Waste Sorting Neural Network Architecture Optimization
Kirill Akhmetzyanov, Alexander Alexandrovich Yuzhakov · 2019 International Russian Automation Conference (RusAutoCon) · 2019
The results of experiments on the search of neural network architecture to classify plastic bottles, aluminum cans and other objects are presented in the article. This neural network is a part of a smart container intended for sorting and collecting waste. RaspberryPi microcomputer is used as a computing device in that container, so when designing a neural network it is necessary to take into account the device's computer power limitation. A smart container is aimed to sort correctly both flat and crushed bottles and cans. The results of previously conducted object classification using neural networks AlexNet, SqueezeNet and MobileNet experiments are also presented. However, the disadvantage of these networks is the need to create a large training set for their training. Moreover, it is necessary to create a set with all the possible crushes of bottles and cans, which will require a huge amount of time to recognize the crushed bottles and cans. Therefore, a small set should be required to train the network being searched for. The search for the architecture was carried out for the original neural network (approach to the development of which was proposed in the article) using the hyperparametric optimization algorithm and manual search of architectures. The developed neural network must meet the following requirements: high speed of image processing, a small amount of required memory, high accuracy rate of object recognition under various lighting, backgrounds, angles and geometric deformations.