The results of experiments on the development of neural network architecture to classify plastic bottles, aluminum cans and other objects are presented in the paper. The neural network is part of a smart container – a device for automated collection and waste sorting. This container consists of three waste cans (for collecting bottles, cans and other objects), the sorting actuator and RaspberryPi microcomputer. When developing a neural network, it is necessary to take into account the limitations of the computing power of this microcomputer with a high speed of image processing at any inclination and any distances to the object. Also, the neural network must properly classify crushed bottles and cans. The article presents the results of previous experiments conducted on the choice of a neural network among AlexNet, SqueezeNet and MobileNet. MobileNet neural network has achieved the highest accuracy. But the disadvantage of this neural network is the need to create a large training set for accurate recognition of crushed bottles and cans. The creation of such a training set with all the possible crushes would have required an huge amount of time. The article describes the original idea of the neural network, as well as the search for the most optimal architecture of this neural network, both manually and automatically.
Authors: K. R. Akhmetzyanov, A. A. Yuzhakov
Direction: Informatics, Computer Technologies And Control
Keywords: Object classification, neural networks, convolutional neural networks, deep learning, computer vision, waste sorting, RaspberryPi
View full article