Computer Vision-based Waste Detection and Classification for Garbage Management and Recycling

Abstract

Waste management systems and their inherent problems are still a matter of great concern amid this cutting edge of science and technology. Piles of untreated waste in open locations contribute to a variety of issues, including air and water pollution, messy and unpleasant workplaces, waste of recyclable products, potential health risks for waste management workers, and so on. The root cause of this problem points to a single fact—too much manual labour involved in the garbage treatment process (collection, separation and recycling)—that can’t keep up to the pace with which garbage generation happens. An efficient recycling method is imperative to solve this problem, which can be achieved by a fast, real-time garbage detection and classification system. In this research, we will propose a novel Deep Learning-based approach for automatic detection and classification of five kinds of waste materials, namely Kitchen Waste, Glass Waste, Metal Waste, Paper Waste and Plastic Waste, from the garbage dump for an efficient recycling process, which not only improves the efficiency of the current manual approach but also provides a scalable solution to the problem. The contributions of this paper include a fully human labelled dataset consisting of 2,200 images of a garbage dump with 135,541 annotated objects from the aforementioned categories and a real-time garbage object localisation and classification framework based on a lightning-fast, fairly accurate and end-to-end deep learning algorithm. For the baseline, we have articulated a single-stage object detection and classification framework and initialised the detector training process with pre-trained weights (trained on MS COCO Dataset) of the feature extractor. Then, with some fine-tuning and employing a few transfer learning tricks, we have proposed a waste object detection framework that yields a mAP of 66.08% at an IoU threshold of 0.35 with an inference speed of roughly 55–58 milliseconds in a single GPU environment on both images and videos. It surpasses the performances of all the contemporary frameworks which deal with waste separation tasks.