Automatic weld seam positioning on sheet metal enclosures by semi-supervised deep learning

Deep learning in computer vision has set new standards in mobile and web-based applications. The power of learning-based computer vision has also tremendous potential in machine vision. Traditionally, machine vision in manufacturing employs analytic solutions often resulting in excellent accuracy but poor robustness. The goal of this project is to increase robustness of a vision-based measurement process in sheet metal manufacturing using deep learning. The ability to accommodate variations in manufacturing enables a manufacturer to provide customized solutions in a more time efficient and cost effective way. One of the major challenges in machine vision is the lack of appropriate large-size training data for supervised learning. This project will train a deep learning algorithm based on all kind of data including expert-labelled images, existing results of a machine vision algorithm and unlabeled images. The project is to provide an effective solution for the industrial partner and general research results.

Faculty Supervisor:

Jochen Lang

Student:

Wenbin Zhang

Partner:

Enclosures Direct Inc

Discipline:

Engineering - computer / electrical

Sector:

Information and communications technologies

University:

University of Ottawa