Litterlens – A deep learning based urban waste detection tool

Abstract

The increase in solid waste pollution and its accumulation due to population expansion is becoming a significant threat in developing countries like Pakistan. Pakistan ranks as the 5th most populated country in the world and produces almost 49.6 million tons of solid waste annually, which has been increasing at the rate of 2.4% each year. Efficient solid waste identification and collection strategies still need to be improved in Pakistan in contrast to conventional methods, which result in inadequate allocation of resources to areas in need and ineffective waste collection operations. This research bridges the gap by providing a deep learning-based solution for the efficient identification of areas with high waste volume by developing a data-driven tool that promotes community involvement to apply a customized waste management approach in Pakistan. A local dataset of 3693 waste images was collected from different cities in Pakistan to train the deep learning models, as no specialized waste dataset was available for Pakistan. This study used three deep learning models, i.e. Deep CNN, You Only Look Once (YOLO) v8 classification model, and Visual Geometry Group (VGG)-19. Overall, all models achieved more than 90% accuracy when trained at 20 and 50 epochs. However, among these networks, YOLOv8 was the highestperforming model with an accuracy of 99.5% at 50 epochs. Furthermore, a functional prototype for the tool was created using the Python Tkinter package, which integrated the best-performing model, allowed the upload of images, and provided classification results in an inference time of under 1 second. Hence, this deep learning-based solution is an efficient approach to waste management in Pakistan and has the potential to be implemented with further improvements along the way.