Comparative analysis of ResNet-18 and efficientNet-B0 for lightweight deep learning in smart waste classification

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Waste mismanagement continues to represent a major environmental and operational challenge in urban areas, particularly due to inefficient waste segregation practices that reduce recycling effectiveness and contribute to increasing landfill accumulation. Although recent advances in deep learning have shown strong potential for automating waste classification, many existing approaches still rely on computationally intensive or hybrid architectures that are impractical for deployment in resource-constrained environments such as edge devices, embedded systems, and smart waste bins. This study presents a comparative analysis of two lightweight convolutional neural network architectures, ResNet-18 and EfficientNet-B0, for multi-class smart waste classification using the TrashNet dataset. Both models were trained and evaluated under a standardized experimental framework employing identical preprocessing pipelines, data augmentation strategies, optimization methods, and hyperparameter configurations to ensure a fair and reproducible comparison. The dataset consisted of 2,527 waste images categorized into six classes: cardboard, glass, metal, paper, plastic, and trash. Experimental results demonstrated that EfficientNet-B0 achieved superior overall performance, attaining 92% test accuracy and a weighted F1-score of 0.92, outperforming ResNet-18, which achieved 90% accuracy and a weighted F1-score of 0.90. Class-wise evaluation further revealed that EfficientNet-B0 performed better in well-represented categories such as cardboard, metal, paper, and plastic, whereas ResNet-18 exhibited stronger recall and F1-score in the minority “trash” category, indicating greater robustness in handling heterogeneous and underrepresented samples. These findings highlight the trade-off between computational efficiency and minority-class sensitivity in lightweight deep learning architectures. Furthermore, the study demonstrates that lightweight CNN models can provide a practical balance between predictive accuracy and deployment efficiency for real-time smart waste management applications in edge-based and IoT-enabled environments.