Abstract:
Growing landfill volumes and slow, labour-intensive sorting have made urban waste one of today’s
least glamorous bottlenecks. To tackle that gap, this study builds an image-based classification
pipeline that recognises eleven common rubbish types ranging from PET bottles to greasy cardboard
in a single camera shot. Four convolutional backbones were examined: ResNet-50, MobileNet V2,
EfficientNet-B0 and DenseNet 121. Each network was fine-tuned with transfer learning on a
purpose-built collection of roughly 30 560 images per class (≈ 336 k photographs overall). Prior to
training, every picture was resized to 224 × 224 px and augmented through random flips, rotations,
colour-jitter and brightness shifts to mimic curb-side variability. Model performance was judged
with accuracy, F1 score and class-level confusion matrices. EfficientNet-B0 delivered the best
balance of speed and precision, attaining 94 % validation accuracy and a macro-F1 of 0.93, while
MobileNet V2 finished close behind but trained 40 % faster an advantage for edge deployments or
rapid re-training cycles. All weights were exported as compact .h5 files and tested on unseen street
images, confirming real-time inference on desktop GPUs and Raspberry Pi boards alike. By
removing much of the manual effort from waste sorting, the proposed system cuts worker exposure
to hazardous material and improves the purity of downstream recycling streams. The results
demonstrate that modern, lightweight vision models can be integrated into smart-bin infrastructure
or material-recovery facilities with only modest computational budgets, offering a practical route
toward cleaner, more sustainable cities.
