Salvato in:
| Autore principale: | |
|---|---|
| Natura: | Recurso digital |
| Lingua: | inglese |
| Pubblicazione: |
Zenodo
2025
|
| Accesso online: | https://doi.org/10.5281/zenodo.17037320 |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
Sommario:
- <div> <div>Industrial environments increasingly require machine learning systems that adapt to evolving data and new object classes. This thesis presents a modular, offline-compatible MLOps pipeline for class-incremental image classification, demonstrated on waste detection. Built with Apache Airflow, MLflow, DVC, PyTorch, and MinIO, the solution emphasizes reproducibility, scalability, and maintainability. It supports both initial and continual training using exemplar-based methods such as Dark Experience Replay (DER) and iCaRL, is fully containerized, and configurable via CLI, environment variables, and centralized constants for edge or on-premise deployment.</div> <br> <div>Experiments on annotated metal waste datasets achieved a mean Average Precision (mAP) of approximately 0.24 after incremental updates with DER, showing strong performance on distinctive components but limitations on smaller or visually ambiguous ones. The iCaRL method yielded lower mAP (~0.002), affected by class imbalance and feature ambiguity, highlighting challenges in exemplar selection and representation. Close collaboration with RIWO's R&D and robotics engineers informed design choices and ensured alignment with operational constraints. The resulting framework is reusable for industrial settings where data evolves and full retraining is impractical.</div> </div>