Second Prototype MileStone by 2/15/18: This is a non-robotic system to find similar looking, but small rare parts. The focus will be on fasteners, but generic functions will handle all kinds of small parts. Rare parts are hand picked from tray on computer display with a camera above it.
What will this do:
- It’s quick low cost solution for something not worth the time and effort of automation
- ID: What type of fastener is this? What type of thread?
- QA: Find rare bad parts
- Search: Finding the last parts that are really needed
- Manual sorting: Raises quality and flexibility with some speed gains
Usage:
- Put parts on the the tray
- Shake the tray a bit to separate them
- It’s ok if they are touching, but they can’t be stacked
- Parts will be highlighted by the display underneath if they need to be hand picked
Speed gains by using the system: Search speed will depend on the rareness of the parts and visual difference. A great example is rarity of one in one hundred. It might take someone 1 second a part vs 6 parts a second using the system. I have worked with coins, screws and legos so far to come up with that number.
With 99.99% accuracy:
- It will find parts that don’t belong
- It will find burrs, damaged, or unfinished areas 1/3rd of a thread diameter and up.
- Users can train their own items to classify as a set(the normal way)
- Users can quickly self-train their own items to classify as one. (train this part to itself)
- It’s not a sorting system, but it can speed hand sorting, by verifying the result
- It will ID all common threads from about 4 to ⅜, 2mm-10mm, machine, wood, sheet metal, etc
- You can get an inventory of parts on trays(just as a demo)
- Lengths on non-touching screws
99.99% accuracy applies to sizes larger than 1/3rd of a thread diameter:
- For example
- It will disciminate between 9 and 10mm long M3 screws
- It will notice a 1mm x 1mm area of crushed M3 thread
- It will highlight a 2mm x 5mm chip burr on a M6 screw
What it does with 99% accuracy on first pass:
- 99% will slow people down, but it’s still very useful
- Basically this means you will need to jiggle the tray more that once
- The operator might find part before the system does
- Lengths on non-touching screws
- It will find parts and areas of parts that just don’t match as well(for whatever reason)
- Internal thread ID
Notes:
- It’s not a precise measuring system
- The system is intended to stand on it's own without robotics
How does this work:
- It uses deep learning a ton of lighting and camera angle augmention
- See Repos:
- Lighting: lighting-augmentation repo
- SAIL: Unsupervised Labeling: sail repo
First proof of concept history:
- I build the first version in summer 2017
- I proved out this setup physically works(the ergonomics of it)
- It works and it's fun it use
- Used traditional CV functions
- But... it will not scale due to touching and training issues