Related Experiment Video
Updated: Jul 3, 2026
08:25
Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment
Published on: May 7, 2019
An enhanced EfficientNet framework for automated waste classification using cosine annealing and label smoothing
P Siva Arun Kumar1, Sriramakrishnan Pathmanaban1, J Mahalakshmi2
1Department of Mathematics, Amrita School of Physical Sciences Coimbatore, Amrita Vishwa Vidyapeetham, Coimbatore, India.
Scientific Reports
|July 1, 2026
Summary
A new EfficientNet-B1 plus Transformer model (EN-CALS) with optimized training achieves 93.68% accuracy for waste classification. This efficient model is suitable for resource-constrained hardware, addressing the global waste management crisis.
Area of Science:
- Environmental Science
- Computer Science
- Machine Learning
Background:
- Effective waste management is critical for environmental sustainability and public health.
- Increasing urbanization and industrialization have escalated waste generation complexity.
- Technology-based waste classification is essential to tackle the global waste crisis, especially on resource-limited hardware.
Purpose of the Study:
- To develop an accurate and computationally efficient waste classification model.
- To address the challenge of deploying models on resource-constrained edge devices.
- To establish a scalable model template and deployment pipeline for waste classification.
Main Methods:
- An EfficientNet-B1 plus Transformer model (EN-CALS) was developed, incorporating Cosine Annealing Learning Rate Scheduling (CALRS) and optimized Label Smoothing (LS).
- The EN-CALS model was trained on a dataset of 2,527 images across six waste classes: cardboard, glass, metal, paper, plastic, and trash.
- Performance was benchmarked against various fine-tuned models including ResNet-50, ViT-B/16, MobileNet variants, ShuffleNet-V2, EfficientNet-B0/B2, and DenseNet-121.
Main Results:
- The EN-CALS model achieved a top test accuracy of 93.68%.
- Large-scale models like ResNet-50 (87.37%) and ViT-B/16 (61.58%) exhibited overfitting and were unsuitable for edge hardware.
- DenseNet-121 reached 92.63% accuracy, but EN-CALS demonstrated superior efficiency and accuracy for the target application.
Conclusions:
- The EN-CALS model offers a highly accurate and efficient solution for waste classification on resource-constrained hardware.
- Ablation studies confirmed the significance of individual components within the EN-CALS system.
- This research provides a reusable model and deployment pipeline for scalable waste classification systems.
