- University of Ljubljana, Faculty of Computer and Information Science
- Kolektor d.d.
- ARRS (J2-3169)
With the introduction of the Industry 4.0 paradigm, machine vision systems are becoming an indispensable part of modern industrial production lines. They enable the digitization of perception and ensure the robust acquisition of visual information, which is crucial for decision-making processes. The main functional goal of the proposed project is to change the way machine vision systems are developed and implemented. Our goal is to shift the prevailing paradigm of manual development of specific solutions in the direction of data-driven design and development, which will enable more general, efficient, flexible and economical development and maintenance of machine vision systems. To achieve this objective, the main application goal of the project is to develop a software framework that will enable such development with the least possible and undemanding intervention of the human operator, which would reduce the need for manual data acquisition and labelling by automating the entire data preparation and model learning process. This goal requires solving a number of research problems, such as the development of basic deep learning methods and procedures for iterative, active, robust, weak, self-, unsupervised and few-shot learning of visual models, so that models achieve the required results with minimal the number of (manually annotated) learning images.
The expected contributions of the project are therefore:
- Advanced data synthesis and augmentation for supervised learning with unlimited annotated data.
- Efficient few-shot, active and robust learning with limited available annotated data.
- Unsupervised learning without annotated data.
- Application of the developed methods to the specific problems of 6DOF object pose detection and visual surface-defect detection.
Work packages: The work programme will be divided into six work packages. Research will be conducted in the first four work packages that will address the following objectives of the project:
- Development of advanced methods for data augmentation and generating synthetic data to enable supervised learning with unlimited annotated data (WP1).
- Development of new efficient, active and robust methods for learning from a limited amount of annotated data (WP2).
- Development of unsupervised learning methods for modelling visual appearance without annotated data (WP3).
- Application of developed methods on two applications: detection of the position of the object and **detection of surface defects?? (WP4).
- The remaining two work packages relate to the dissemination and exploitation of results (WP5) and project management (WP6).
- Year 1: Activities on work packages WP1, WP2, WP4, WP5, WP6
- Year 2: Activities on work packages WP2, WP3, WP4, WP5, WP6
- Year 3: Activities on work packages WP3, WP4, WP5, WP6
ARRS, Slovenian Research Agency