Transfer Learning LinkNet with SegNet for Waste Sorting

To make the cities more sustainable and less
damaging to the environment, they should control their
rubbish accordingly. Sorting trash into biodegradable and
non-biodegradable categories requires a lot of time and effort;
otherwise, errors occur frequently when done manually. This
research paper provides a hybrid deep learning system which
combines Modified LinkNet and Enhanced SegNet and a
transfer learning approach to achieve the goal of automated
waste classification using high accuracy. The improved SegNet
has the ability to clearly differentiate garbage objects on
complex backgrounds, unlike the modified LinkNet that uses
EfficientNet/MobileNet transfer learning techniques to classify
objects. A trained and tested model was trained using over
2,500 labelled images consisting of the TrashNet and TACO
datasets. The data was supplemented to strengthen it. Based on
the experiment findings, this model is 98 per cent accurate, 97
per cent precise, 96 per cent recall, and 98 per cent specific,
which is greater compared to the standard CNN, VGG-16, and
ResNet-50 architectures. The analysis conducted by ROC
indicated that the AUC equals 0.98, and it is a good indicator
that the model is very well able to distinguish between various
types of data. The training-validation curves indicated that
overfitting was not taking place too much. Being able to
leverage the advanced method in relation to smart waste
systems that can make the use of rubbish sorting scalable,
sustainable, and efficient in real time is a promising idea. The
future research will be done on the improvement of
categorizing different types of waste and real-time usage optimization on edge devices.