Paper Details

Name: Shruti Handa & Mandeep Kaur
Country: India

Year: 2025
Volume: Volume-12, Issue-2 (July-December)
Page Number: 129-152
DOI: https://doi.org/10.5281/zenodo.17338006

ABSTRACT
With the expansion of urban and economic landscapes, the volume of solid waste generated
globally surges, posing significant environmental and public health challenges. Sustainable
waste segregation is essential for proper disposal, promoting recycling, and reducing landfill
accumulation, thereby supporting ecological balance. Existing studies leverage deep learning
for solid waste classification, but mostly datasets consist of single-object images on plain
backgrounds, which limits real-world applicability. To address this gap, a diverse dataset of
22,000 images spanning 12 waste categories is compiled from multiple public sources. Six
state-of-the-art pre-trained convolutional neural networks—DenseNet201, ResNet101,
EfficientNetB7, ConvNeXtBase, MobileNetV2, and InceptionV3—are fine-tuned using
transfer learning. Among these, ConvNeXtBase achieves the highest individual test accuracy
of 98.13%. To further improve performance, a hybrid model combining DenseNet201 and
ConvNeXtBase is developed using an attention-based fusion mechanism. This model
achieves a test accuracy of 98.45%, outperforming all single models. The results demonstrate
the effectiveness of attention-driven ensemble learning in complex waste classification tasks.
Future research emphasizes real-time deployment, adaptability across diverse waste streams,
and integration with edge devices while promoting sustainable waste management practices.
To further enhance accuracy, the study suggests expanding datasets, optimizing attention
mechanisms, and experimenting with architectures such as Vision Transformers.
Keywords: deep learning, hybrid model, transfer learning, ensemble learning, solid waste,
sustainable waste management