AI for Glass Defect Detection: Scratches, Bubbles, and Inclusions

Glass manufacturing produces a material where visual perfection is often the primary quality requirement. Whether it is float glass for windows, container glass for bottles, or optical glass for lenses, defects that are invisible in opaque materials become obvious and unacceptable in glass. A single bubble the size of a pinhead can reject an entire sheet of architectural glass.

Human inspection of glass is inherently limited. The products are often large, the defects are often small, and the production rates are fast. AI-based machine vision closes these gaps.

Types of Glass Defects

Glass defects fall into several categories, each with different formation mechanisms and visual characteristics:

Gaseous inclusions include bubbles (blisters) and seeds (tiny bubbles). These form when gases do not fully escape the melt or when chemical reactions generate gas during forming. They appear as round or elongated voids within the glass.

Solid inclusions are particles of unmelted batch material (stones), refractory material from the furnace lining, or foreign contaminants. These appear as opaque specks within the otherwise transparent glass.

Surface defects include scratches, chips, and surface cracks from handling and processing. These are introduced after forming, during cutting, grinding, tempering, or transportation.

Optical distortion from thickness variations, surface waviness, or internal stress patterns. These may not be visible as discrete defects but cause the glass to distort transmitted or reflected images.

How AI Vision Systems Inspect Glass

The transparency of glass is both the challenge and the advantage for inspection. Because light passes through glass, you can use transmitted light to reveal internal defects and reflected light to reveal surface defects.

For flat glass inspection, the typical setup uses a line scan camera looking at the glass as it moves past on a conveyor. A bright, uniform light source on the opposite side of the glass provides transmitted illumination. Internal defects like bubbles, seeds, and stones scatter or block the transmitted light, creating contrast that the camera captures.

Surface defects are detected using reflected light from angled sources. Scratches scatter light differently than the surrounding surface, creating bright or dark lines depending on the illumination angle. The AI processes both transmitted and reflected images to build a complete defect inventory.

For container glass, the inspection geometry is more complex because the product is three-dimensional. Multiple cameras and light sources arranged around the container capture images from different angles. The AI assembles these into a comprehensive inspection that covers the entire surface including the bottom, sidewall, shoulder, finish, and sealing surface.

AI Advantages Over Conventional Systems

Older glass inspection systems used fixed threshold algorithms that compared pixel brightness to a reference value. These worked for large, high-contrast defects but struggled with small defects near the detection limit and generated excessive false alarms from normal variations in glass thickness or surface texture.

AI models handle these ambiguities much better. They learn to distinguish between a tiny bubble and a dust particle on the surface. They accommodate normal thickness variations that cause gradual brightness changes across the image. They recognize that a scratch on one side of the glass has a different optical signature than an internal inclusion at the same location.

The result is higher detection rates with lower false alarm rates, which is the fundamental tradeoff in any inspection system. More real defects caught, fewer good parts rejected.

Process Feedback

Defect data from AI inspection feeds directly back to process control. A sudden increase in seed count might indicate a temperature problem in the melter. Stones appearing from a specific location in the glass ribbon indicate refractory erosion at a particular point in the furnace. Surface scratches that correlate with a specific cutting station identify a handling problem.

This connection between inspection and process makes AI vision a process improvement tool, not just a quality gate. The goal is not just to reject bad glass but to make less of it.

For more on AI quality systems in manufacturing, visit the FirmAdapt manufacturing analysis page.