<![CDATA[The significant increase in global waste requires innovative and accessible solutions, which aligns with Sustainable Development Goal (SDG) 12, which focuses on reducing the environmental impact of human activities. Automatic waste sorting using Computer Vision and Deep Learning offers a promising alternative to labor-intensive and risky manual methods. This study presents the design, implementation, and comprehensive performance analysis of an automated waste classification system, with a specific focus on evaluating its feasibility on hardware without specialized GPU accelerators. By leveraging transfer learning on a lightweight Convolutional Neural Network (CNN) architecture, MobileNetV2, a model was trained to classify six common waste categories: cardboard, glass, metal, paper, plastic, and other waste. The public “Garbage Classification” dataset from Kaggle, consisting of 2,527 images, was used as the basis for training and validation. The experiment was conducted using the tensorflow-cpu library, which does not require a dedicated GPU accelerator. After 10 training epochs, the model achieved a significant validation accuracy of 86.73%. Computational performance analysis showed an efficient average training time of 31.17 seconds per epoch and a fast average inference time of 14.47 milliseconds per image (~69 FPS) on the validation dataset. These findings demonstrate the feasibility of developing an effective AI-based waste classification system on hardware without a GPU accelerator, providing a realistic performance benchmark for the development of low-cost smart bins with embedded waste sorting solutions in the future, thereby contributing to sustainable waste management practices.]]>
Copyrights © 2025
