteristics of paper currency while taking into account several aspects such as picture size, edge enhancement, and correlation analysis, all of which are significant in the recognition process.
6. In 2010, Junfang Guo et al. proposed an approach for recognising paper currency that is based on the standard local binary pattern (LBP) method and an upgraded LBP algorithm known as the block-LBP algorithm. The author's suggested android paper money identification system had a greater retention rate and was simple to use, but it had certain restrictions because it was only applicable to Saudi Arabian papers. With a two-phase classification strategy employing template matching, methods for recognising paper cash mainly depended on some properties and correlations between two currency pictures.
7. In 2009, Hamid Hassanpour et al. using Hidden Markov Models, he devised a method for recognising paper cash from various nations. Their approach is based on three features of currency notes, including size, colour, and texture, and it is a valuable tool in the modelling of random processes. They showed a paper money model with texturing based on the Markov Chain idea.

III. ARCHITECTURE AND LIMITATIONS

When developing on a mobile platform, there are a few things to keep in mind. The three primary limits are program size, memory, and processing time. To function without interfering with other programmes, an app should not utilise more than 100 megabytes of storage and 50 megabytes of RAM on a mobile phone. The banknotes are recognised by our application in two processes. To begin with, we separate the bill from the rest of the mess. Then we examine the bill in the information base that is the most practically identical. Several state-of-the-art computer vision algorithms can efficiently handle both of these problems, but they are not mobile-friendly. If implemented directly, the recognition model and other critical information for our application would normally need additional storage and compute capacity. By a significant margin, this surpasses practical bounds. The application's reaction time should be fast and accurate.

The challenge is compounded by the fact that the intended audience is visually impaired. The user has no idea how the surrounding environment is, including other objects, lighting, contrasts, or if the bill is inside the camera's field of vision. When it comes to task execution, the system should be extremely user-friendly and robust. Using the application should be simple with no authentication or login and no internet connection required.

Various security measures recommended by the RBI for currency notes are being evaluated in order to identify the existence of currency denominations, and they may also be expanded to detect whether the money is genuine or counterfeit. Watermark, security thread, and intaglio printing are some of the characteristics used in this project.

A.  METHODOLOGY

The Currency Denomination Identifier Application is based on machine learning and Android development techniques, which are mentioned in the following steps:

1. TensorFlow: TensorFlow is a machine learning framework that automates the entire process. It contains a wide scope of toolsets, modules, and public resources that empower researchers to upgrade the best in class in machine learning. TensorFlow is utilised to fabricate and prepare ML models effectively utilising intuitive significant level APIs like Keras, and it assists with the simple preparation and organisation of models in programs or on mobile devices.
2. Keras: Keras is a profound learning API written in Python that chips away at the top of the TensorFlow AI structure. It was made fully intent on working with expedient trial and error. In Keras, prototyping takes less time. Subsequently, your thoughts will be carried out and conveyed in a more limited timeframe. Additionally, Keras offers a choice of organisation strategies based on the requests of the client. Keras is a quick structure that works on top of TensorFlow. Moreover, it is unequivocally incorporated with TensorFlow, permitting you to quickly make custom cycles. Keras and TensorFlow are two widely used libraries for machine learning and deep learning applications.
3. Fastai: Fastai is a deep learning library that gives professionals access to significant level parts for delivering best in class, bringing about traditional, profound learning spaces quickly and easily. It gives low-level parts that might be consolidated and matched to make new procedures for specialists. It aims to fulfil both objectives while maintaining usability, flexibility, and performance.
4. PyTorch: PyTorch is a free and open-source AI (ML) structure in light of the Torch library and the Python programming language. It's one of the most well-known deep learning research stages. The structure was designed to facilitate the progress from research model to execution. Fastai and PyTorch Lightning are built on PyTorch, but Keras' backend is mostly TensorFlow.
5. MobileNet: MobileNet, TensorFlow's first mobile computer vision model, is optimised for usage in mobile apps. For picture characterization and versatile vision, MobileNet is a CNN compositional model. MobileNet are low-inactivity, low-power models that have been defined to fit different use cases' asset requirements. Characterization, identification, embeddings, and division may be in every way developed on top of them. We are presently using the MobileNet model for this since it is more proficient and quicker on cell phones than ResNet.

The steps involved in the development of the Currency Denomination Identifier Application are as follows:

• Dataset preparation and pre-processing
• Choosing the right model
• Build and train model
• Deploy the model and finally, the output is given in form of computer-generated audio.

The process of separating distinct attributes or traits of a currency note has a direct influence on currency identification ability, and currency recognition is always dependent on the characteristics of a specific country's note. To extract the characteristics, many image processing techniques have been presented over time. The security thread, the note's length and colour, the RBI logo, the identifying mark, and other security measures are among them. The identification of money notes relies heavily on feature extraction. As a result, several feature extraction techniques are applied. In this paper, we will look at feature extraction and identification techniques and libraries.

A. Dataset Preparation and Pre-processing

Aside from the stated value and other data, the different divisions of Indian Rupee notes contrast in size and variety, making them easy to recognise visually. For the visually impaired, however, text and colour are worthless, and the similar dimensions of the several banknotes may cause confusion. There is presently no dataset of pictures of Indian Rupee banknotes in different arrangements that is sufficient for the use cases that a visually impaired user may face. As a result, establishing such a collection was a part of our job. For this initiative, we generated the Indian Currency Dataset, which now has about 200 images per class. The dataset is a fairly large dataset with a wide range of photos.

While collecting the dataset, we examined different banknotes for each denomination, in different indoor and outdoor situations. In terms of lighting, backdrop, and position, this provides a lot of variability to the dataset. The collection includes images of clean and worn-out money, along with ones with scribbles. There are many classes for different denominations of currency pictures (including both the front and back of the currency note), as well as a class for "background." Each class comprises photos with notes positioned in various locations and at varied angles. For “background” class images are taken from ImageNet Samples. In order to get decent performance from the dataset model in real-life, the samples in the dataset should be illustrated and experimented on in all conditions accordingly. 

B. Choosing the Right Model

For the past several years, researchers have been working on fine-tuning deep neural networks to achieve the best mix of accuracy and performance. This becomes significantly more difficult when the model must be deployed to mobile devices while still being high-performing. When developing Seeing AI apps, the currency detection model, for example, must operate locally on the phone. To provide the optimal user experience and without losing accuracy, inference on the phone must have a short latency.

We chose MobileNet for Seeing AI because it is fast enough for cell phones and gives adequate performance based on our empirical tests.

C. Build and Train Model

Since the dataset is of around 200-250 images per class, we employ two techniques to achieve the required solution -

1. Transfer Learning: Transfer learning is a machine learning technique that starts with a pre-trained model and then adapts and fine-tunes it for use in different domains. Because the dataset isn't large enough for use cases like Seeing AI, beginning with a pre-trained model and fine-tuning it later can minimise training time and avoid overfitting. In reality, employing transfer learning typically necessitates "freezing" the weights of a few top layers of a pre-trained model before allowing the remainder of the layers to be trained normally (so back-propagation process can change their weights). It's simple to use Keras for transfer learning: simply set the trainable parameter of the layers you wish to freeze to False, and Keras will stop updating those layers' parameters while keeping back propagating the weights of the other layers:
2. Data Augmentation: Because we don't have enough input data, another option is to reuse as much as feasible existing data. Shifting pictures, zooming in or out of photos, rotating images, and other typical image manipulation methods can all be done simply with Keras.

D. Deploy the Model

We want to run the models locally for applications like Seeing AI so that the app may be utilised even when there is no online connection. The main modules used are:

1. Camera
2. Tflite
3. Tts
4. Path_provider

The backend is made up of the tflite package and a tflite model, as well as labels that define all of the classes. The Tts module converts text to speech and is used here so that the result may be accessed as an audio file. The MobilenetV2 version can also be used as it is a newer version of the previous V1 version. The test results are 35% faster than the V1 version and its accuracy has also improved over the previous one.

V. EXPERIMENTATION AND RESULTS

In an assortment of studies, bills were held at various ranges and areas and compared to the device's camera in a variety of trials. They demonstrate that for the whole arrangement, solid outcomes are by and large recorded assuming the bill picture incorporates something like 40-60% of the general region of the procured picture, or on the other hand, in the event that the bill is a way off, not more than an arm's manageable distance from the camera.

We further attempt to illustrate why our strategy doesn't work in unambiguous conditions (see Figure 4 for instances of failure). Each banknote has an image of Mahatma Gandhi written on the front-side (front) (see Figure 4). When examining the bill's half-fold, there are few distinguishing features, and even fewer when the user's fingers cover a section of the surface area. Because colour is so susceptible to light and fading, it isn't a dependable feature in this situation. As a result, such postures frequently provide inaccurate or unclear outcomes. When various denominations of money are in the camera's view, the result is sure to be unclear because the person may be oblivious to his or her surroundings.

The test scenarios presented above are related to the suggested method's real-time operation. More data is necessary to increase the models' performance. We should study new techniques to synthesise more data while collecting more real-life data. In practise, we train the algorithm on both real and synthetic data, then test it against the real-world data we gathered.

Conclusion

Using our technique, we were able to get a recognition accuracy of 94.6 percent on an experimental collection of Indian rupees, and the mean computation rate was faster on a standard smartphone. We were successful in achieving our aim of building a system that could be used by visually impaired people to recognise currency. We moved the system to a mobile platform, overcoming obstacles such as limited computational power and storage while retaining high accuracy and quick announcement times. In the great majority of circumstances involving images taken on a cell phone, the approaches used are effective.

References

[1] Kaushiki Gautam Singh, Shweta Yadav, Zulfikar Ali Ansari. Currency Detection For Visually Impaired, © 2020 JETIR May 2020, Volume 7, Issue 5 (ISSN-2349-5162). [2] Paper Currency Identification Using image processing and radial basis functions (Rbf), Prakhar Chaturvedi, Harshdeep Kalra, and Ritu Raj Madhup, International Journal of Recent Technology and Engineering (IJRTE), ISSN: 2277-3878, Volume-7, Issue-6, March 2019. [3] P. Dhar, M. B. Uddin Chowdhury and T. Biswas, \"Paper Currency Detection System Based on Combined SURF and LBP Features,\" 2018 International Conference on Innovations in Science, Engineering and Technology (ICISET), 2018, pp. 27-30, doi: 10.1109/ICISET.2018.8745646. [4] Hameed, Shaimaa & Alwan, Mohammed. (2018). Paper Currency Detection based Image Processing Techniques: A review paper. Journal of Al-Qadisiyah for Computer Science and Mathematics. 10. 10.29304/jqcm.2018.10.1.359. [5] Althafiri, E., Sarfraz, M., & Alfarras, M. (2012). Bahraini Paper Currency Recognition. Journal Of Advanced Computer Science And Technology Research, 2(2). [6] J. Guo, Y. Zhao and A. Cai, \"A reliable method for paper currency recognition based on LBP,\" 2010 2nd IEEE InternationalConference on Network Infrastructure and Digital Content, 2010, pp. 359-363, doi: 10.1109/ICNIDC.2010.5657978. [7] Hassanpour, H. & Masoumifarahabadi, Payam. (2009). Using Hidden Markov Models for paper currency recognition. Expert Systems with Applications. 36. 10105-10111. 10.1016/j.eswa.2009.01.057. [8] A. Bhatia, V. Kedia, A. Shroff, M. Kumar, B. K. Shah and Aryan, \"Fake Currency Detection with Machine Learning Algorithm and Image Processing,\" 2021 5th International Conference on Intelligent Computing and Control Systems (ICICCS), 2021, pp. 755-760, doi: 10.1109/ICICCS51141.2021.9432274. [9] Kiran, Chinthana K, Mahendra K N, Sahana Y S, 2018, Feature Extraction and Identification of Indian Currency for Visuallu Impaired People, INTERNATIONAL JOURNAL OF ENGINEERING RESEARCH & TECHNOLOGY (IJERT) NCESC – 2018 (Volume 6 – Issue 13). [10] Reddy, Jagan & Rao, K.. (2020). Identification of Indian Currency Denomination Using Deep Learning. Journal of Critical Reviews. 7. 2020.

Copyright

Copyright © 2022 Raparthi Vivek, Soujenya Voggu, D. Vasu Mitra, N. Ganesh Gautham. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.