GlossaryKind of term: Manufacturing termComputer vision
What is computer vision?
Computer vision is the field in which software identifies objects, shapes, text or defects in photos, camera images or scans.
The information software extracts from an image falls into four questions: which object is present, where it lies, what shape or size it has and whether a defect is visible. The image can be a photo, a frame from a measuring or inspection station or a scan of a drawing. OCR, converting characters into text, is one branch of it. The output is a classification or a value with an uncertainty.
What does computer vision depend on?
In manufacturing, computer vision appears in three forms. Industrial vision systems use a camera and controlled lighting to check whether a feature is present, whether a surface is damaged or where a part lies for a robot to pick. Optical measuring machines determine dimensions from an image. And at the request stage, scans and PDF drawings are analysed to locate views, dimensions, feature control frames and the title block. Some systems work with fixed rules and thresholds, others with a model trained on examples. Results depend on resolution, lighting, contrast, focus, the variety of what the system is given and, for a trained model, the examples it was trained on. On drawings, look-alike symbols, crowded dimensions and crossing lines make the task harder.
How does computer vision work in practice?
A camera image is a measurement only when the system is calibrated and its measurement uncertainty is known; otherwise it is an indication. For visual inspection you define what counts as a defect, with boundary samples, and how often the system may miss a defect or wrongly reject a good part. For drawings, every value found is a proposal that someone confirms against the drawing, because a missed symbol is a missed requirement. The cleaner the source, the less doubt: a digital drawing from the CAD system usually does better than a scan of a print.
What is computer vision often confused with?
Feature recognition works on the geometry of a 3D model, computer vision on pixels; they are different techniques. OCR falls within computer vision, aimed at characters; working out what a symbol or frame means takes more than converting characters into text.
Frequently asked questions
What is computer vision on the shop floor?
Software that analyses images: a camera that checks whether a feature is present or damaged, an optical measuring machine that determines dimensions, or a system that finds dimensions, tolerances and the title block in a scanned drawing. The output is a value to check.
Can computer vision handle a scanned drawing?
Yes, but the result depends heavily on the scan. A skewed, blurred or repeatedly photocopied drawing produces more doubtful cases than a PDF exported straight from the CAD system.
Can a camera measure dimensions?
Yes, with an optical measuring system that is calibrated and whose measurement uncertainty is known. Whether that uncertainty is small enough for the tolerance depends on optics, resolution, lighting and the part. An uncalibrated camera gives an indication, not a measured value.