Upload
others
View
7
Download
0
Embed Size (px)
Citation preview
© 2019 Cognex Confidential1
How to use Cognex VisionPro ViDi to solve difficultapplications on the Factory floor?
Olivier Despont – Field Product Marketing Manager | July 4th 2019
© 2019 Cognex Confidential2
WHAT IS COGNEX DEEP LEARNING TECHNOLOGY?
human
performance
and flexibility
Reliability and
consistency
© 2019 Cognex Confidential3
WHAT IS COGNEX DEEP LEARNING TECHNOLOGY?
Different approach to solving vision problems
▪ Example-based training modeled on human learning
▪ Not a rigid rules-based solution
© 2019 Cognex Confidential4
How it works - Active Perception
CNN
© 2019 Cognex Confidential5
END-TO-END LEARNING & BLACK BOX RESULTS
Traditional Deep Learning
approaches
Cognex ViDi approach
scratch
dent
stain
dust
© 2019 Cognex Confidential6
COGNEX DEEP LEARNING FOR FACTORY AUTOMATION
Works on a commercial IPC
with high resolution images
Works with limited data sets
Bundled with VisionPro™ for the best of Deep
Learning and traditional visionDoesn’t need a Deep Learning Ph.D. to
configure & maintain
Image
Informat
ion
Confusi
on
Matrix
Score Plots
∆𝒘𝒋 =
−𝜼𝝏𝑱
𝝏𝒘𝒋= 𝜼σ𝒊(𝒕𝒂𝒓𝒈𝒆𝒕
(𝒊) - 𝒐𝒖𝒕𝒑𝒖𝒕(𝒊))(𝒙𝒋(𝒊)
)
© 2019 Cognex Confidential7
WHY VIDI?VS. TRADITIONAL MACHINE VISION
Human-like judgment Handles variable part appearance and unpredictable defects
Train by exampleField maintainable
© 2019 Cognex Confidential8
WHEN SHOULD WE USE DEEP LEARNING VISION?
ViDi Solves problems traditional vision cannot
Inspection
Deformed characterOCR
ViDi for unpredictable features
Distorted part location
VisionPro™ classic vision is best for
Precision alignment
1D & 2D code reading
VisionPro for consistent features
Gauging
© 2019 Cognex Confidential9
WHY VIDIVS. MANUAL INSPECTION
Faster▪ Keep pace with in-line production
Consistent results▪ Every shift, every line, every factory
Continuous improvement▪ Additional examples after initial deployment
Documented results
▪ Yield improvement via real-time process control
▪ Analytics data for offline improvement initiatives
▪ External reporting to end customer
▪ Traceability, for future failure analysis
© 2019 Cognex Confidential10
SolveComplex Inspection Human-like algorithm allows to tackle formerly hard-to-solve application by tolerating natural parts variation while focusing on what matters
ReduceMaintenance costEasy set-up and powerful algorithm configurable on Factory floor by operator
AT A GLANCE – BENEFIT OF VISIONPRO VIDI
Speed up DevelopmentNo programming thanks to easy set-up and powerful algorithm configurable through Intuitive GUI
Train & ProcessImages faster Powerful and Optimized Algorithms need low number of images and processed on single GPU
BenefitBest-in-class software & worldwidesupportBest Software for image analysis in Factory automation and supported by Cognex High-skilled engineers
ImproveQuality & Defects traceabilityConsistent 24/7 documented inspection results
IncreaseProduction YieldAutomatization of human low value and low performing tasks
6 -12 Months ROIPayback when deploying ViDi in the factory
© 2019 Cognex Confidential11
What applications can ViDi solve?
© 2019 Cognex Confidential12
Cosmetic InspectionSurface Inspection
Functional defect detection
Part locationDeformable part location and counting
Path following for robot
Pre-Assembly verificationPre-assembly clearance check
Part Correctness and Orientation
Kitting & Palletizing
ClassificationBulk/batch identification
Defect classification
Hard-to-solve OCRDistorted character detection
Post AssemblyPlacement check
Final assembly & packaging
verification
WHAT APPLICATIONS CAN VIDI SOLVE?
© 2019 Cognex Confidential13
Applications
Optimized tool for assembly verification
Confirm that all parts are present & correct
Share components across many layouts
For Automotive, Medical, Consumer Electronics, Food, …
BenefitsHandle 3D and deformed parts
Re-use component models without re-training
Optimized tool for OCR
Out-of-the-box reading
Retraining capabilities to be more application specific
Benefits
Facilitate image labelling
Read hard-to-tackle applications
Optimized tool for well-defined cosmetic inspection
Unsupervised/supervised inspection
Defects classification
Benefits
Train on few Good images
Handles variations without programing
© 2019 Cognex Confidential14
Automotive Fusebox Assembly
Wrong ComponentPass
© 2019 Cognex Confidential15
Electronics Assembly
Pass Missing Screw
© 2019 Cognex Confidential16
Food Packaging
Pass Missing Salami Wrong Placement
© 2019 Cognex Confidential17
Assembly Check tool demo
© 2019 Cognex Confidential18
Pre-closing verification and qualityinspection
Presence / absence and counting Cosmetic Inspection
© 2019 Cognex Confidential19
Metal Parts OCR
Trained on 183 images
Validation than 1000+ images
Results: > 99.5%
Time Cycle : 3 months
OCR examples
Generalized OCR for non-flat poorly inkjet printed label package
Trained on ca. 1000 images (99.5%)
Results after deployment: > 99.8%
Time Cycle : 5 months
Reading station images fed by DataMan fixed-mount readers
© 2019 Cognex Confidential20
Rotated Characters
▪ Added ability to read rotated characters (curved strings)
▪ Better out-of-the-box performance
▪ Easier and faster to develop OCR solution
Read Tool Enhancements with 3.4
ViDi 3.4
ViDi 3.3
© 2019 Cognex Confidential21
Well-defined Cosmetic Inspection Success Stories – Weld beads
Large variability across good parts
Large variability across bad parts but obvious differences compared to good parts
© 2019 Cognex Confidential22
Feasibility on 100s images
Applications building on 6535 images and Training the applications on 50% good (2969 images)
Fine tuning and re-training on reference set #1 Validation/Deployment/Optimization on 32,901 images
100% defect detection
99.82% good detection
(0.18% good parts consider as bad)
Well-defined Cosmetic Inspection Success Stories - Weld beads
Description Quality % Good as good 32,240 99.82%
Good as intermediate 20 0.06%
Good as bad 38 0.12%
Total good images 32,298
Bad as good 0 0.00%
Bad as intermediate 2 0.33%
Bad as bad 601 99.67%
Total bad images 603
Total images 32,901
Description Quantity %
Good as good 5937 99.82%
Good as intermediate 3 0.05%
Good as bad 8 0.13%
Total good images 5,948
Bad as good 0 0.00%
Bad as intermediate 1 0.17%
Bad as bad 586 99.83%
Total bad images 587
Total images 6,535
© 2019 Cognex Confidential23
Full comprehensive range of machine vision software platform allowing to combine best of both world to solve MV applications
What makes ViDi different from the competition – 8 reasons to convince
Easy to use interface helping
to reduce development time
Unsupervised Anomaly Detection
permitting to teach on good
images only
Most Powerful OCR and easy to
set up to solve complex
character reading applications
Chainable tools in the same
workspace to facilitate the
development of application
and speed up time to market
Pretrained font OCR to
reduce time to label
images
3x-5x Faster in training and
processing images
No programming required
allowing building
complex application
quickly
© 2019 Cognex Confidential24
How to start with Deep Learning ? Cognex can assist & support you
© 2019 Cognex Confidential25
A well-designed pilot project is key▪ Not too easy & Not too hard
▪ Constrain the application (Prioritization)
Clear understanding ▪ Acceptance criteria (well-defined, image-based, globally
agreed)
▪ Current process (Material handling, Inspection decisions,
Manufacturing yield, rework, throughput)
▪ Resources, schedule and budget
Integration in Production ▪ In-line Inspection (Simple step, high speed, camera
available)
▪ Final Inspection (Metrics exists, ROI, complex tasks, high
value station)
What makes a good first project ?
© 2019 Cognex Confidential26
Team : 4 core roles required ▪ Vision Developer (Implemented SW and build/optimize the
application)
▪ Quality Expert (Analyzes images to determine correct “classification” and determine Acceptance Criteria)
▪ Image Labeler (Marks defects and training features on the image database – CONSISTENCY)
▪ Data Collector (Record projects test and production data (images, results,…) and organize them (for debugging, yield calculation, validation)
Schedule :▪ Allocate sufficient time for factory-based
development (Image acquisition / Golden database definition may require several phases (system tested & product variations identified)
▪ Allow for AI learning curve (it is a NEW approach)
Statistically valid data sets are the only way to validate Deep Learning in production
Development Resources & Schedule
© 2019 Cognex Confidential27
Project phases
Prototyping
Understand the current process and determine if ViDi is a good candidate to solve it
Image Data Collection
Integrate the system on the production line and begin gathering and organizing data
Optimization (Lengthiest step)
Improve the ViDi solution until it meets the performance target
Validation & DeploymentQualify the solution and begin using it in production
Determine application requirements & prioritize.
Acquire a small database of graded & labeled
images.
Build a proof of concept system to test the
approach.
Integrate the camera and lighting on the
production line.
Begin logging image data and manual
inspection results (if any)
Establish ground truth data, optimize and
labelled image sets (CONSISTENTLY)
Run ViDi against your image data sets and in
production
Compare the results to ground truth and to
manual inspection results (if any)
Adjust the system and re-train as needed
(steady improvement over time)
Pass factory acceptance tests and lock
configuration (statistically proven)
Integrate into production and expand to
additional lines, prepare for future changes
Establish continuous monitoring process
Tasks / Actions
© 2019 Cognex Confidential28
Traditional
Machine VisionDeep Learning