Machine learning for visual inspection systems

点击翻译按钮获取中文摘要

Sightline Innovation is a company that provides artificial intelligent products and solutions for businessesaround the world. One area of interest for the company is visual inspection for quality control inmanufacturing. The goal is to detect defects in manufactured parts and classify the type of defects. Since visualinspection by humans is time-consuming and costly, it is imperative to develop visual inspection systems thatcan automatically detect and classify defects from images. Traditional visual inspection systems are designedusing hand-engineering features. However, since Sightline has clients from multiple industries, the visualinspection tasks of these clients can be vastly different depending on the industry. It is nearly impossible todesign visual features that can work effectively across various visual inspection tasks. So engineers at Sightlineend up spending a lot of time hand-tuning the system for each task.Currently, the company has access to a large amount of images for various visual inspection tasks fromdifferent clients. The company is interested in exploiting these data to develop machine learning basedsolutions for visual inspection systems that can be quickly adapted to specific clients.In the proposed research, we will study two specific problems. First, we will develop algorithms for localizingdefects in images from weakly labeled data. Second, we will develop domain adaptation techniques for visualinspection. This project will produce a set of algorithms and prototypes for visual inspection that the companycan build upon and integrate with its products.