Abstract
The process of selecting a convolutional neural network for the RaspberryPi microcomputer, which is part of a sorting machine – a device for collecting and sorting plastic bottles and aluminum cans, is shown. Three neural networks were selected for training: AlexNet, SqueezeNet and MobileNet. The training was conducted using Transfer Learning in two ways: replacing the last level of the classifier, then training and configuring the entire network. The Caffe framework was chosen as the basis for launching neural networks as the most popular. For the verification sample consisting of 2300 photos, which includes photos of cans, bottles and "other garbage", the correct recognition results are given, and the verification result is described in more detail in a small trial sample of 30 photos. The work gives time characteristics of neural networks on a computer and RaspberryPi. Experiments show that the MobileNet network has the highest percentage of correct recognition, while RaspberryPi has a network where the minimum time required to process one image belongs to SqueezeNet.