Visual inspection of glass bottlenecks by multiple-view analysis

The narrow structure of bottlenecks poses a very challenging problem for automated visual inspection systems and surprisingly, this issue has received little attention in literature. Bottleneck inspection is highly relevant to the fabrication of glass bottles, e.g., for the wine and beer industry. Defects in glass bottles can arise in various situations such as an incomplete reaction in a batch, batch contaminants and interactions of the melted material among others. This paper presents an inspection approach that utilises geometry of multiple views along with a rich set of feature descriptors to discriminate real flaws from false alarms in uncalibrated images of glass bottlenecks. The proposed method is based on an automatic multiple view inspection (AMVI) technique for the automatic detection of flaws. This technique involves an initial step that extracts numerous segmented regions from a set of views of the object under inspection. These regions are subsequently classified either as real flaws or as false alarms. The classification process considers that image noise and false alarms occur as random events in different views while real flaws induce geometric and featural relations in the views where they appear. Therefore, by analysing such relations it is possible to successfully localise real flaws and to discard a large number of false alarms. An important characteristic of the proposed methodology is the complete lack of camera calibration which makes our method very suitable for applications where camera calibration is difficult or expensive to carry out. Our inspection system achieves a true positive rate of 99.1% and a false positive rate of 0.9%.