Dry Waste Segregation using Computer Vision

Waste management has been a constant problem for many countries, regardless of economic and social conditions [1]. The accumulation of garbage and its non-segregation for proper disposal has contributed to environmental and health problems in the most vulnerable population. In Honduras, only 2% of the waste is recycled, carried out by the informal sector, individual collectors, who separate the garbage for sale to the few companies that are dedicated to recycling. This method has proven to be ineffective, so there is a latent need to search for better solutions. In this study, a computer vision system is proposed for the classification of some dry waste, such as aluminum cans, PET bottles and glass bottles. The objective is that the program can detect waste in real time, considering the goal of being able to replace human labor for this work. Results showed that a dataset of 5,981 images, generated from data augmentation of 2,052 images produced significant improvement over the simple dataset. The resulting program meets the stated objective, achieving a mAP of 0.95, and with evaluation metrics such as accuracy and recall greater than 85%. In addition, detections can be made in real time, processing each image in 16.5 ms on average, achieving fast and efficient inference. It is concluded that the program can be implemented for the automatic segregation of aluminum cans, PET bottles, and glass bottles.