Title: Inference of Recyclable Objects with Convolutional Neural Networks

Abstract: Population growth in the last decades has resulted in the production of about 2.01 billion tons of municipal waste per year. The current waste management systems are not capable of providing adequate solutions for the disposal and use of these wastes. Recycling and reuse have proven to be a solution to the problem, but large-scale waste segregation is a tedious task and on a small scale it depends on public awareness. This research used convolutional neural networks and computer vision to develop a tool for the automation of solid waste sorting. The Fotini10k dataset was constructed, which has more than 10,000 images divided into the categories of 'plastic bottles', 'aluminum cans' and 'paper and cardboard'. ResNet50, MobileNetV1 and MobileNetV2 were retrained with ImageNet weights on the Fotini10k dataset. As a result, top-1 accuracy of 99% was obtained in the test dataset with all three networks. To explore the possible use of these networks in mobile applications, the three nets were quantized in float16 weights. By doing so, it was possible to obtain inference times twice as low for Raspberry Pi and three times as low for computer processing units. It was also possible to reduce the size of the networks by half. When quantizing the top-1 accuracy of 99% was maintained with all three networks. When quantizing MobileNetV2 to int-8, it obtained a top-1 accuracy of 97%.