A.N. Kokoulin - Ph.D.(Eng.), Associate Professor, Department of Automatics and Telemechanics, Perm National Research Polytechnic University
A.I. Tur - Post-graduate Student, Perm National Research Polytechnic University
A.I. Knyazev - Student, Perm National Research Polytechnic University
A.A. Yuzhakov - Dr.Sc.(Eng.), Professor, Head of Department of Automatics and Telemechanics, Perm National Research Polytechnic University
A reverse vending machine (RVM) is a machine where people can return empty beverage containers like bottles and cans for recycling. The machine often gives back a deposit or refund amount to the end user. This is what makes it a «reverse» vending machine: instead of the user putting in money and getting out a product (like at a candy vending machine), the user puts a product in and gets out a monetary value. Reverse vending machines are a key part of container deposit systems in Europe and United States, which takes 70% to almost 100% of all drink containers returned for recycling.
For instance, RVM machines do it as follows: control of the material of the container (e.g. by IR-spectrometer), control of the shape of the container, control of the barcode. These three basic control-procedures prevent fraud or attempted fraud and make any attempt of the fraud completely impossible. But the same time it makes the RVM too expensive. With the modern computer vision technologies we can design another kind of efficient and non-expensive RVM having the same functionality. In our paper we consider some approaches in computer vision and image processing and their application to the problem of automatic recognition of empty recyclable containers (bottles and cans) and detecting fraud.
We have to restrict the list of the available methods and frameworks because of IoT controllers and tiny single-board computers we use have memory and computational restrictions. We cannot apply all the possible neural network and we decided to use some of the most popular networks based on Keras over Tensorflow and Caffe frameworks. They are LeNet, AlexNet, MobileNet. But also we have our first implementation of object detection algorithms based on Histogram of Oriented Gradients (HOG) image descriptor and a Linear Support Vector Machine (SVM) and we should start with this method. We should note that our aim is to classify the image inside the RVM by three possible classes: PET bottle, aluminum can or fraud (everything that doesn’t match PET bottle or can). We take into attention that those cans or bottles could be twisted or jammed and we included corresponding images into training and test sets.
For lowering of prime cost an attempt to refuse expensive electronics was made, using only visual recognition of an object on the basis of neural networks. Some of the most popular networks, are carried out for comparative tests and conclusions are completed.
Also this article analyzes the performance of image recognition programs using the MobileNet neural network and the OpenCV computer vision library in Python and C ++ languages. The purpose of the research is to identify the best option for implementation on low-power machines for application in the field of reverse vending machines.
In this research we noted that the program written in C++ best matches the image processing purposes in small micro-controllers than the Python program, because it is faster and uses less memory. According to this the C++ programming suits container identification purposes in developed reverse vending machines better. All the conclusions of this research also can be used for information security systems development, especially for AI security cameras and identifying devices.
- Reverse vending 101: a beginner’s guide. URL = https://www.tomra.com/en/collection/reverse-vending/reverse-vending-news/2017/how-does-a-reverse-vending-machine-work/ (accessed at 30 December 2017).
- Cost of a Deposit Return Reverse Vending Machine. URL = http://www.zerowastescotland.org.uk/sites/default/files/reverse%20v%20-%20CFE%20response.pdf (accessed at 30 December 2017).
- Rosebrock A. Histogram of Oriented Gradients and Object Detection. URL = https://www.pyimagesearch.com/2014/11/10/histogram-oriented-gradients-object-detection/ (accessed at 30 December 2017).
- Rosebrock A. Practical Python and OpenCV + Case Studies. URL = https://www.pyimagesearch.com/2016/11/20/Python-OpenCV/ (accessed at 30 December 2017).
- Iandola F. Forrest, Han Song, Moskewisz W. Mattewet. et al. SqueezeNet: MobileNet-level accuracy with 50x fewer parameters and <0.5 Mb model size // Proc. of ICLR’17 Conference. 2017. P. 207−212.
- LeCun Y., Bottou L., Haffner P. Gradient-based learning applied to document recognition // Proc. of the IEEE. 1998. № 86(11). P. 2278−2324.
- Krizhevsky A., Sutskever I., Geoffrey E. Hinton ImageNet Classification with Deep Convolutional Neural Networks. URL = https://papers.nips.cc/paper/ 4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf (accessed at 30 December 2017).
- Ogrel L.D. Ozenka nakoplenya, sbora i pererabotki othodov PETF v Rossii. [Estimation of gathering, utilization and recycling of PETF wastes in Russia] // Ekologicheskij vestnik Rossii. 2012. № 4. P. 26−31.
- А.N. Kokoulin, A.I. Tur, A.A. Yuzhakov Convolutional Neural Networks Application in Plastic Waste Recognition and Sorting // Proc. of the 2018 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering (2018 ElConRus). P. 127−134.
- Lasoff M.A. An RV By Any Other Name // Waste Age. 2000. V. 31. № 7. P. 34.