Baid idu Clo loud In Industry ry Quali lity In Inspection Solu - - PowerPoint PPT Presentation

baid idu clo loud in industry ry quali lity in inspection
SMART_READER_LITE
LIVE PREVIEW

Baid idu Clo loud In Industry ry Quali lity In Inspection Solu - - PowerPoint PPT Presentation

Baid idu Clo loud In Industry ry Quali lity In Inspection Solu lution Baidu Inc. Lei Nie 2019.3.18 Highest Labor Cost in Complex Cosmetic Defect Inspections Previo vious us machine hine visio sion n automat mation ion soluti


slide-1
SLIDE 1

Baidu Inc.

Baid idu Clo loud In Industry ry Quali lity In Inspection Solu lution

Lei Nie 2019.3.18

slide-2
SLIDE 2

Highest Labor Cost in Complex Cosmetic Defect Inspections

Target product(e.g.) Human inspections

industry daily output current cycle time labor amount

  • avg. labor cost

Monthly inspection labor cost % of inspection labor # 3C 2.3 million pcs 19 pcs/person/min 261 7800 RMB 2.03 million RMB from 20% to 40% in past 3 yrs. Previo vious us machine hine visio sion n automat mation ion soluti lution n still ll canno not t solv lve this s problem lem The comp mple lexity ity and varia iabil ility ity of cosmet smetic ic defe fect ct The comp mple lexity ity and varia iabil ility ity of products’ surface Limit itatio tions ns of traditi itional nal rule le- base sed machine hine- visio sion n algorithm hms

slide-3
SLIDE 3

Unified Terminal-Cloud Quality Inspection Solution

Human inspections

Intelligent Quality Inspection Integrated Machine Quality Inspection Box (Embedded Terminal) Model Store Model Synchronization AI model training closed loop Terminal- Cloud Cooperativ e Control Baidu Industry Quality Inspection Platform(Public Cloud) Multiple Computation Powers BIE runtime package Model Downloads Data Uploading Industrial Personal Computer Application Application 3C Iron Terminal Registration & Configuation Heterogeneous Model Optimization Device Admittance Intelligent Annotation Defection Synthesis New Defection Discovery Data Uploading Image Annotation Model Training Model Testing Model Integration Model Downloading Data Communication Etc..

slide-4
SLIDE 4

5 Key Features on Baidu Industry Quality Inspection

Human inspections

Fast Product Line Translation

Over 80% Data Labeling Time Reduced

Best Performance-Cost Balance

Intelligent Data Annotation Defection Synthesis for New Product Line New Defect Discovery

Discovery Unclassified Defect

Edge-Cloud Association

Model and Data Cycle

Heterogeneous Inference Engine

slide-5
SLIDE 5

Intelligent Data Annotation

Input Image Predefined Defected Area Interactive Segmentation Annotation : Slow( usually > 50s per object) → Potential Defect Suggestion ( > 80% precision ) Coarse ( approximated polygon) → Automatically Segment Foreground (only 4 corners)

slide-6
SLIDE 6

Intelligent Data Annotation- Interactive Segmentation

Interactive Segmentation: class-agnostic semantic segmentation with global context heatmap to acquire high accuracy Deep Segmentation from Extreme Points

slide-7
SLIDE 7

Heterogeneous Inference Engine

1、Multiple Deep Learning Framework 3、Address Pre-Mapping 2、Graph Merging and memory re-using

Anakin- Baidu Heterogeneous Inference Engine

4、 Heterogeneous Assembly Optimized

slide-8
SLIDE 8

Intelligent Defection Synthesis

New Product Line Cold Start: Transfer time-consuming labeled data collection (> 3 months) to 30 minutes

Mono Defect-Free Sample Mono Defect Sample Poly Defect-Free Sample Poly Defect Synthesised Sample Encoding Transformation Decoding Decision [0,1] Discriminative Network

slide-9
SLIDE 9

Intelligent Defection Synthesis

X G(X) F(G(X)) F(Y) G(F(Y) Y

slide-10
SLIDE 10

New Defection Discovery

Maintenance : Discovery un-classified defect for model optimization Novel Defect discovery: discover un-predefined defect by modeling defect-free image

Apple iWatch Charger Textured Anomaly Dataset

slide-11
SLIDE 11

New Defection Discovery

Normal samples Anomaly samples

DAE DAE

Network: ⚫ Reconstruction Network: Denoising/Variation AutoEncoder ⚫ Discrimination Network: Detect Anomaly Samples Loss Function: ⚫ Encoding Loss: latent variable loss ⚫ Content Loss: Ensure the reconstructed image ⚫ Adversarial Loss: Ensure the authenticity of samples generated

slide-12
SLIDE 12

Edge Cloud Synchronization

Openedge Remote mqtt

master

Agent

Reload

Engine (docker client)

Module Start/Stop

Api

container

Openedge Hub

container container

Openedge_Function

instance pool Routing Rule Cloud Management (iothub)

Docker network:openedge

mqtt mqtt grpc http mqtt mqtt/https function runtime instance container

1、Hardware and Framework Transparency

3、AI on the Edge Device

2、Edge-Cloud Association

4、Seamless Integration with Industrial Inspection Baidu Intelligent Edge(BIE) Framework

Developed in the cloud, Run on the edge Results upload for a further analysis Monitor, Config, Control

slide-13
SLIDE 13

Baidu Cloud ABC Integrated Industrial Inspection

Developed in the cloud, Run on the edge Results upload for a further analysis Monitor, Config, Control

High-Performance Industry Inspection Device Feature Strictly Meet AQL and FP for super-star Scenario Replace 100% labor + Computer Center + Quality Tracing Center Customer Digital Industrial Inspection for large company ( > 200+, 1 millon RMB per Machine) High Performance-Cost Industry Inspection Feature Basically Meet the AQL and FP Scenario Replace the Entire- workers(Verification worker reserved) ,rapidly deployment Customer Rapid labor-saving for small incorporation with sailing and renting (> 2000, renting for 10,000 per month)

slide-14
SLIDE 14