A Comparative Study in Usage of Convolutional Neural Network Models in Waste Classification
U.M.M.P.K. Nawarathne, Chamila Walgampaya · 2024
Waste accumulation has become a concerning problem over many years for the human race. Unmanaged waste poses risks to life, including hazardous effects, improper disposal, worker fatigue, and costly costs due to manual labor usage. Currently, in many countries, waste picking and classification are done manually. However, manual waste classification can lead to physical strain and health issues for human laborers. Considering these pressing issues it is necessary to introduce procedures that can automate manual waste classification. This research work proposes a waste classification solution using pre-trained convolutional neural network models. The Waste Recycling Plant Dataset-Classification (WARP-C) was used as the dataset, which contained pictures of bottles, cardboard, cans, canisters, and detergent materials. Before the model training phase, image interpolation methods such as inter-linear, inter-area, inter-cubic, inter-nearest, and lanczso4 filters were also used to upscale the images. Pre-trained models, MobileNet, MobileNetV2, Xception, InceptionV3, EfficientNetB0, EfficientNetB1, EfficientNetB2, EfficientNetB3, DenseNet121, DenseNet169, and DenseNet201 were used for classification. The DenseNet169 model outperformed the rest of the classifiers when applied with inter-linear interpolation with a model accuracy of 87.68% along with best scores for precision, recall, and f1-score.