COMET-Project transMVS (2019-2022)
Transfer Learning Based Machine Vision Systems
This project addresses the efficient semi-automatic setup of a knowledge base of domain-specific image data for training and analysis of deep models for machine vision applications. Transfer learning approaches are studied for efficient reuse of available data and pretrained models in similar domains. In particular, this project addresses the small lot-size problem in quality inspection in metal processing and injection molding industries and real-time tracking systems by exploiting deep transfer learning techniques.
The goal is to boost machine learning driven machine vision by exploiting domain adaptation and co-training techniques for the problems above
Other expected results are:
- benchmark datasets of AOI images and annotations for small lot size test cases from metal processing and injection molding industries, including the norm-eye problem
- Python software prototype for central moment discrepancy (CMD) based method for leveraging unlabeled data and data from different tasks for training neural models
- Python software framework for hyper-parameter optimization, testing and benchmarking selected state-of-the-art methods as outlined in the approach
- Python software prototype for transfer learning approach for the reuse of behavioral models
scientific analysis of outlined approach
- proof of concept and demonstrator of the norm-eye model
- feasibility study for its extension as a predictive model
- peer-reviewed scientific publications, e.g., Journal of Electronic Imaging, IEEE transactions on industrial informatics
This project is subsidized in the frame of COMET – Competence Centers for Excellent Technologies by BMK, BMDW, State of Upper Austria and its scientific partners. The COMET program is handled by FFG.