A Weakly-Supervised Glass Bottle Defect Detection System Based on Multi-View Analysis

Defect inspection is an essential process in glass bottle production. Machine vision has shown great potential as an alternative to human inspection. In this paper, we present two open challenges in developing effective machine vision system, i.e., the constraint on an availability of training datasets with all relevant ground truth labels due to high annotation cost; and the difficulty in detecting defects in the texture region which may only be observable at certain angles. A novel weakly-supervised glass bottle defect detection system is proposed. The model features a two-stage learning process to recover missing labels, and train classifiers. A n IoT-based image acquisition system was designed to provide multi-view images of the bottle using a single camera. The proposed defect detection system is evaluated on glass bottle images acquired by our designed apparatus. The experimental results validate the effectiveness of the proposed system on a dataset with only single-positive labels with up to 96% accuracy.