EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Performance Evaluation of an Enhanced EfficientNet-B3 Model for Intelligent Waste Classification

Dongjin Fu, Alexis John M. Rubio

European Journal of AI, Computing & Informatics · 2026

DOI: 10.71222/52hpc271 ↗

Kurz gefasst

Fachartikel in der Zeitschrift European Journal of AI, Computing & Informatics, 2026. Verfasst von Dongjin Fu und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate and scalable waste classification is critical for advancing automated recycling infrastructure. This study presents a rigorously optimized variant of EfficientNet-B3 tailored for high-fidelity intelligent waste sorting under real-world constraints. The proposed model integrates architectural refinements—including adaptive input resolution scaling, channel-wise attention recalibration in the MBConv blocks, and class-balanced focal loss training—while preserving computational efficiency. A curated dataset of 42,860 high-resolution RGB images across seven waste categories (paper, cardboard, plastic bottles, plastic containers, aluminum cans, glass bottles, and organic waste) was assembled, augmented with physically plausible photometric and geometric transformations to simulate lighting variability, occlusion, and viewpoint shifts. Model evaluation employed stratified five-fold cross-validation on held-out test partitions, with performance benchmarked against baseline EfficientNet-B3, ResNet-50, and MobileNetV3. The enhanced model achieved a weighted F1-score of 96.7%, top-1 accuracy of 95.9%, and inference latency of 18.3 ms per image on an NVIDIA T4 GPU—surpassing the baseline by +3.2 percentage points in accuracy and reducing misclassification of visually similar classes (e.g., plastic bottles vs. glass bottles) by 41%. Confusion analysis revealed consistent improvement in fine-grained discrimination, particularly for translucent and reflective materials. These results demonstrate that targeted architectural and training optimizations yield substantial gains in operational robustness without compromising deployment feasibility, offering a validated pathway toward embedded vision systems for smart waste management.

Zitieren

APA

Fu, D., & Rubio, A. J. M. (2026). Performance Evaluation of an Enhanced EfficientNet-B3 Model for Intelligent Waste Classification. European Journal of AI, Computing & Informatics. https://doi.org/10.71222/52hpc271

Vancouver

Fu D, Rubio AJM. Performance Evaluation of an Enhanced EfficientNet-B3 Model for Intelligent Waste Classification. European Journal of AI, Computing & Informatics. 2026. doi:10.71222/52hpc271

BibTeX
@article{Fu2026,
  author = {Fu, Dongjin and Rubio, Alexis John M.},
  title = {Performance Evaluation of an Enhanced EfficientNet-B3 Model for Intelligent Waste Classification},
  journal = {European Journal of AI, Computing & Informatics},
  year = {2026},
  doi = {10.71222/52hpc271}
}
Kostenlos starten →