CONVOLUTIONAL NEURAL NETWORKS COMPARISON FOR WASTE SORTING TASKS

We show convolutional neural network choosing process for RaspberryPi microcomputer. These networks are run on RaspberryPi, which is reverse vending machine part – device for collect and sorting of aluminum cans and plastic bottles. We chose and train three neural networks such as AlexNet, SqueezeNet, and MobileNet. For net-work training purpose, we use transfer learning, that is divided into two approaches: classifier replacement and retraining of convolutional neural network and weights fine-tuning of the pretrained network. As for framework, we use Caffe, because it is most popular. We evaluate these neural networks on validation set, which consists of 2300 photos. The validation set includes bottles, cans, and «other» trashes photos. We present the results of this evaluating. Also we evaluate networks on test set, which consists of 30 photos, and we present the results of these evaluating in detail. In this paper, the neural networks operation time on a computer and on RaspberryPi are given. Experiments are shown that MobileNet network has the highest accuracy and SqueezeNet has the mini-mum operation time per image on RaspberryPi.

Authors: K. R. Akhmetzyanov, A. A. Yuzhakov

Direction: Automation and Control

Keywords: Object classification, Caffe, Transfer Learning, neural networks, convolutional neural networks, Deep Learning, computer vision, waste sorting, RaspberryPi

---

View full article