Litter Detection from Digital Images Using Deep Learning

Machine learning to support citizen science in urban environmental management

“…These scores are higher than in other ML studies on imagery analysis, as discussed in Section 1.2 (except the work in Ref. [ 28 ] which also reported 0.98 accuracy), although there are several essential distinctions in these different ML applications. Other studies’ targeted problem with ML is litter detection; our goal uses litter data collected by humans to predict Litter Index.…”

“…The use of Convolutional Neural Network (CNN) models for deep learning on trash images was proposed by Mittal et al [ 24 ], who developed an Android app called SpotGarbage, which used CNN to produce a coarse-grained segmentation of garbage area and non-garbage area in imagery from heavily polluted regions. Other works have used CNN in various waste contexts, including to support identification and sorting in managed waste streams [ 25 , 26 ] as well as for in-situ litter monitoring using mounted cameras in vehicles [ 27 ] and surveillance videos [ 28 ]. To increase the imagery coverage for detecting litter in the environment using ML, unoccupied aerial vehicles (UAVs, aka drones) have also been used [ [29] , [30] , [31] ].…”

“…While this may be a cost-effective approach to expand the geographic scope and frequency of litter monitoring, it has its limitations. As discussed by Liu et al [ 28 ] about this type of approach, “small-sized objects in picture, luminance changes, moving occlusions and public facilities in ground all cause unexpected problems for litter detection.” Compared to image recognition, CS-based walking surveys can provide more accurate quantitative assessment, especially in the following conditions in cities: 1) when trash is partially degraded and blended within dry grass or bark, 2) areas not visible from common camera placement, e.g. behind bushes, 3) items that are smaller than surrounding items, especially in a low light situation.…”