logo SBA

ETD

Archivio digitale delle tesi discusse presso l’Università di Pisa

Tesi etd-06102021-143011


Tipo di tesi
Tesi di laurea magistrale
Autore
RITACCO, ANTONIO
URN
etd-06102021-143011
Titolo
Automatic Scrap Classification using Deep Learning Techniques
Dipartimento
INFORMATICA
Corso di studi
DATA SCIENCE AND BUSINESS INFORMATICS
Relatori
relatore Prof.ssa Monreale, Anna
tutor Prof.ssa Colla, Valentina
controrelatore Prof. Bacciu, Davide
Parole chiave
  • convolutional neural networks
  • deep learning
  • image classification
  • instance segmentation
  • metal
  • object detection
  • scrap
  • semantic segmentation
Data inizio appello
25/06/2021
Consultabilità
Non consultabile
Data di rilascio
25/06/2091
Riassunto
This work aims at investigating the use of Deep Learning techniques to automatically detect and classify, through CCTV images, the scrap material entering the industrial site using a set of images collected by Danieli Automation.
The primary objective is evaluating the performances of Deep Learning models and libraries in scrap object segmentation and non-conforming object detection tasks. The secondary objective is to ensure the possibility of integrating such models in a web-based online tool that can be queried by technical personnel.
The object segmentation task was approached both as an Instance segmentation and a Semantic segmentation problem.
The instance segmentation problem was tackled by implementing a pipeline that includes a DetectoRS model with a mask branch plus a post-processing stage. The semantic segmentation problem comprises the generation of suitable semantic annotations from the original dataset and a UPERnet model with a Vision Transformer backbone that outperforms the instance, segmentation model.
The non-conforming object detection task was conducted in two successive phases. The first comprises a careful annotation of the non-conforming dataset and subsequently, a DetectoRS model was implemented whose performance are much higher than a baseline obtained with a Faster R-CNN
All the developed models are integrated into an inference pipeline and served by a dash based interface that can be queried by uploading a CCTV image and performing the required task
File