Recyclable Waste Classification Using a Deep Learning Vision System

Resumo

One of the biggest environmental problems that humanity have been facing is the amount of waste
generated. The disorderly growth of large cities combined with consumption and industrialization are substantially
increasing the amount of solid waste generation. Recycling is an essential resource for sustainable development of
societies whereas, as it reuses waste and decreases the accumulation of garbage. Brazil is one of the countries in
the world where most generate solid urban waste and which has one of the largest numbers of people who separate
this waste manually, many times in deplorable working conditions. Selective waste sorting basically consists of
segregation of recyclable waste. However, when performed manually, the practice of segregation may not be
followed evenly. The use of automated systems are an alternative to make the waste segregation more efficient
and safer. This paper describes the use of a computer vision-based method to detect 4 main types of solid waste:
glass, metal, paper and plastic. To classify and detect each type of solid waste was used a Convolutional Neural
Network Based on the Mask Region (Mask R-CNN). The classification generates a mask and bounding box. After
training the model using TrashNet dataset was achieved 0.80 for mAP. The use of those information can provide
the position of the objects at the scene and robotic arms can make the automatic waste sorting.