Comparative Analysis of CNN Architectures for Recyclable Waste Image Classification

Abstrak

Manual recyclable waste management has a low level of efficiency, prompting the need for Computer Vision-based automation. This study aims to analyze and compare the performance of Convolutional Neural Network (CNN) architectures in classifying five types of recyclable waste (glass, cardboard, paper, metal, and plastic). A controlled experiment was conducted using a Basic CNN model as a baseline, along with three transfer learning models (MobileNetV2, DenseNet121, and ResNet50) by applying uniform data preprocessing. The models were evaluated using a combined dataset of 15.501 images sourced from open-source repositories, including Kaggle and Google Dataset Search. The evaluation results show that MobileNetV2 achieved the highest performance with an accuracy of 91.54% and a stable F1-Score, demonstrating a strong capability in extracting complex visual features. DenseNet121 followed with an accuracy of 90.70%, while Basic CNN reached 81.25%. Conversely, ResNet50 experienced severe underfitting (43.64%) because the architecture expects a zero-centered input distribution and failed to adapt to standard normalization (1./255 scale). This study concludes that MobileNetV2 is the most ideal and efficient model to be implemented in real-world automated waste-sorting systems that demand both operational accuracy and computation simplicity.

Identifier

• DOI: https://doi.org/10.31154/cogito.v12i1.1094.81-91
• OpenAlex: https://openalex.org/W7202178965