INTEGRATED BIBLIOMETRIC AND EXPERIMENTAL ANALYSIS OF DEEP LEARNING ARCHITECTURES FOR AUTOMATED WASTE CLASSIFICATION

Abstract

Accurate waste classification is a critical enabler of automated recycling and sustainable solid waste management (SWM). In this study, we present a comparative analysis of deep learning architectures for image-based waste classification using two benchmark datasets: (1) a large-scale Kaggle garbage dataset containing more than 13,000 images and 10 classes of waste and (2) the Trash Net dataset containing approximately 3,000 images and 6 classes of waste. Multiple convolutional neural network (CNN) architectures, including Efficient Net, Mobile Net, Res Net, Dense Net, and Inception, were eval­uated in terms of validation accuracy and model complexity. In addition to classification accuracy, the trade-off between predictive performance and model complexity was investigated to assess the suitability of different architectures for practical deployment for example, in the large-scale garbage dataset, the best-performing model achieved approximately 95.9% validation accuracy, while on Trash-Net the best result was approximately 94.3%. In addition to the experimental analysis, we conducted a bibliometric study of 84 research papers chosen using the PRISMA and Kitchenham’s guidelines using VOS viewer to identify major research themes, temporal trends, and collaboration patterns in AI-driven waste classification research. Bibliographic maps revealed that deep learning, waste classi­fication, machine learning, and computer vision constitute the dominant research themes, with recent studies increasingly emphasizing transfer learning and lightweight architectures. The paper contributes both an empirical benchmark across two datasets and a structured research landscape analysis, offering guidance for future work on efficient, scalable, and deployable waste classification systems.