What is Large-Scale Machine Vision (LSMV)?
Large-Scale Machine Vision(LSMV) or Large-Scale Automated Visual Inspection (LS-AVI) is the sub-field of Machine Vision (MV), and it refers to the use of a range of Machine Vision technologies for complex object inspections in the industry. Check What is Machine Vision? - #1 for more details.
When to use LSMV?
The source that leads to the difference between Large-Scale Machine Vision and traditional Machine Vision is the to-be-inspected object. The objects that require LSMV are typically large in physical dimensions; they are also complex in structural geometry; more than one optical characteristic is involved from a material, surface finish, and color; the ratio between features (typically as defects) and the required surface coverage is extreme; and mostly a direct online feedback loop within the quality/control process is required.
Here are some examples of objects/projects mostly involving traditional Machine Vision:
Circuit board inspection, wood/glass/metal panel inspection, consumer products inspection, barcode reading, assembly verification, tracking components, bin-picking, and fix-point surveillance.
Here are some examples of objects/projects mostly involving Large-Scale Machine Vision:
Full aircraft composite structured part inspection, car full surface inspection, construction inspection, patrol surveillance.
Let’s break down some examples to have a closer look at the difference.
For most consumer products, circuit boards can be completely captured by one or two images through a properly designed industrial camera system. This means that the inspection system can determine the condition of the goods after a few milliseconds of analysis. The under-quality products will be marked and picked from the conveyor belt. The good products leave the quality control cell within seconds, and inspection is done.
In contrast, inspecting a full building or a full aircraft need hours. Multiple vision and motion sub-systems work together to cover every corner and every sub-component. Hundreds of gigabytes, if not terabytes, of data, are acquired, synchronized, and analyzed. It is almost guaranteed that no 100% good aircraft or building exists. Most of them lie in the gray area—it is ok but needs some rework. Throwing away an aircraft or a wing makes no economic or financial sense; even hundreds of millimeter-range defects are detected. Repair requests are made, reworks are done, and revisitings of the defect area are planned.
Ten steps to start and finish an LSMV project?
#1. Be sure you have a full and practical list of requirements.
#2. Ensure you understand your object well enough.
#3. Decompose the object into basic pieces based on its appearance.
#4. Define the golden reference of the data quality for every basic piece.
#5. Merge the basic pieces into groups if they share the same (similar) golden reference.
#6. Design the vision systems for every group, and define the motion strategy for every group.
#7. Verify the systems individually against decomposed non-fake objects.
#8. Link groups into an LSMV system and optimize the transitions (software, hardware, motion) between groups.
#9. Verify the system against full-scale non-fake objects.
If necessary, and highly possible, you will need this additional step:
#10. Iterate between Step #3 and Step #9.
We will break down these ten steps later one by one.
How do I know whether my LSMV system is done right?
There is no hard line to tell when is right or wrong. The only metrics you can base yourself on are the number and the frequency of curse words shouted by the data analyzing team.
Joke aside, as explained in another article What is Machine Vision? - #1 , data analysis is the exit gate of all MV projects. It produces the final quality report. Implicitly, it also measures the quality of all MV components before itself. The longer the data analysis takes, the poorer the system is designed and the less reliable is the analysis report. You should consider redesigning the light system if you constantly try pre-processing techniques to remove image noise.